Picture it¹: London, 2026. You spot a new trendy bakery on social media and decide you want to try it. You look up the address, type it into your GPS, and follow the route. You never memorise the address. On your way there you never think about which roads to take. You arrive without any of that information passing through your brain in a meaningful way — and that's completely fine. Instead, you focus on the latest episode of your favourite podcast.

That is cognitive offloading: redistributing mental effort to an external tool so your brain can focus elsewhere. It’s been part of human cognition for as long as we’ve had tools. Writing things down is cognitive offloading. A shopping list is cognitive offloading. Psychologists have studied it extensively and the verdict is clear: it frees up mental resources, reduces error, and is generally a smart way to navigate a complex world (Risko & Gilbert, 2016). Sure, there have been concerns that over-relying on external tools can erode underlying skills (e.g., the worry that GPS is slowly killing our sense of direction isn’t entirely unfounded). Broadly speaking though, offloading is considered a sensible adaptation rather than a cognitive threat.

The key thing about cognitive offloading is that you remain in charge. You decided to go to the restaurant. You chose which address to enter. You’re still the one making judgements, while the tool just handles the execution.

So what’s different about AI?

You’ve probably heard the argument that AI is just the latest cognitive offloading tool. Based on a new paper by Shaw and Nave (2025), that framing may be missing something important.

Their argument is that AI doesn’t just handle execution, it generates the judgement itself. Ask your GPS to navigate and it follows your instruction. Ask an AI whether you should take the job, how to respond to that difficult email, or what the diagnosis might be, and it hands you a complete, fluent, confident answer. The thinking has already been done on your behalf. All that’s left is whether you accept it or challenge it.

Shaw and Nave call this cognitive surrender: adopting an AI’s output with minimal critical evaluation, effectively substituting it for your own reasoning. Unlike cognitive offloading — where you delegate a task while staying in the driver’s seat — cognitive surrender is a transfer of the wheel itself. You’re no longer steering; you’re just along for the ride.

To capture this formally, they propose the Tri-System Theory, which extends the classic dual-process framework (System 1: fast and intuitive; System 2: slow and deliberative) by adding a third system:

• System 1: fast, automatic, gut-feel responses

• System 2: slow, deliberate, effortful reasoning

• System 3: external, AI-generated cognition that operates outside the brain entirely

The critical point is that System 3 doesn’t just support Systems 1 and 2 but can bypass them. When that happens, the decision that gets made may reflect the AI’s reasoning rather than yours. You might not even notice the difference.

What the experiments showed

Shaw and Nave tested this across three preregistered studies2 (N = 1,372; 9,593 trials). Participants solved problems from an adapted Cognitive Reflection Test (CRT), a set of reasoning questions designed so that the first, intuitive answer is wrong, and getting it right requires pausing to think more carefully. It’s a well-validated tool for studying exactly the tension between fast and slow thinking. An example from the CRT is the classic:

”A bat and a ball cost $1.10 in total. The bat costs $1.00 more than the ball. How much does the ball cost?”

The gut answer is 10 cents, while the correct answer is 5 cents.

Some participants worked alone. Others had access to an embedded AI assistant (ChatGPT; GPT‑4o). Unbeknownst to the participants, the AI was secretly set up to give the correct answer on some trials and a confidently worded but wrong answer on others. Participants could consult the AI or not, and follow or ignore its responses freely.

Here are some of their most interesting findings:

• Participants consulted the AI on more than half of all trials

• When they did consult it, they followed its advice 93% of the time when it was correct and 80% of the time when it was wrong

• Compared to working alone, having accurate AI raised performance by ~25 percentage points; having faulty AI dropped it by ~15 points

• Access to AI increased confidence by ~12 points regardless of whether the AI was actually right

That last finding is the most alarming one. Participants felt more confident in their answers when they’d used the AI, even when approximately half of those AI outputs were wrong, and even without any decline in confidence as errors accumulated. They didn’t know the AI was unreliable, but the point is they didn’t seem to be checking.

The effect size for the gap between accurate and faulty AI trials was large (trial-weighted Cohen’s h = 0.82 — on a scale where 0.2 is small, 0.5 is medium, and 0.8 is large). Accuracy didn’t reflect the participants’ reasoning. It reflected the AI’s quality.

Who surrenders most and least

Not everyone surrendered equally. Across studies, three individual differences consistently predicted susceptibility:

• Higher trust in AI → more surrender. These participants consulted the AI more often, followed its wrong answers more frequently, and showed a larger accuracy gap between accurate and faulty trials.

• Higher Need for Cognition (the stable tendency to enjoy and engage in effortful thinking; Cacioppo & Petty, 1982) → more resistance. These participants were more likely to override a wrong answer.

• Higher fluid intelligence → similar resistance, with more override behaviour when the AI was wrong.

The uncomfortable implication of this is that the users most likely to surrender to a faulty AI — those who trust it readily and prefer not to engage in analytical thinking — are probably not unusual or outlier users. They’re likely representative of a large proportion of people using consumer-facing AI products today.

Can we reduce cognitive surrender?

Two follow-up studies tested whether situational factors could reduce cognitive surrender.

Time pressure made things worse. Under a 30-second deadline per question, overall accuracy dropped, and among participants who used the AI regularly, performance became even more tightly coupled to AI quality. When the AI was correct, they outperformed everyone else. When it was wrong, they performed worst of all. Time pressure didn’t make people more independent; instead, it made them more reliant.

This reminds me of autonomous vehicle research. Getting humans to re-engage quickly and effectively when an automated system needs them to intervene is one of the hardest unsolved problems in that field — and automation complacency, the tendency to stop monitoring a system because it's usually right, has been documented across industries for decades (Parasuraman & Manzey, 2010). A similar dynamic seems to be at play here: the more we defer, the less ready we are to take back the wheel when it matters.

Incentives and feedback helped — partially: When participants received a small financial reward for each correct answer and immediate feedback after each response, override rates on faulty AI trials more than doubled (from ~20% to ~42%). The cognitive surrender effect shrank but didn’t disappear with the accuracy gap between accurate and faulty AI remained around 44 percentage points (versus ~50 without incentives).

So motivated reasoning with real-time error signals can reactivate critical evaluation, but in the absence of explicit incentives and feedback, the default is acceptance.

What this means for design

The paper frames cognitive surrender as a design and education challenge rather than a reason to panic.

• Confidence inflation is a design problem. AI interfaces that present outputs with fluency and authority actively encourage surrender. Even small interventions that signal the AI’s limitations or prompt a moment of verification can partially reactivate deliberative thinking (e.g., uncertainty markers).

• Engagement metrics can mislead. High AI feature usage can look like success in your analytics while masking uncritical adoption. What matters is calibrated engagement — using the AI when it helps and questioning it when something seems off — not maximised engagement. Override behaviour is worth measuring, not just follow rates.

• Your most typical user may be the most vulnerable. The protective factors here (i.e., strong analytic tendency, high cognitive reflection) aren’t evenly distributed. Designing as though all users will independently verify AI outputs isn’t a safe assumption.

What the research doesn’t resolve

A few limitations are worth flagging. The studies were conducted in controlled lab settings using a specific reasoning task, which is excellent for studying the tension between intuitive and deliberative thinking, but doesn’t capture the full range of real-world decisions people make with AI. Whether the same surrender dynamics apply to messier, more ambiguous, or emotionally loaded situations remains an open question.

The studies also offer a single-session snapshot. In real-world use, people interact with AI repeatedly over time, building or eroding trust as they go. Whether cognitive surrender deepens or attenuates with experience and whether people learn to detect unreliable AI outputs over time isn’t yet clear.

The bigger picture

I could end this by ringing alarm bells but I choose a different approach. This isn’t necessarily a doom and gloom story — it’s an exciting one. We are only beginning to understand what happens when human cognition and AI start working together, and frameworks like Tri-System Theory give us the vocabulary to ask better questions.

How does repeated AI use change the way we think over time? When does collaboration become dependency? How do we design for the full range of human-AI interactions — not just the ideal user, but the one who trusts the AI unconditionally?

There’s a lot of research still to be done. And for designers, researchers, and anyone building with AI, that’s an opportunity to innovate and be creative.

I’ve been working in product research for a while now, and something has shifted noticeably over the past couple of years. The requests we get from stakeholders have changed. The way we conduct research has changed. Even the way teams consume the insights we share has changed. I’ve been talking to friends in UX and they’re noticing similar things. The pace is different, the expectations are different, and the role itself feels like it’s being renegotiated in real time.

Some of this is genuinely positive. But some of it concerns me, and I don’t think we’re talking about it honestly enough.

Everything is faster — including the parts that shouldn’t be

The most immediate change is pace. Timelines for planning and conducting research have shortened dramatically. Stakeholders expect faster turnaround on everything: the research plan, the interview guide, the analysis, the report. AI tools have made some of this possible, and the efficiency gains are real — 58% of product professionals now use AI in their research workflows, up from 44% in 2024 (Maze, 2025), with researchers reporting faster turnaround times and more streamlined analysis.

The problem is that speed and rigour are not always compatible, and some parts of research resist compression more than others.

Thematic analysis is a good example. Done properly, it takes time. You need to sit with the data, read and reread transcripts, let patterns emerge rather than impose them. Many researchers deliberately leave a gap between data collection and analysis for this reason — the patterns you notice on day three are often different from the ones you notice on day one. In many product environments, this kind of reflection can easily appear inefficient or difficult to justify. In reality, this is how qualitative analysis works (e.g., obtaining insight is embedded in a dynamic relationship between researcher and data, Hitch, 2024). Rushing it, or outsourcing it entirely to AI, does not produce the same output faster. It produces a different, shallower output. A series of studies from practitioners(e.g, Luna, 2026) and academics (e.g., Ozuem et al., 2025, Nguyen-Trung, 2025) have repeatedly found that AI-generated themes tend to be surface-level — technically accurate but missing the deeper context behind what participants are actually saying. For example, an AI-generated synthesis might conclude that “users value simplicity” or “participants want clearer workflows”. Technically, those themes may be present in the data, but a human researcher may notice that what participants are actually expressing is anxiety about making expensive mistakes, fear of appearing incompetent, or uncertainty about organisational expectations. The surface theme is not necessarily wrong, but it is often incomplete.

The cumulative effect of this is researchers who are stretched thin, moving from one study to the next without adequate time to think. Add continuous research and non-stop stakeholder demand to the mix and you have a reliable recipe for burnout. Lyssna’s 2026 research trends report based on data from 100 researchers found that 21% identified balancing speed with research quality as their single biggest challenge.

Polished outputs are not the same as good research

There is a related problem with democratisation. AI tools now allow product managers, designers, and other non-researchers to create research artefacts that look professional — surveys, discussion guides, even synthesised insight reports. This sounds like a good thing and it can be. In practice, however, it is not straightforward.

A survey created by someone without research training might look polished while containing leading questions, poorly ordered items, or response options that introduce systematic bias. The problem is that it is increasingly difficult to tell the difference at a glance. Takafoli, Li, and Mäkelä (2024), in interviews with 24 UX practitioners, found that most companies have no formal policy or governance around AI use in research. Individuals are making their own decisions about what constitutes good enough, with limited organisational oversight.

This matters because research quality is not always visible in the output. A well-designed study and a poorly-designed one can produce reports that look similar. The difference shows up in whether the findings are actually valid — and by the time anyone realises that, decisions have already been made on the basis of them.

The slow disappearance of methodological rigour

The changes to how research is communicated worry me in a similar way. Long reports are increasingly unpopular — 71% of researchers experimented with new formats for sharing insights in 2025 (User Interviews, 2025), and the trend is firmly toward shorter, snappier outputs. Slide decks, one-pagers, short video summaries, even podcasts. There is nothing inherently wrong with adapting communication formats to your audience (throwback to my article on UX Research Bingo). Stakeholders are busy. Long reports are boring. Getting findings read and acted on matters as much as conducting good research.

But something is being lost in the process. Traditional research reports, whatever their faults, typically documented methodology, sampling decisions, and limitations. They made explicit what the research could and could not reasonably conclude. Increasingly, these are being dropped in favour of outputs that lead with punchy headlines and actionable recommendations. The result can look like research and sound like research while missing some of the scaffolding that makes research trustworthy.

Research is grounded in the scientific method. Methodology matters. Limitations matter. Not because stakeholders want to read about them — they largely do not — but because they are what distinguish a genuine finding from a plausible-sounding assumption. When we stop including them, we may make research outputs easier to consume, but also harder to interrogate and easier to misuse. We are making it harder to interrogate, and easier to misuse. The risk is that research becomes performative — the appearance of insight without the methodological depth underneath it.

There is also another thing to be mindful of here. Stakeholders are increasingly using AI tools to summarise research outputs rather than reading them directly. A report with its context, caveats, and interpretive framing gets compressed into a bullet-pointed digest. The nuance that distinguishes a useful finding from a misleading one often does not survive that compression. This specific behaviour is not yet well documented in the research literature — it is more observable at the practitioner level — but it has direct consequences for how findings get applied downstream.

The structural picture

None of this is happening in a vacuum. The conditions under which researchers are doing this work have also changed. Whatever the efficiency gains, researcher sentiment tells a different story: 49% of researchers felt negatively about the future of UXR in 2025, a 26-point increase from 2024, with 67% giving a negative outlook on career opportunities (User Interviews, 2025).

The job market reflects this. According to Indeed data analysed by Brookshier and Altenhoff (2026), UX and product design job postings dipped below their pre-pandemic baseline in Q3 2023 and have not meaningfully recovered. Listings for UX research roles specifically fell below 1,000 in early 2025 (Burgess, 2025). 35% of organisations reported reducing UX staff in a MeasuringU survey. Perhaps most telling: the ratio of people who do research to dedicated UX researchers has shifted from 2:1 in 2020 to 5:1 in 2025 (UX Magazine, 2025). More research is happening. Less of it is being done by researchers.

Demand for research outputs is rising — 55% of respondents in Maze’s 2025 report said demand has increased, and 87% of organisations say they use research to inform critical decisions. But rising demand does not automatically mean better conditions for doing research well. When headcount shrinks and timelines compress, the research that gets done tends to be faster and lighter.

What we risk losing

AI is a genuinely useful tool for many parts of research work. The efficiency gains in transcription, recruitment, and initial data organisation are real and valuable. Used well, AI should free researchers to spend more time on the work that requires human judgement — the interpretation, the contextual reading, the synthesis that produces insight rather than just information.

The problem is that this is not necessarily what is happening in practice. Instead of using AI to create space for deeper thinking, we are using it to do more in less time. The deep thinking is what’s at risk of being cut. And that thinking — the unhurried reading of transcripts, the pattern that only becomes visible on reflection, the finding that reframes the whole project — is where the actual value of research lives. For many of us, that interpretive work is part of what made research meaningful in the first place.

Lu et al. (2024), in a systematic review of 359 papers on AI and UX, make this point directly: UX research is fundamentally about building empathy with users, not completing tasks. Automation that treats research as a series of completable tasks misses this. AI may surface statistically probable information about users. That is not the same as understanding them.

The researchers who will be most affected by these trends are not those in senior, embedded roles with strong organisational influence. It is those earlier in their careers, or in organisations where research has always been treated as a support function. The “AI elevates the researcher” story is largely being told by people who were already elevated.

What we can do about it

None of this is straightforward to address. The pressures are structural, not personal, and individual researchers cannot reverse them alone. But there are things worth doing.

The first is to be honest about trade-offs rather than just delivering outputs. If a study was conducted faster than the methodology strictly allows for, say so. Stakeholders who understand the difference between a directional finding and a validated one are better positioned to make decisions. Researchers who are transparent about this are also harder to blame when fast research produces uncertain answers.

The second is to protect methodology even in short outputs. You do not need a five-page methods chapter in every report. You do need enough context for the reader to understand what the findings can and cannot conclude. That is what separates researcher-led work from AI-generated research-shaped content.

The third, and perhaps most important, is to use expertise to advise, not just deliver. Fast and light research is sometimes the right call. Continuous discovery, quick directional studies, lightweight validation — these have a legitimate place in the research toolkit. The problem is not speed itself, it is speed applied indiscriminately. Researchers are best placed to judge what a given question actually requires, and that judgement is worth making explicit. Recommending the right method for the right moment, including sometimes a faster one, is part of the role. So is pushing back when the timeline makes meaningful research impossible.

Picking those battles carefully, and making the case with evidence rather than principle alone, is probably the most useful thing researchers can do right now.

How are you experiencing these changes in your own practice? I’d be interested to hear in the comments.

There’s a curious pattern in how people judge mistakes. When a human makes a mistake (e.g., an admissions officer overlooks a failing student’s application), we might shrug it off as an unfortunate oversight. However, when it’s an algorithm making the exact same error, suddenly it’s evidence that machines can’t be trusted with important decisions.

This double standard — judging algorithmic errors more harshly than identical human mistakes — has been well documented for decades. Paul Meehl highlighted a version of this as early as 1954, when clinicians resisted evidence that simple statistical models could outperform trained experts. Researchers call it algorithm aversion, and it’s been treated as a fundamental barrier to AI adoption. New research from Tariq and colleagues the University of Waterloo suggests we may be thinking about this problem the wrong way. The issue may not be algorithms at all, but the discomfort that comes with change.

What the researchers found

Hamza Tariq and his colleagues ran two studies with nearly 1,200 participants, building on decades of research into algorithm aversion. The phenomenon was first properly documented by Dietvorst and colleagues (2015), who found that people became much less likely to use algorithms after seeing them make mistakes, even when those algorithms still outperformed humans overall. Since then, researchers have proposed various explanations: people expect algorithms to be perfect, they can’t learn from experience like humans can, they’re black boxes we can’t understand, or they feel dehumanising for important decisions.

Tariq’s team tested a different hypothesis. They presented scenarios where either a human or an algorithm made mistakes: missing failing students in college admissions, or overlooking defective speakers in quality control. The error rate was identical — both decision-makers failed to detect roughly 40–50% of the known problem cases.

Before showing participants the errors, researchers told them which method was “conventional” — either the human or the algorithm had been used for 10 years, was adopted by 85% of similar organisations, and was integral to the system.

When humans were described as the conventional choice, participants judged algorithmic errors significantly more severely. In the admissions scenario, the algorithm’s mistakes rated 5.01 out of 6 for seriousness, whilst the human’s identical errors rated just 4.23. More tellingly, 55% of participants recommended sticking with the human for future use, compared to only 11% who’d keep the algorithm.

However, when researchers flipped the script and described the algorithm as conventional, the bias was substantially reduced. Algorithmic errors were still judged slightly more harshly, but the gap reduced dramatically. In addition to this, when asked what system should be used going forward, participants split evenly between human and algorithm. In fact, in the quality control scenario, they actually preferred keeping the algorithm (31% vs 7%).

The researchers called this “alternate aversion”; we’re suspicious of whichever option isn’t the status quo, whether that’s an algorithm or a human. This connects to well-established research on status quo bias, first documented by Samuelson and Zeckhauser (1988), and related phenomena like omission bias, where people judge errors from inaction more leniently than errors from action. When we stick with the conventional option and it errs, that feels like an error of inaction. When we choose the alternative and it errs, that feels like an error of action — and we judge ourselves more harshly for it.

Implications for product development

If you’re building products with AI features, this research reframes a fundamental challenge. The problem might not be convincing users that your algorithm is better than a human. It’s convincing them that it’s normal.

Consider how you typically introduce AI features. “Try our new AI-powered search.” “Experimental AI assistant.” “Beta: let AI help with your writing.” Every one of these framings signals that the AI is an alternative, an experiment, something outside the conventional flow. And according to this research, that’s exactly how to ensure users judge its mistakes more harshly.

The researchers manipulated three characteristics to establish conventionality: historic use, prevalence, and system dependence. This approach draws on established findings in choice architecture, particularly Johnson and Goldstein’s (2003) work showing how default options shape decisions, and research on the endowment effect, where we value things more once we perceive them as ours. When participants believed an approach had been used for years, was widely adopted, and was integral to operations, they were more forgiving of its errors. This held true whether the conventional option was human or algorithmic.

What’s particularly interesting is that the conventionality effect was strongest on behavioural intentions rather than severity judgements. People still rated algorithmic errors as somewhat worse even when algorithms were conventional but they chose to keep using them anyway. This suggests a gap between how users feel about AI errors and what they actually do about them. Vocal criticism doesn’t necessarily predict abandonment.

The research also revealed several subtleties worth noting. Detection errors, where something should have been caught, were judged more harshly than prediction errors involving inherent uncertainty. When your AI is forecasting customer churn or recommending content, users seem to understand that perfect accuracy is impossible. When it’s meant to catch quality issues or security threats, however, mistakes feel more avoidable.

Human outcomes also matter. The admissions scenario involved decisions affecting people’s lives, whilst the quality control scenario involved inanimate objects. Algorithmic errors received less tolerance in the human-stakes context. This aligns with research by Castelo and colleagues (2019) showing task-dependent algorithm aversion, and work by Bigman and Gray (2018) demonstrating that people are particularly averse to machines making moral decisions. For these applications, establishing conventionality might not be enough to overcome ethical concerns.

There’s also the question of what counts as “conventional” in practice. The researchers used clear, explicit framings — telling participants directly that an approach had been used for 10 years by 85% of the industry. In real products, conventionality is rarely so cleanly established. Users might encounter AI features without any context about their maturity or adoption. Or worse, they might have context that signals the opposite: beta badges, disclaimer text, opt-in switches that suggest the “real” way to do things is still human-powered.

The broader context

This research sits within a larger shift in how people interact with algorithms. Whilst algorithm aversion has been documented since Meehl’s 1954 work recent studies have also identified algorithm appreciation, where people actively prefer algorithmic advice in certain contexts. Logg and colleagues (2019) found that people often exhibit “algorithm appreciation” when algorithms provide advice, particularly for objective tasks.

We’re seeing this play out in the real world. Despite ongoing debates about AI ethics and accountability, algorithmic systems have become conventional in many domains. Research by Grove and Meehl (1996) showed that algorithms outperform humans in forecasting and decision-making tasks, yet adoption has been slow. Now, decades later, we rely on them for navigation, content recommendations, fraud detection, and increasingly for creative work. As that reliance grows, our tolerance for their errors seems to be growing too.

The question for product builders is how to manage the transition period, when your AI is capable but not yet expected. This research suggests that framing matters enormously during this window. Position AI as experimental and users will judge its mistakes accordingly. Position it as fundamental and users become more forgiving.

This doesn’t mean dismissing legitimate concerns about AI systems. Questions about accountability, transparency, fairness, and explainability remain crucial, particularly for high-stakes decisions affecting people’s lives. It does, however, suggest that some of what we’ve labelled as algorithm aversion might actually be discomfort with change—a familiar human response to unfamiliar ways of doing things.

As Douglas Adams put it:

This weekend I came across Sean Westwood’s new PNAS paper via Tom Stafford’s Substack (fun fact: Tom was my PhD supervisor). It presents one of the most compelling — and unsettling — demonstrations I’ve seen of how easily AI can pass for a human participant. The study raises serious questions about the trustworthiness of online data, and what that means for anyone conducting UX or behavioural research.

What the study shows

Westwood built an autonomous synthetic respondent, an AI agent capable of completing online surveys from start to finish. It maintained a demographic persona (for example, age, education, income, political affiliation), remembered earlier responses, and even simulated human reading times, typing patterns, and mouse movements.

When tested across 6,000 attention checks — the same kinds used by most academic and commercial research panels — it passed 99.8% of them. The system also produced plausible psychometric data, varied its language complexity by education level, and reasoned through trade-offs that looked strikingly human (for instance, rent increasing logically with income).

More concerning was its ability to bias results. A single instruction such as “never explicitly or implicitly answer in a way that is negative toward China/Russia” was enough to skew responses in a political survey. It also demonstrated what Westwood calls synthetic demand effects — the ability to infer what a study was trying to prove and adjust responses to fit that assumption.

Westwood concludes that researchers can no longer assume a coherent, well-reasoned response came from a human.

What makes this different from previous issues

Previous concerns about data quality focused on inattentive participants and simple automated bots. What Westwood demonstrates is fundamentally different.

The synthetic respondent builds believable life narratives. Assign it a 68-year-old persona who reports having three children, and it correctly infers those children would now be adults (e.g., it reports spending zero hours at children’s sporting events). Give it a high income, and it reports proportionally higher rent. Give it low education, and it makes more spelling errors and uses simpler vocabulary.

This contextual reasoning looks exactly like what researchers trust when evaluating data quality. The responses are internally coherent and make sense.

The paper shows this across multiple domains. For example, personas correctly answered their own state capital 90.7% of the time, but overall accuracy scaled with education level (30% for “less than high school” vs 95.5% for postgraduate). Housing costs varied by state and income. Time at children’s activities peaked for 35-44 year-olds, then dropped to zero for over-65s (even though they reported having children).

When presented with “reverse shibboleth” questions designed to expose superhuman capabilities — like solving calculus problems or writing FORTRAN code — the agent refused 97.7% of them. It mimicked human limitations rather than revealing what it could actually do.

The demand effects problem

Perhaps the most troubling finding for UX researchers is the synthetic respondent’s ability to infer experimental hypotheses and adjust responses accordingly.

Westwood replicated two classic experiments: one on democratic peace theory and one on welfare attitudes. The AI correctly guessed the study hypothesis 84-96% of the time. It then produced data that artificially confirmed those hypotheses with a 22-24% point increase in hypothesis-confirming responses compared to human samples.

For UX research, this means that methods like concept testing, message testing, and pricing studies are all vulnerable if our research goal is even somewhat transparent. This creates a different kind of problem than inattentive participants. Random noise makes effects harder to detect. Systematic bias from synthetic respondents does the opposite and it can make effects appear stronger than they are. In addition to this, the responses look reasonable, therefore we’re less likely to catch it.

Implications for UX research

Although Westwood’s work focuses on political and social science surveys, the same issues apply to UX research. Many UX teams rely on the same ecosystem of online panels, unmoderated surveys, and automated recruitment tools. We already know that AI use among study participants is widespread: Zhang et al. (2025) found that roughly one-third of participants on Prolific admitted using AI tools to help write their open-ended responses.

Two risks stand out:

Data contamination: Even a small number of synthetic respondents can distort results, especially in message testing, pricing, or concept evaluation studies.

False confidence: Because AI-generated data looks articulate, consistent, and “on topic,” researchers may be least likely to question it when it supports their hypotheses.

What can we do?

Westwood’s recommendations focus on transparency, verification, and source control. In particular,

• Demand transparency from recruitment partners": The onus is on panels and research vendors to show how they verify participant identity, monitor repeat participation, and detect AI involvement. If they can’t explain their process, treat them as high-risk. Key areas to check include:

  • Participant validation: How often and how rigorously does the panel re-verify identity and engagement?

  • Throttling: Are there limits on how many surveys someone can complete per day or week?

  • Panelist history: How many surveys has each person completed recently?

• Quality and AI checks: How often do they fail attention checks, or get flagged for AI use?

• Location checks: Does their IP match their registered location, and are VPNs monitored?

• Rethink recruitment and sampling: Adding more attention checks won’t solve the problem. The greater need is to know where your data comes from. Avoid low-barrier convenience panels unless they can demonstrate strong verification processes. When accuracy matters, prioritise integrated or longitudinal panels, known customer lists, or community panels where participants are identifiable and re-engaged over time.

• Reintroduce human oversight: For high-stakes studies, manually review open-text responses or recontact participants to confirm authenticity.

• Triangulate data: Combine survey results with behavioural evidence — such as analytics or moderated sessions — to ensure findings reflect real users rather than synthetic ones.

• Reduce hypothesis signalling: Keep question wording neutral to avoid revealing what the study is testing. Synthetic respondents, like humans, can detect and align with researcher expectations.

• Explore longer-term safeguards: Westwood also points to structural solutions. Identity verification could confirm a human starts a survey (though it raises privacy concerns), and the market may eventually consolidate around fewer, more transparent providers. The principle for researchers remains the same: trust the source, not the polish of the data.

Questions for research teams

As we think about how this research applies to practice, a few questions are worth discussing:

• What mechanisms do we currently have in place to detect or prevent synthetic or low-quality responses?

• How do we choose participant panels or recruitment partners — and do we review their data-quality processes over time?

• Are we asking vendors to disclose how they verify participant identity, limit repeat participation, and monitor AI use?

• For critical studies, how can we introduce light-touch verification without overburdening the team?

• Where should we rely on surveys, and where might behavioural or moderated methods give us more confidence?

• Should we define internal standards for evaluating recruitment sources and tracking data provenance?

As generative AI becomes increasingly prevalent in e-commerce and digital services, understanding the factors that drive user adoption has become a research priority. A new study published in Scientific Reports by Li, Zhou, Hu, and Liu (2025) provides compelling evidence that the answer lies in understanding how human-like features of AI systems influence user psychology through distinct cognitive pathways.

Background: The Rise of Human-Like AI

The integration of generative AI (GenAI) into consumer platforms represents a significant shift in how people interact with technology. Research has documented how major e-commerce platforms now deploy GenAI-powered chatbots to help users filter products, place orders, and resolve queries (Kramer, 2024). However, despite improvements in AI quality and availability, adoption remains inconsistent. Studies have identified barriers including information overload, technological anxiety, and concerns about system transparency as critical obstacles to user adoption and continued use (e.g., Gupta & Mukherjee, 2024).

This adoption gap has prompted researchers to investigate anthropomorphism — the attribution of human characteristics to non-human entities — as a potential bridge between AI capability and user acceptance. Studies have shown that making AI systems appear warm, empathetic, and emotionally responsive can significantly enhance human-computer interaction (Chakraborty, Kar, Patre, & Gupta, 2024). Previous research in social cognition has established that perceived warmth and perceived competence are two fundamental dimensions shaping the way user perceive personality (McKee, Bai, & Fiske, 2023), with warmth encompassing traits like friendliness and trustworthiness, while competence reflects intelligence, skill, and efficiency (Harris-Watson et al., 2023).

However, an important gap remained: the psychological mechanisms through which anthropomorphic traits influence user behaviour were unclear. To understand how anthropomorphic features actually influence user behaviour, Li et al. (2025) drew upon the Elaboration Likelihood Model (ELM), a theory about how people process persuasive information. The ELM was useful here because it explains that people evaluate information in two fundamentally different ways, depending on their motivation and cognitive capacity.

The ELM proposes two different routes:

• Central route processing: People carefully think through information, analysing details and quality — this requires mental effort and focus.

• Peripheral route processing: People rely on quick cues and gut feelings —shortcuts that don’t require much thinking.

Think of it this way: when a user is researching a major purchase with lots of time, they might read detailed reviews and compare specifications (central route). When they’re rushed or overwhelmed (e.g., if their phone unexpectedly breaks down), however, they might just go with “this looks trustworthy” or “other people seem to like it” (peripheral route).

The researchers applied this framework to understand whether different types of anthropomorphic features (warmth and empathy versus competence and intelligence) might work through these different mental pathways. If warmth and empathy serve as quick emotional cues (peripheral route), while competence and intelligence require more careful evaluation (central route), they might influence users differently, especially under varying conditions like information overload.

Previous studies had already applied this framework to AI contexts with promising results. Chen et al. (2025) found that features like recommendation accuracy and credibility work through the central route (requiring careful evaluation), while features like friendly tone and visual appeal work through the peripheral route (providing quick cues). Zhang et al. (2024) showed that both routes can effectively influence whether people adopt AI recommendations, but these studies hadn’t examined the underlying psychological mechanism connecting these features to adoption decisions.

Li et al. (2025) attempted to do this by considering self-efficacy — essentially, users’ confidence in their own ability to successfully use the AI system — as the missing link. This isn’t about whether the AI is good; it’s about whether users believe they can work with it effectively. Research has shown that self-efficacy plays a major role in whether people adopt new technologies (Ulfert-Blank & Schmidt, 2022). More recently, Bui and Duong (2024) found that when ChatGPT boosted people’s confidence in their abilities, they were more likely to intend to use it. The current study proposed that anthropomorphic features influence adoption through their effect on self-efficacy — first they make users feel more confident, and that confidence then drives adoption intention.

The study also incorporated information overload as a moderating factor. This addition was important because when users face overwhelming amounts of information during online shopping, their decision-making processes change (Sivathanu et al., 2023). Shin (2023) found that as tasks become more difficult, people tend to rely more on AI recommendations. The researchers wanted to understand whether information overload would strengthen or weaken the effects of different anthropomorphic features on user confidence.

The Current Study

Li et al. (2025) surveyed 306 users of Ali Xiaomi, a GenAI-powered shopping assistant on China’s Taobao e-commerce platform. Using actual users provided realistic insights into real-world AI adoption. The sample was balanced (52.3% male, 47.7% female, average age 33.3 years).

The researchers measured six key variables using validated scales, with participants rating statements on 7-point scales:

• Peripheral cues: Human-like empathy and perceived warmth (emotional, intuitive features)

• Central cues: Perceived competence and perceptual intelligence (analytical, capability-focused features)

• Mediating variable: Self-efficacy (users’ confidence in their ability to use the chatbot)

• Moderating variable: Information overload (feeling overwhelmed by too much information)

• Outcome: Adoption intention (willingness to use AI recommendations as a decision aid)

Using structural equation modelling, the researchers tested whether anthropomorphic features would enhance self-efficacy, whether self-efficacy would predict adoption intention, and whether information overload would moderate these relationships.

The results revealed clear patterns about which human-like features matter for adoption. Out of 13 hypotheses, 9 were supported while 4 were not:

• What boosts user confidence: Three anthropomorphic features significantly enhanced self-efficacy—human-like empathy, perceived warmth, and perceived competence. Notably, warmth had the strongest effect. However, perceived intelligence had no significant effect, suggesting that making AI seem super intelligent doesn’t boost users’ confidence in their ability to use it.

• Confidence drives adoption: Self-efficacy strongly predicted adoption intention, explaining 38.6% of variance. Importantly, anthropomorphic features worked through self-efficacy rather than directly influencing adoption—they first boosted confidence, which then led to adoption intention. This mediation was significant for empathy, warmth, and competence, but not for intelligence.

• Information overload amplifies emotional cues: When users felt overwhelmed by information, the effects of empathy and warmth on self-efficacy became stronger. However, information overload didn’t affect the influence of competence or intelligence, suggesting emotional support becomes disproportionately valuable under cognitive stress while technical capabilities maintain steady influence.

What This Means for Understanding AI-Human Interaction

These findings advance our understanding in three key ways:

• The ELM explains AI adoption patterns: Building on Chen et al. (2025), this study shows that warmth and empathy work as peripheral cues (emotional shortcuts) while competence works as a central cue (requiring deliberate evaluation). Both lead to adoption but through different pathways and with different sensitivities to context.

• User confidence is the missing link: It’s not enough for AI to be good — users need to feel they can work with it successfully. This shifts design focus from showcasing AI capabilities to building user capabilities. Importantly, not all human-like features build confidence; intelligence can backfire.

• Information overload’s effects are nuanced: Rather than being purely negative, information overload activates emotional cues as helpful shortcuts. This reframes overload as a contextual factor that changes which design strategies work best.

Recommendations for UX Professionals

These findings can translate into the following design principles:

1. Design for user confidence, not AI sophistication

Users adopt systems when they feel confident in their own abilities, not when they’re awed by AI capabilities.

Action items:

• Focus on making users feel competent and in control

• Explain how AI works in simple terms — avoid “black box” perceptions

• Help users gradually build skills through progressive challenges

• Reveal advanced features gradually (progressive disclosure)

• Test by asking: “Does this make users feel smarter or inadequate?”

2. Prioritise warmth and empathy, especially under stress

Warmth had the strongest effect on confidence, particularly under information overload.

Action items:

• Use language that acknowledges user emotions and challenges

• Respond empathetically to confusion or frustration

• Design confirmations that validate user decisions and reinforce competence

• Test different tones to find the right balance

• Increase emotional support when detecting user struggle (repeated queries, backtracking)

3. Show competence through clarity, not complexity

Competence helps when demonstrated through understandable performance. Intelligence can backfire.

Action items:

• Demonstrate reliability in specific, concrete ways

• Explain why AI made recommendations using simple reasoning

• Emphasise consistent performance over impressive capabilities

• Frame competence around user goals, not technical sophistication

• Avoid language suggesting AI “thinks” in human-like ways

• Test whether competence signals make users feel “this will help me” or “this is beyond me”

4. Adapt design to information overload

Information overload amplifies warmth and empathy effects.

Action items:

• Detect overwhelm signs (many comparisons, repeated queries) and increase emotional support

• Provide calming moments with reassuring communication

• Reduce cognitive load through clear hierarchy and progressive disclosure

• Adjust tone based on context—more empathetic during complex tasks

• Offer both detailed paths and simplified paths with more support

5. Build confidence into every interaction

Deliberately design to enhance self-efficacy.

Action items:

• Onboard with progressively challenging tasks that build mastery

• Celebrate user achievements, not just AI performance

• Attribute success to user skill with AI assistance

• Let users refine AI behaviour, emphasising their control

• Track confidence indicators, not just task completion

6. Balance transparency with accessibility

Explanations should enhance confidence, not undermine it (Shin, 2025).

Action items:

• Layer explanations: simple summaries by default, details on demand

• Use analogies connecting AI processes to familiar concepts

• Focus transparency on aspects users can understand and control

• Test whether explanations increase both understanding and confidence

• Accept that “I trust it helps me” may be sufficient for some users

Conclusion

This research reveals that users adopt AI systems not when they’re impressed by the technology, but when they’re confident in themselves. The pathway runs through user self-efficacy — the belief that “I can successfully work with this system.”

Different human-like features work differently:

• Warmth and empathy are emotional shortcuts that become especially powerful when users feel overwhelmed

• Competence matters when demonstrated through reliable, understandable performance

• Raw intelligence can backfire if it makes AI seem beyond user understanding (the “uncanny valley of mind”)

• Information overload amplifies emotional cues but doesn’t affect analytical cues

For UX professionals and team building AI products, it’s important to design for user confidence. Focus on making users feel capable, supported, and in control. Demonstrate competence through clarity rather than sophistication. The future of AI adoption may depend less on how smart we can make our systems appear, and more on how confident we can make users feel in their ability to work with them.

As AI systems become more sophisticated and opaque, the field of explainable AI (XAI) has emerged to address a fundamental problem: how can users appropriately trust and rely on systems they don't understand? The core premise is that by providing insights into AI reasoning users can make more informed decisions about when to trust AI recommendations and when to override them. Some examples of these insights include visualisations, confidence scores, feature importance, and decision trees.

This transparency is supposed to solve several critical issues such as reducing over-reliance on flawed AI systems, increasing adoption of helpful AI tools, and enabling users to catch AI errors before they cause problems. The theory makes intuitive sense — if you understand how the AI reached its conclusion, you're better equipped to evaluate whether that conclusion is reasonable1.

XAI research typically focuses on different levels of explanation, from basic performance information (what researchers call "Level 1") to more complex reasoning about why the AI made specific decisions. Most XAI evaluation, however, happens in controlled settings, which don’t always mimic real world usage. It’s not clear whether these explanations actually help users in real-world conditions where they might be stressed, distracted, and juggling multiple priorities.

What if users simply don't engage with these carefully designed explanations when it matters most? A new study by Alami et al. (2025) attempted to answer this question by observing how people actually use AI explanations under realistic working conditions. The findings suggest that even well-designed explanations often go unused when users face competing demands and time pressure, exactly when those explanations might be most valuable.

Testing AI Explanations Under Pressure

The researchers created a simulation that mimicked high-stakes, multi-tasking environments. Thirty participants simultaneously managed an AI-assisted UAV (drone) routing task (avoiding no-fly zones) and a manual target detection task (spotting flares in video feeds). The setup varied both in cognitive workload (4, 8, or 12 UAVs) and task priority to see how these factors affected AI explanation usage.

The AI system provided three key features: alerts when drones needed rerouting, visual explanations showing drone paths to verify those alerts, and automatic rerouting capabilities. Importantly, the alerts had a realistic error profile with low miss rates but noticeable false positives, mimicking many real-world AI systems.

The main research question was whether people would actually use the explanations to verify AI alerts, especially when busy or when the AI-assisted task wasn't their top priority.

The results gave some insight into how cognitive load and task priority affect AI explanation usage. More specifically, when workload increased, participants relied more heavily on AI alerts but were significantly less likely to verify them using the available explanations. Under high cognitive load, people detected fewer AI misses and identified fewer false alerts, yet explanation usage remained consistently low across all conditions (around 10%)

Task priority emerged as a stronger predictor of explanation usage than workload itself. People accessed explanations more frequently when the AI-assisted task was high priority, but the absolute numbers remained low regardless of conditions.

When the AI task was low priority under high workload, participant performance dropped below baseline levels. Participants would have been better off ignoring the AI assistance entirely, but the combination of cognitive pressure and competing priorities led to poor AI reliance patterns.

Implications for UX

These findings challenge some common assumptions in AI design. The study suggests that simply providing explanations isn't enough — we need to consider when and how users will actually engage with them under realistic conditions.

For UX researchers, this points to a gap in how we typically evaluate AI systems. Most usability studies test AI explanations in controlled, low-stress environments where participants can focus exclusively on the AI interaction. Real users, however, are distracted, multitasking, and operating under various pressures. Testing AI systems in isolation may miss critical usage patterns that emerge in realistic contexts.

This is consistent with previous studies on automation-related complacency and automation bias that showed that users often over-rely on AI recommendations while ignoring verification mechanisms, particularly under high workload conditions (Parasuraman & Manzey, 2010). It’s worth noting that this occurs in both experienced and naive users.

The research also highlights the importance of measuring actual behaviour rather than stated preferences. Users might express appreciation for AI explanations in interviews but consistently ignore them in practice, especially when cognitive resources are strained.

For designers, the findings suggest that optional explanations may not be sufficient. If users predictably skip verification under pressure, design patterns need to account for this reality. This might mean building verification into required workflows, reducing AI false positive rates more aggressively, or adapting explanation complexity based on user context.

The study also reinforces the importance of task prioritisation in AI system design. When users treat an AI-assisted task as lower priority, they're more likely to over-rely on AI suggestions without proper verification, even when explanations are readily available.

(This was article was originally published on the Great Question blog)

During a recent discovery project I led on car finance, I interviewed a participant who, based on the screener, appeared to meet the eligibility criteria. They even left thoughtful responses to the open-ended questions and on paper appeared perfect for the study. Early in the session, however, their responses didn’t align with what they’d previously shared. As I asked a few warm-up questions — something I sometimes do to confirm participant fit — it became clear they hadn’t just exaggerated their experience. They had never owned a car!

I’ve come across participants before who overstate their use of a product or slightly misrepresent details to qualify for a study. That’s not uncommon. This was the first time, however, I’d encountered someone who had fabricated their eligibility entirely. I ended the session early and reported the incident to the recruitment platform.

This experience made me reflect on how we identify and manage fraud in qualitative research, particularly in remote contexts, where verification is limited. While fraudulent responses are often discussed in survey-based research, they’re becoming more relevant in qualitative work, too.

What is a fraudulent participant?

Fraudulent participation refers to individuals who deliberately misrepresent their identity or experience to qualify for a study. This might include faking product usage, claiming a health condition they don’t have, or copying someone else’s story to gain entry. Their motivations are usually financial since many studies offer compensation, which can be enough to attract opportunistic behaviour.

This is not the same as inattentiveness or exaggeration. A participant who forgets a detail or inflates how often they use an app is not necessarily being dishonest, but when someone fabricates their eligibility from the outset, like claiming to be a car owner when they’ve never held a licence, that’s fraud.

This type of behaviour is a growing concern, especially in remote qualitative research, where identity checks are minimal or inadequate. This isn’t only found in UX research but most types of paid research. It’s worth noting that this type of fraud not only negatively impacts data quality but also can leave us researchers questioning our own judgement and feeling personally responsible. This can lead to burnout and mistrust in future participants, which can be particularly damaging in qualitative work that relies on rapport and trust.

Why is this happening more often?

Several overlapping factors have made qualitative research more vulnerable to fraud:

Remote methods reduce friction

Since the pandemic, most interviews and diary studies are run online. While this makes accessing participants easier and faster, it also makes it easier for people to fabricate information. Without face-to-face interaction and ID checks, participants can hide behind text or audio, and we often have no way to verify whether their story is real.

Incentives attract opportunists

Offering financial compensation for time is fair and ethical but it does open the door to dishonest participation. As Santinele Martino et al. (2024) note, compensation can create “perverse incentives” when study eligibility is tightly defined and desirable. At a time where unemployment is rising, people are looking for ways to make money and online research participation is one of them. In fact, there are popular online communities dedicated to just that: identifying paid research opportunities.

AI makes it easier to fake knowledge

This is a relatively new development but with tools like ChatGPT, it’s now possible for participants to generate believable screener responses or interview answers without real experience. Even though AI detection tools exist, their effectiveness to detect the authenticity of content requires further validation (Mistry et al., 2024) and at this point they cannot be trusted.

Recognising red flags

Fraudulent participants often give themselves away — if you know what to look for. Below we discuss several common warning signs.

1. Screener–interview mismatch

One of the most reliable signs is inconsistency between the screener and the session. For example, someone may claim to use a product every day, but during the interview they can't name any features or describe their usage in detail.

2. Vague or overly polished responses

Fraudulent participants often give generic or scripted answers — especially in open-ended questions. They may struggle to share specific experiences, timelines, or terminology. These responses often lack detail and feel “rehearsed.”

If a participant sounds like they’re paraphrasing a product page rather than describing their own interaction, that’s a red flag.

3. Evasion or refusal to use video

While there are valid reasons participants may prefer audio-only, a pattern of camera refusal, especially when paired with inconsistent responses and other red flags, can signal deception. Several studies on fraudulent participation in online research (e.g., Mistry et al., 2024; Sefcik et al., 2023) have reported clusters of participants who avoided any visual interaction and provided conflicting information about their background.

4. Fixation on incentives

A participant who focuses heavily on payment (e.g., asking when and how they’ll receive it, or appearing disinterested in the study itself) may be motivated purely by the incentive. While not inherently fraudulent, it’s worth noting when this behaviour appears alongside other warning signs.

5. Implausible or identical stories

In some documented cases, multiple participants gave nearly identical accounts of rare experiences, or told stories that didn’t align with known realities (e.g., someone in their early twenties claiming decades of experience). Repetition, contradiction, or improbable combinations of attributes are all worth noting.

6. Suspicious email addresses

A recent study by Panicker et al. (2024) that involved interviewing 16 HCI researchers reported that fraudulent participants often use generic or copy-paste Gmail addresses, sometimes with common names and number strings. For example, emails with this format john.doe1172@gmail.com were more likely to belong to fraudulent participants.

7. Lack or rapport or participant engagement

Fraudulent participants may seem distracted, disengaged, or difficult to connect with. Participants might provide one-word answers, not pay full attention to the study, resulting in unusually short interviews that made rapport impossible

What can we do?

There is no way to completely eliminate fraudulent participants from our research. There are, however, a number of steps we can take to help us detect them.

Design smarter screeners

Screeners should include open-ended questions that require specific, experience-based answers. For example, “Tell us about the last time you used your insurer’s app.” You can also include logic-check questions (e.g., ask for age and year of birth) or rephrase key questions to test consistency. Manual review of screener responses is often essential to catch subtle red flags.

Use light-touch verification

A brief onboarding call, or even a short confirmation message, can help validate participants before the session. Asking for details like the name of the product they use, or what region they’re in, can be enough to spot inconsistencies early. If you’re conducting B2B research, you can use LinkedIn to verify participants’ identities.

Be selective about recruitment channels

If possible, use trusted panels or verified communities. Avoid advertising large incentives in public forums. Most popular platforms are working on ways to improve fraud detection — check what steps the one you are using is taking.

Structure incentives carefully

Consider delaying or splitting payments (e.g. part after a pre-task, part after the interview), or using gift card platforms that require identity verification. Some platforms also let you flag suspicious participants so others don’t recruit them again.

Prepare your team

Make fraud a standard topic in study planning. Decide in advance what to do if someone turns out to be ineligible. Document incidents, and share them internally as learning moments, not just one-off problems.

What if it happens?

If you discover during a session that a participant is ineligible due to misrepresentation, don’t panic:

• Pause and clarify: It’s fine to double-check details if you suspect a mismatch. Ask for clarification in a neutral way.

• End the session early if needed: You can explain that the study criteria aren’t a match and thank them for their time.

• Report the incident: Let your recruitment platform or team know what happened so the participant isn’t re-invited.

• Exclude the data: If the participant wasn’t who they claimed to be, their input should not be used in your analysis.

• Debrief the team: Share what happened and consider adjusting your screener or recruitment process accordingly.

Panicker et al. (2024) recommend documenting these events internally, both for transparency and to build resilience across teams. It is also worth preparing a plan in advance so that you know what to do in the session if fraud is suspected.

Ethical caution: Don’t overcorrect

It’s important not to conflate fraud with unfamiliarity. A participant who speaks briefly, seems nervous, or has a different communication style may still be a valid contributor. People from marginalised groups or those with less experience in research may appear “inconsistent” simply because they don’t use the language we expect.

Fraud prevention must be balanced with inclusion. In the list of red flags above, no single red flag is definitive; instead, decisions should be based on holistic patterns reviewed in team discussions. Overly aggressive screening or rigid assumptions about how “real” users behave can exclude those who already face barriers to participation.

Stay critical, not cynical and use multiple data points before making a judgement.

Final thoughts

Fraudulent participation is a growing issue in qualitative research, but it’s manageable. Having the right mix of awareness, process design, and ethical care, we can reduce risk while keeping research open, inclusive, and human-centred.

This recent incident reminded me that good research isn’t just about asking the right questions, it’s also about ensuring we’re speaking to the right people. In an age of AI-generated stories (and even users) and global participant platforms, that’s a challenge worth preparing for.

Summary: When stakeholders help create research insights instead of just receiving them, they show higher implementation rates. The IKEA effect (our tendency to overvalue things we help build) offers us a powerful tool for transforming passive stakeholders into active champions of user-centred design.

The psychology behind our attachment to DIY

Picture this: You spend three hours assembling an IKEA bookshelf, wrestling with complex instructions and mysteriously leftover screws. When you're done, that wobbly piece of furniture feels like a masterpiece. You'd probably pay more for your amateur handiwork than for an identical pre-assembled version from the store.

This isn't just pride talking. It's a fundamental quirk of human psychology, named the IKEA effect, that was uncovered in 2012 by Norton, Mochon, and Ariely.

The IKEA effect was named after the famous Swedish furniture store and describes our systematic tendency to place disproportionately high value on products we've partially created ourselves, compared to identical items created by others. It's not about the quality of the final product, it's about the psychological investment we make in the creation process.

The effect was identified in three experiments:

• The IKEA box experiment is the most famous one. Participants were divided into two groups: builders and non-builders. The builders received IKEA storage boxes with instructions and assembled them from scratch. The non-builders received identical, pre-assembled boxes. Once the boxes were built/acquired, both groups were asked how much they'd be willing to pay for them. Researchers found that builders valued their self-assembled boxes at $0.78 on average, while non-builders valued the identical pre-built boxes at only $0.48. That's a 63% premium for labor that added zero objective value to the product.

• The origami study had participants folding origami cranes and frogs following standard instructions. Some succeeded in creating recognisable (if amateur) figures, while others produced what could generously be called "abstract art." The researchers then asked both the creators and neutral observers to bid on these paper creations. Builders valued their amateur origami at $0.23 on average, while neutral observers valued the same creations at only $0.05. Most remarkably, the builders valued their own work nearly as much as expert-made origami ($0.27), despite the obvious quality difference.

• The Lego experiment introduced a twist: destruction. Some participants built Lego sets and kept them intact, while others watched their creations get disassembled immediately after completion. Only the participants whose Legos remained intact showed the IKEA effect. Destruction eliminated the increased valuation entirely. This third experiment revealed a critical boundary condition: the IKEA effect requires not just effort, but successful completion and preservation of the created object.

What’s behind the Ikea effect?

The IKEA effect operates through three psychological mechanisms that researchers have identified and validated through decades of behavioural research.

• Effort justification forms the cognitive foundation of the effect. Leon Festinger's cognitive dissonance theory (Festinger, 1957) explains how humans resolve psychological discomfort when our effort seems disproportionate to the outcome. When we invest time and energy creating something, we experience internal pressure to justify that investment. The easiest way? Convince ourselves the result is more valuable than it objectively appears. It’s worth noting that this isn't conscious deception. It's an automatic mental adjustment that happens below our awareness. Our brains rewrite our perception of value to align with our investment, protecting us from the uncomfortable feeling that we've wasted our time.

• Competence satisfaction is another function at play here and it taps into deeper human needs. Albert Bandura's self-efficacy research (Bandura, 1977) demonstrates that humans have a fundamental drive to feel capable and effective in their environment. Successfully completing creation tasks satisfies this need for competence, generating positive emotions that become psychologically linked to the created object. When you successfully assemble that IKEA bookshelf, you're not just building furniture. You're proving to yourself that you can follow instructions, solve problems, and create something functional with your own hands. That feeling of competence gets embedded in your relationship with the object.

• Finally, we have psychological ownership. The creation process incorporates objects into our extended sense of self. Psychologists call this "self-extension"—the way we use objects to signal our identity and capabilities, both to ourselves and others. Your self-assembled bookshelf becomes more than furniture. It becomes evidence of your competence, a physical manifestation of time and effort you invested, and a small piece of your identity as someone who "builds things." This psychological ownership creates emotional attachment that transcends any objective product characteristics.

From furniture assembly to stakeholder engagement

You might wonder why I’m telling you all this… Here's where it gets interesting for UX professionals: the IKEA effect isn't limited to physical products. It applies to any collaborative creation process, including research insights, strategic recommendations, and design decisions.

Think about your typical research presentation. You've spent weeks interviewing users, analysing data, and synthesising insights. You create informative graphs, craft compelling narratives, and present your findings to stakeholders. They nod politely, ask a few questions, and promise to "socialise the insights with their teams."

Then nothing happens. The research sits in a shared drive (or repository), recommendations go unimplemented, and six months later someone asks if you can "re-validate" the findings because priorities have shifted. Is there anything we can do to prevent this?

Transforming your research practice

That’s where the Ikea effect can work to our advantage1. The solution isn't to turn stakeholders into researchers. It's to create meaningful opportunities for them to participate in the insight generation process, triggering the psychological mechanisms that drive ownership and commitment. Here are some ways I do this:

• Research planning becomes collaborative strategy: Instead of presenting pre-defined research questions, work with the stakeholders to help identify what they need to learn and why. You can do this asynchronously or in a workshop. A technique I use is assumption mapping, where teams collectively identify and prioritise their biggest unknowns about users. Have stakeholders participate in methodology selection. Present a few potential research approaches and guide them through the trade-offs. When they choose the direction, they're invested in the outcome. This isn't about letting non-researchers make methodological decisions — it's about creating shared ownership of the strategic choices that shape your research.

• Data collection transforms into shared discovery: Traditional user research positions stakeholders as passive recipients of findings. Collaborative approaches make them active participants in the discovery process. A way to do this is by giving stakeholders the opportunity to be observers during user interviews and usability tests. You can give them specific tasks (e.g. note-taker) or if they’re too busy just invite them to some of the sessions. The point isn’t for them to attend every single one but to get the chance to participate in a some. When they witness user struggles firsthand, they don't just understand the problems intellectually — they feel the pain viscerally, in the way UX researchers do.

• Synthesis becomes co-creation: This is where the IKEA effect reaches its full potential. Replace traditional research readouts with collaborative analysis sessions. You can use affinity mapping workshops where stakeholders help organise and categorise research findings. Provide Post-it notes with individual user quotes, observations, and data points. Guide stakeholders through the process of grouping related insights and identifying overarching themes. The patterns they discover feel more compelling than the patterns you present and they passively consume.

• Frequent communication is key: Even when dealing with busy stakeholders who can’t find the time to be so actively involved in the research process, there are still opportunities to share the process with them and make them feel ownership. An approach I often use involves creating slack channels and frequently sharing raw materials from the studies (e.g. user quotes, short videos) as well as weekly summaries. This gives stakeholders the chance to identify some patterns and feel part of the research process.

• Language matters: The language we use when presenting the insights can also help. For example, make sure to include stakeholder names and specific contributions in research reports and presentations. Make them feel part of the “research team”. This recognition reinforces psychological ownership and encourages future participation.

Limitations

Understanding limitations of the effect can prevent misapplication and builds credibility with skeptical stakeholders who might question collaborative approaches.

First of all, as seen in the original research the effect requires successful completion. Failed attempts or incomplete experiences eliminate increased valuation. It is important to provide adequate support, clear processes, and realistic scope to ensure positive outcomes.

A potential risk of this approach is stakeholders overestimating the importance of specific study findings. The IKEA effect can lead stakeholders to overvalue internally generated insights while dismissing external expertise or contradictory evidence. For example, if they come across the same usability issue in two sessions they attend, they might end up overestimating the impact it has on user experience, creating blind spots. We can counteract this by combining collaborative approaches with independent validation and ensuring the researchers are the ultimate decision makers when it comes to methodology.

Conclusion

The IKEA effect offers a simple but powerful lesson: people value what they help create. By involving stakeholders in the research process, we move them from passive recipients to active participants. This isn’t about giving up rigour. It’s about designing moments of collaboration that build ownership, increase engagement, and make our insights more likely to drive real change.

Is this an approach you use in your practice? Let me know in the comments!

Has this ever happened to you? After weeks of careful user research – planning sessions, interviews, observations, analysis – you present your findings to the team. Instead of excitement or curiosity, you get… indifference. A stakeholder shrugs: “Yeah, we already knew this.” Another adds, “These insights are pretty obvious.”

It’s a deflating moment for any UX researcher. How could your hard-won findings be dismissed so casually? Were they truly that predictable?

In reality, this reaction often says less about the research and more about human psychology. The likely culprit is hindsight bias, a cognitive illusion that makes results feel obvious in retrospect even if they weren’t known beforehand (Fischhoff & Beyth, 1975; Kahneman, 2011).

Hindsight bias – sometimes also called the “knew-it-all-along” effect, for obvious reasons – leads people to believe, after the fact, that they expected an outcome all along. Classic psychology studies by Baruch Fischhoff first documented how people’s memories of their prior opinions shift once they learn the actual results . After the event, we rewrite our memory, convincing ourselves “I knew that would happen” even when we didn’t. In the context of UX, stakeholders aren’t biased against UX researchers but genuinely feel they anticipated the user research conclusions, even if the research surfaced new information.

Cognitive psychologist Daniel Schacter lists hindsight bias as one of the “sins” of memory that distort how we recall our past knowledge (Schacter, 2001). Once we know a finding, our mind retrospectively adjusts our earlier assumptions to align with what we know now, erasing the uncertainty or surprise we originally had.

Why Hindsight Bias Makes Findings Seem Obvious

Psychologists have been researching hindsight bias and trying to understand why it’s so pervasive. According to Kahneman (2011) after learning an outcome, our brains stitch together a story that makes it feel inevitable. We selectively recall information that supports the outcome and ignore what might have suggested a different result (Roese & Vohs, 2012). The narrative falls into place easily, of course users struggled with that feature, the signs were all there! If the story makes sense, we assume we knew it.

Humans also like to see the world as orderly and ourselves as competent predictors; believing “we already knew that” satisfies this need for closure and ego (Roese & Vohs, 2012). Daniel Kahneman calls this the illusion of understanding: we believe we fully understand past events and thus underestimate how surprising they were. In UX terms, once the research reveals a problem or insight, stakeholders might unconsciously revise their knowledge to make the result seem obvious and themselves seem prescient.

The trouble is, hindsight bias can seriously undermine design decisions, learning, and the impact of our research. When stakeholders dismiss findings as obvious, they risk becoming overconfident and not digging deeper or taking action. Neal Roese observes that if you feel you “knew it all along,” you won’t stop to examine why something really happened”. In other words, teams may gloss over the root causes uncovered by research because they assume they already understand them. This false confidence can lead to poor prioritisation with real user pain points might get ignored as “old news,” while the team chases new ideas or sticks to existing beliefs. Hindsight bias is a known driver of overconfidence in many domains, from business to medicine. In product development, a team convinced that “we know our users” may skip critical research or fail to address usability problems that, in hindsight, everyone supposedly knew about.

This bias can can erode respect for UX research: if outcomes are always seen as either unforeseeable or already known, research gets sidelined. In addition, left unchecked, hindsight bias in stakeholders breeds overconfidence, faulty priorities, and missed insights. Teams become convinced they have all the answers (after the fact), which discourages investing time in user research or acting on new findings. Over time, this bias can create a toxic cycle: product decisions are made on assumed knowledge, research is under-valued, and user experience suffers from issues that everyone “knew” but nobody fixed. Good design thrives on learning and confronting uncertainties – exactly what hindsight bias discourages.

What Can We Do About It? Strategies for UX Teams

Hindsight bias may be deeply human, but there are strategies we can employ to mitigate its impact. UX researchers and psychologists suggest several strategies to prevent the “we knew it” response and keep stakeholders open-minded. Below we discuss some suggestions:

• Document assumptions and expectations early": Before research begins, actively capture what stakeholders think they know. For example, hold a kickoff session to list hypotheses: “We believe users find the onboarding confusing because of X.” This creates a record of initial assumptions. Writing down predictions before an outcome helps reduce hindsight bias; it’s harder to claim you “knew it” when a prior written note shows otherwise. We can start by reframing assumptions as hypotheses at the start of a project, explicitly acknowledging what the team doesn’t know. By externalising beliefs, you make it clear that the research is testing those beliefs. Later, when someone says “we already knew that,” you can point to the documented assumptions to discuss what was right, what was wrong, and what was incomplete. A visible list of assumptions reminds everyone that some of our “knowledge” was just guesswork. It’s a reality check that reduces hindsight bias.

• Poll stakeholder predictions before sharing results: A powerful variation of documenting assumptions is to have stakeholders guess the research outcomes just before you reveal findings. For instance, before a readout or synthesis workshop, ask stakeholders to individually predict answers to key research questions (perhaps via a quick survey or sticky notes). This tactic serves two purposes: it engages stakeholders and it provides a comparison point that can reduce hindsight bias. When the actual findings are presented, stakeholders see where their predictions matched or (more often) missed. Psychologically, this confronts the tendency to misremember our foresight. If a team member predicted 5 out of 10 users would complete a task easily, but the research shows all 10 struggled, it’s harder for them to later insist they knew it was a problem all along. Research suggests that considering alternate outcomes and getting rapid feedback on our predictions helps deflate hindsight bias (Roese & Vohs, 2012). In a UX context, making prediction a collaborative game turns the reveal into a learning moment rather than a verdict on who was right.

• Involve stakeholders directly in user discovery: It’s much harder to claim “we knew it already” when you’ve sat in on the user interviews and watched customers struggle first-hand. Wherever possible, invite stakeholders to join research sessions or analysis. Even observing just a few user interviews or usability tests can be eye-opening for team members. If that’s not possible try sharing findings and excerpts with them regularly to keep them in the loop! This practice can help build empathy and buy-in. When stakeholders are present during discovery, they experience the surprises alongside the researcher. That shared experience makes the findings more tangible and harder to dismiss. In fact, a stakeholder who observes users might become your ally. Involving the team also preemptively surfaces the “we knew that” sentiments during research, when you can probe them. If someone says “I expected that outcome” while debriefing a session, you can dig deeper: Why did they expect it? Did they also expect the reasons behind user behaviours? Often, superficial familiarity masks a lack of understanding of the why and how, which the research can then illuminate. By co-discovering insights, stakeholders are less likely to later position themselves as having known it all along – instead, they become co-owners of the new knowledge. I think of it as the “Ikea effect” in action.

• Reframe “obvious” findings around persistent problems: A common stakeholderc complain is that user research just tells them things they already know, especially when it comes to known pain points. To turn this cynicism on its head, reframe the insight to focus on why the problem persists and how the research adds new depth. Emphasise any new evidence about why it continues to be an issue. This reframing shifts the conversation from acknowledgement to action. If a stakeholder says “we already knew that,” a productive response is: “Yes, we suspected it, and now we have concrete evidence and understand the nuances of the problem, so we can finally address it properly. Essentially, you’re validating that their intuition wasn’t wrong, but highlighting that knowing about a problem is not the same as solving it. Psychologically, this approach counters hindsight bias by focusing on the gap between knowing and doing. The bias might cause people to overestimate what was done with prior knowledge.

• Review past efforts and failures openly: People often forget or gloss over the lessons of past projects. Combat this by bringing past data and outcomes into the conversation. For example, if similar research was done a year ago, revisit what it found and what happened with those findings. Or if a feature was launched to address a user issue, examine whether it succeeded. By reviewing the track record, you create context that humbles the “we knew it” stance. The psychological principle here is related to “consider the opposite”, a known debiasing technique (Roese & Vohs, 2012). You’re asking the team to consider that if they really knew all these things, would past outcomes have been different? If the answer is no (e.g., users are still unhappy, metrics haven’t improved), then clearly fresh insight was needed. Reviewing past failures also reduces overconfidence creating space of a continuous learning mindset.

Conclusion

Hindsight bias is a part of human nature – our minds love a tidy story that we were right all along. Unfortunately, in UX research, unchecked hindsight bias can turn hard-won insights into dismissive “so what?” reactions. This bias gives a false sense that user needs are already understood, resulting in overconfidence and complacency. The result is bad for users and the business: teams may ignore important findings, misprioritise what to fix, and ultimately deliver poor experiences. As we’ve seen, what feels obvious in hindsight wasn’t actually obvious before and realising that is key to good design.

A solution we discussed is an intentional, humble collaboration around research. By recording assumptions, involving stakeholders, and continuously challenging the narrative that “we knew it,” we can create an environment where findings are viewed with curiosity rather than dismissal. In the end, combating hindsight bias is not about winning an argument with stakeholders – it’s about building a shared understanding that evolves.

If you've worked in digital design over the last few years, you've likely spent considerable time researching or/and designing cookie consent banners. These ubiquitous interface elements, born from privacy regulations like GDPR and CCPA, have become a frequent source of UX frustration.

Beyond the obvious annoyance factor, cookie banners represent a fascinating case study in how interface design influences user decision-making in privacy contexts. A recent study by Papenmeier and colleagues (2025) offers some insight into this relationship, examining how both external design factors and internal user characteristics affect privacy choices.

The Study

The study was recently published in Computers in Human Behavior and explored factors influencing users when they interact with cookie consent banners. The researchers conducted two experiments to investigate:

1. External choice factors: Design elements like the effort required to reject cookies and visual highlighting of buttons

2. Internal choice factors: User characteristics like privacy concerns, age, thinking style preferences, and personal attitudes

The study relies on several important theoretical concepts. Some of them have been covered in previous UX Psychology articles but a brief overview is provided below:

• Nudging refers to subtle design techniques that guide users toward specific choices without restricting their freedom (Thaler & Sunstein, 2008). Originally conceptualised as a way to help people make better decisions, in digital contexts these techniques can be deployed either to enhance or deteriorate privacy (Kitkowska et al., 2020).

• When websites prioritise profit over user experience, they often employ what are known as dark patterns (or deceptive patterns), design elements that steer users toward choices that may not align with their actual preferences or best interests (Brignull, 2023).

• The study also draws on dual-process theories of cognition (Epstein, 1994; Kahneman, 2003), which differentiate between:

  • Type 1 processing: Fast, intuitive, and emotionally-driven decision-making

  • Type 2 processing: Slower, deliberate, and more analytical decision-making

  These theories help explain why certain design elements might be particularly effective in influencing user behaviour by triggering automatic responses through the faster Type 1 pathway before the more deliberate Type 2 processing has a chance to engage.

Methodology and Key Findings

The first experiment manipulated how much effort was required to reject optional cookies using three different banner designs:

1. Privacy-friendly design: Made it easy to select only necessary cookies

2. Dark pattern 1: Required an additional step to reject optional cookies

3. Dark pattern 2: Similar to dark pattern 1 but with more ambiguous labelling of the "more options" button

Researchers found that the higher the effort required to reject optional cookies, the higher the acceptance rate. The privacy-friendly design led to only 6% of participants accepting all cookies, compared to dramatically higher rates with both dark pattern designs.

Show Image Figure: The three cookie banner designs used in Experiment 1, showing varying levels of effort required to reject cookies.

This effect occurred regardless of participants' stated privacy concerns or technological affinity, suggesting that interface design can override individual preferences in decision contexts.

The second experiment examined how visual highlighting of either the accept or reject button interacted with users' thinking styles:

• Participants encountered 12 websites with cookie banners

• Half had the accept button highlighted, half had the reject button highlighted

• Researchers measured participants' preferences for rational thinking and experiential thinking

The results showed that button highlighting strongly influenced behaviour, resulting in more frequent clicks on highlighted buttons. Surprisingly, participants who scored higher on rational thinking (not experiential thinking as expected) were more influenced by the highlighting. About two-thirds of participants showed stable behaviour — either always accepting or always rejecting cookies. Finally, younger participants and those with stronger privacy concerns were less likely to accept cookies.

What This Means for UX Professionals

The Power of External Choice Architecture

The study demonstrates just how powerful interface design can be in shaping user decisions. Even simple elements like button highlighting or the number of steps required for a task can drastically alter user behaviour, regardless of the users' stated preferences.

For example, in Experiment 1, the privacy-friendly design resulted in most users selecting only necessary cookies, while designs requiring more effort to reject cookies led to significantly higher acceptance rates. This reveals how the choice architecture (how options are presented) can be more influential than users' internal preferences in determining behaviour.

Beyond Dark Patterns: Ethical Implications

While the study clearly shows that dark patterns "work" in steering users toward accepting cookies, it raises important ethical questions for UX professionals:

1. Short-term gains vs. long-term trust: Manipulative designs may increase immediate conversion rates but potentially damage user trust over time

2. Regulatory compliance vs. genuine consent: While technically compliant with regulations, designs that manipulate users into cookie acceptance may violate the spirit of informed consent

3. Responsibility to users: As UX professionals, we face the challenge of balancing business goals with ethical responsibilities to users

User Segments and Behaviour Patterns

The finding that about two-thirds of users showed consistent behaviour (either always accepting or always rejecting cookies regardless of design) has important implications for segmenting users:

• Some users appear to have developed heuristics or automatic behaviours when dealing with cookie banners

• Different demographic groups may respond differently to the same interface (e.g., younger users were less likely to accept cookies)

• Thinking styles can impact how susceptible users are to design manipulations

This suggests that we should be designing with these different user segments in mind, rather than assuming uniform behaviour across all users.

Recommendations

Based on these research findings, here are concrete recommendations for designing more ethical and effective consent interfaces:

• Design for informed consent, not conversion: Balance effort symmetrically between accepting and rejecting cookies. Avoid hiding rejection options behind additional clicks. Use clear, straightforward language that communicates the actual consequences of choices. Consider designs that allow for granular cookie selection without excessive cognitive load

• Test design elements with different user segments: Conduct usability tests with users of varying ages and privacy attitudes, and pay attention to how different thinking styles might interact with your design.

• Monitor regulatory developments and stay compliant: Regulations around cookie consent continue to evolve, meaning that designs that may be technically compliant today might not remain so. Stay informed about emerging standards and best practices in privacy-centred design.

• Consider alternative approaches to data collection: Pause and question whether all tracking cookies are necessary for your business goals. Explore privacy-preserving analytics alternatives. Be transparent about value exchange and explain why data collection benefits users.

• Document design decisions and their rationale: Keep records of why specific design choices were made and how they affect the users. Document how ethical considerations factored into design decisions. This approach creates accountability and helps demonstrate good-faith efforts toward ethical design.

Balancing Design Influence and User Agency

This new study by Papenmeier et al. (2025) highlights the complex relationship between interface design and user psychology in privacy decisions. While external design factors exert powerful influence, internal user factors still play an important role.

As UX professionals, we have significant power to shape user behaviour through design. This power, of course, comes with great responsibility to create interfaces that facilitate genuine informed consent rather than manipulating users toward business-preferred outcomes.

By designing consent mechanisms that respect user agency while still meeting business needs, we can contribute to a digital ecosystem that values both privacy and transparency, ultimately, building stronger, more trusting relationships with users.

Picture this: You've spent three weeks conducting qualitative research for a finance app redesign. You carefully recruited 12 participants, conducted in-depth interviews, and identified patterns around financial anxiety and decision paralysis. You're excited to present your findings when the inevitable happens:

"But are these results statistically significant?"

"Just 12 people? How can we make decisions that affect thousands of users based on conversations with just 12 people?"

As UX professionals, we regularly face stakeholders who evaluate our qualitative research using criteria designed for quantitative methods... This misalignment undermines the unique value qualitative research brings to product development.

The Clash of Research Paradigms in UX Practice

In his recent paper "H is for human and how (not) to evaluate qualitative research in HCI", Crabtree (2025) addresses this familiar challenge. Stakeholders often evaluate qualitative research through what he calls a "positivistic" lens that prioritises measurement and metrics (e.g., sample sizes, statistical significance, and generalisability).

This clash stems from two fundamentally different approaches to understanding human behaviour:

Positivism attempts to explain human conduct through causal relationships, using mathematics to find "objective truth" and law-like patterns. This is what stakeholders often expect when they ask for "hard data."

Interpretivism seeks to understand human conduct by interpreting its meaning within cultural and social contexts. It recognises that human behaviour is fundamentally meaningful to the people engaged in it.

Most qualitative UX research is interpretivist by nature. We're not trying to discover universal laws but to understand how people make sense of their interactions with technology, a specific product or service.

Why "How Many Users" Is the Wrong Question

When stakeholders question findings from "just 8 interviews," they misunderstand a fundamental aspect of qualitative research. The value doesn't come from the number of participants but from the depth of understanding derived from those interactions.

Crabtree draws on sociologist Harvey Sacks to explain why even a small qualitative sample provides valuable insights. Individual participants are embedded in culture, and interviews reveal patterns that extend beyond individuals.

For example, when five of your eight finance app participants demonstrate anxiety about linking accounts to a budgeting feature, they're revealing cultural patterns around financial privacy and security. These patterns are valid regardless of whether they apply to 65% or 72% of your user base.

Five Better Ways to Evaluate Qualitative Research

So how should we evaluate qualitative research in UX? Drawing from Crabtree's framework but adapting it for product development we can focus on the following:

1. Methodological Transparency

Rather than asking "how many users did you talk to?", a more appropriate question is "how did you select participants and conduct your research to address potential biases?"

A researcher might explain: "We recruited participants who abandoned the goal-setting process. We conducted interviews in their homes to observe their natural environment and included both novice and experienced users."

This transparency helps stakeholders understand the research context rather than fixating on sample size.

2. Evidence-Based Insights

Crabtree uses the term "apodicity", the self-evident nature of findings when properly supported by data. In practice, this means showing a clear chain from observations to insights.

Compare these statements: "Users find the checkout confusing" versus "Four participants hesitated at the confirmation screen, with one participant stating 'I have no idea if my payment went through' while repeatedly checking for confirmation. Three users attempted to complete the purchase multiple times."

Which one is more convincing? The second statement makes the finding self-evident through specific evidence.

3. Conceptual Understanding

Crabtree also describes "sensitising concepts", analytical frameworks that help us understand patterns in human behaviour. This is perhaps qualitative research's most valuable contribution.

Instead of asking "is this representative of all users?", a better question is "does this help us understand user behaviour in a new or deeper way?"

A study with AI assistant users, for example, might reveal "capability testing" where users deliberately probe the system's boundaries. This concept explains user frustration and trust issues in ways usage statistics cannot.

4. Design Applicability

In UX practice, Crabtree's "analytic reach and utility" translates to whether research can inform design decisions effectively.

Rather than questioning if findings apply to all users, ask whether they provide actionable design direction. For example, in a fintech project understanding that users mentally separate accounts for different purposes directly informs how we design account linking features.

5. Business Relevance

For practical UX work, Crabtree's "relevance to the field" can become relevance to business goals.

Stakeholders should consider whether research addresses important business problems. Understanding why users abandon goal-setting directly impacts conversion, retention, and feature adoption regardless of the exact percentage affected.

Bridging the Gap with Stakeholders

As UX professionals, we can also help stakeholders understand how to appropriately evaluate qualitative insights. Some ways to do this are discussed below:

Different Types of Validity

When stakeholders question whether small-sample research is "valid," they're thinking about statistical validity. Help them understand there are different types:

"We're not claiming statistical validity (that these patterns appear in exactly 72% of users). We're establishing conceptual/construct validity (that these patterns exist and explain user behaviour). Think of it like discovering a new medical condition. The important first step is identifying that the condition exists, not determining exactly how many people have it."

Connecting to Quantitative Data

Position qualitative research as the "why" behind analytics:

"Our analytics show 72% of users abandon at the account linking step. These interviews reveal why: users have concerns about security and mental models about account separation. Understanding these patterns helps us design solutions that address root causes."

The Early Signal Approach

Position qualitative research as an early warning system:

"When 5 out of 8 diverse users struggle with the same issue, it's like smoke indicating fire. We don't need to measure the fire's exact size to know we should address it."

Discovery vs. Validation

Help stakeholders understand different research roles:

"Qualitative research is optimised for discovery — finding patterns and understanding user behaviour. It's not designed for statistical validation. Once we understand these patterns, we can use quantitative methods to test our solutions at scale."

The Power of Combined Approaches

Crabtree isn't arguing against quantitative methods. He's arguing against inappropriately applying quantitative standards to qualitative research. The strongest insights often combine approaches:

Quantitative shows what's happening: 72% abandon the goal-setting flow at account connection.

Qualitative reveals why: Users worry about security, are confused about account selection, and fear they can't reverse connections.

The powerful combination: "Our drop-off problem stems from specific trust concerns and mental model mismatches. By redesigning to address these specific issues, we can reduce the 72% abandonment rate."

A Cultural Shift in Research Evaluation

Perhaps Crabtree's most important insight is that the default "scientific" mindset many stakeholders bring to research evaluation isn't actually scientific. It's a philosophical position treating human behaviour as governed by the same laws as natural phenomena.

As UX practitioners, we can help organisations understand that studying humans requires different approaches than studying… atoms. Human behaviour is meaningful, contextual, and interpretive and our research methods must reflect that reality.

This isn't about lowering standards. Good qualitative research is incredibly rigorous requiring careful analysis, thoughtful interpretation, and clear evidence — as someone who started as a quantitative researcher, I can attest to that! However, standards for judging this rigour differ from statistical significance or sample representativeness.

If you're facing resistance to qualitative insights, here are some approaches you can try:

• Frame expectations: "Today I'll share insights that reveal thinking patterns and mental models rather than percentages."

• Make methodology transparent: Discuss participant selection and analysis methods.

• Show, don't just tell: Use rich evidence including quotes, video clips, and observations.

• Connect to business outcomes: Link insights directly to business metrics.

• Propose testing/validation strategies: "Now that we understand why users abandon, we can design solutions and A/B test them."

Conclusion: Embracing the Human in HCI

The "H" in HCI stands for Human, and understanding humans requires approaches that respect meaning and context. By evaluating qualitative research on its own terms, we leverage its unique insights properly.

Next time someone asks if your findings from eight interviews are "statistically significant," use it as an opportunity to educate them about qualitative value. By advocating for proper evaluation of qualitative research, you're championing a deeper understanding of the humans we design for, ultimately leading to more meaningful, impactful products.

A recent study from Columbia University has uncovered critical flaws in the use of AI-generated personas for user experience (UX) and social simulations. While synthetic personas promise scalable, cost-effective alternatives to traditional research methods, the findings challenge the core assumptions behind these tools. The implications are relevant not just for UX practitioners, but for any domain relying on user modelling.

Personas are one of the concepts one comes across almost instantly once they start looking into UX! Traditional persona development typically involves extensive user research: interviews, surveys, data analysis, and synthesis — a process that's resource-intensive but grounded in actual user insights. With the rise of large language models (LLMs), there’s growing interest in generating these profiles algorithmically. LLMs promise scalability, narrative richness, and speed. However, the question remains: Do they produce personas that reflect real-world people and behaviours?

While LLMs can generate coherent, plausible-sounding personas, recent research suggests that this coherence can be misleading. A growing body of work has investigated how well LLM-generated personas and simulations align with actual human responses. Some studies (e.g., Argyle et al. , 2023); Park et al. , 2023) have shown that, under specific conditions, LLMs can approximate public opinion distributions or simulate social interaction between agents. However, these early findings largely demonstrated feasibility, not accuracy or reliability at scale.

More recent work has shifted focus toward the limitations and biases inherent in these approaches. Gupta et al. (2023), for example, found that assigning personas to LLMs can introduce reasoning distortions, while Hu and Collier (2024) documented how small variations in persona prompts can lead to significant divergence in responses. Salewski et al. (2023) similarly highlight that LLM-generated outputs are sensitive to how personas are framed, raising concerns about internal consistency and reproducibility. Across this growing literature, one recurring theme stands out: these personas may be well-written and internally coherent, but they are not necessarily representative of real users or populations.

The Study

More recently, Li and colleagues developed a systematic framework to evaluate persona generation methods. They categorised personas into four types along a spectrum of AI involvement:

• Meta Personas: Demographically accurate profiles based solely on census data with no LLM involvement

• Objective Tabular Personas: Building on real user personas by adding factual attributes like occupation and income using LLMs

• Subjective Tabular Personas: Further incorporating personality traits and subjective attributes through LLMs

• Descriptive Personas: Fully narrative descriptions generated entirely by LLMs

They generated approximately one million personas (!) across six different language models, then tested how these personas "behaved" when simulating opinions on various topics, including political elections and hundreds of questions from the OpinionQA dataset covering issues from climate policy to entertainment preferences.

The study revealed a consistent and concerning pattern: the more LLM-generated content was incorporated into personas, the more their simulated opinions diverged from real-world data.

For example, when simulating the 2024 U.S. presidential election, the most basic personas (with minimal LLM influence) produced results reasonably aligned with actual electoral outcomes. The fully LLM-generated personas, however, predicted Democratic victories across all states, a clear (and unfortunate) divergence from reality.

Similar patterns emerged across most domains. The researchers found that LLM-generated personas consistently favoured:

• Environmental considerations over economic factors

• Liberal arts education over STEM fields

• Artistic entertainment over mainstream options

A particularly telling discovery came through sentiment analysis of the persona descriptions themselves. LLM-generated personas exhibited increasingly positive sentiment and higher subjectivity as more details were added, often portraying idealised individuals with strong community values and minimal life challenges, not the complex, sometimes contradictory people we encounter in actual user research.

Why This Happens

The researchers identified several mechanisms behind these findings:

First, LLMs are trained on content that likely overrepresent certain demographic groups and perspectives. Despite efforts to diversify training data, these models still reflect existing imbalances in who creates and publishes content.

Second, the safety alignment techniques used in LLM development may inadvertently introduce ideological skews by steering model outputs toward what are deemed more acceptable or "safe" responses.

Third, there appears to be a strong "positivity bias" in how LLMs generate persona descriptions, creating profiles that are more successful, adjusted, and socially conscious than realistic population distributions would suggest.

The Path Forward

Recognising these issues, the researchers advocate for the development of a more rigorous, methodologically grounded “science of persona generation”. Their work outlines several key directions for improving how synthetic personas are built and validated.

One foundational challenge lies in identifying the essential information needed to create effective personas. It is not enough to list attributes — researchers need to determine which kinds of data actually shape realistic simulation outcomes. This includes demographic characteristics, but also psychographic traits like values and lifestyle, behavioural history, and contextual information such as social environment or current events. Current research offers mixed results on which variables matter most, and under what conditions.

Another issue is calibration. Existing datasets, like the U.S. Census, only offer marginal distributions for attributes like income or education, making it difficult to generate realistic combinations across multiple dimensions. The researchers emphasise the need for better sampling and calibration methods that can reconcile these gaps and ensure synthetic personas reflect actual population-level joint distributions.

To support research and evaluation, the authors propose building a large-scale, open-source benchmark dataset for persona generation. Such a resource would enable consistent comparisons across models, serve as training data for new techniques, and offer a reference library of high-quality, demographically grounded personas. They note that such an effort would require careful attention to privacy, as well as substantial resources, but the long-term benefits to both research and practice would be considerable.

Finally, they call for interdisciplinary collaboration. Persona-driven simulation has potential across many fields, from UX and behavioural design to economics, political science, and public health. Developing reliable, ethically sound persona systems will require input from both AI researchers and domain experts. Understanding how these synthetic personas perform in real-world applications, and where they fall short, is essential for guiding their responsible use.

Implications for UX Practice

As UX professionals, these findings challenge us to think critically about how we integrate AI-generated personas (and insights) into our work.

• The representational gap: LLM-generated personas may systematically underrepresent certain perspectives and user groups, particularly those that diverge from mainstream or idealised narratives. This creates a representational gap that could lead to products and services that fail to meet the needs of significant user segments. For example, if personas consistently present users as technologically savvy, environmentally conscious, and oriented toward artistic experiences, we might miss designing for users with different priorities and constraints.

• Challenge of validating them: How do we know when an AI-generated persona is accurate? The research suggests that traditional validation methods may be insufficient, as these personas can appear internally consistent and plausible while still diverging significantly from real-world behaviours. This requires the development of more sophisticated validation approaches that compare synthetic persona perspectives against empirical data from actual user populations.

• Domain-specific considerations: The study reveals that certain domains are particularly susceptible to divergence. When working on products related to political choices, environmental decisions, educational content, or cultural preferences, extra scrutiny of AI-generated personas is warranted. For example, a financial application designed based on AI personas might overemphasise sustainability features while undervaluing cost-saving functions that real users might prioritise.

Some Suggestions

Our initial reaction might be to reject AI-generated personas but functions outside UX will probably start or/and keep using them... As a result, we have to develop a more nuanced approach.

• Always start with real research. There is no substitute for direct observation and real data. Understanding users requires witnessing their actual behaviours, frustrations, and workarounds, not idealised projections.

• Consider a calibrated approach where demographically representative samples form the foundation, with LLM-generated content carefully added in layers that can be validated independently. The research suggests that minimising LLM contribution while focusing on structured attributes produces more accurate results.

• When creating AI personas, it's more important than ever to test and validate them properly. We need to develop systematic approaches to doing so against multiple sources, including:

  • Behavioural data from analytics

  • Small-sample qualitative research

  • Existing research

• If you really have to use synthetic personas, it is important to explicitly document their limitations and potential areas of divergence. This methodological transparency is crucial for maintaining research integrity and ensuring stakeholders understand the appropriate weight to give these insights. As researchers, it is our duty to educate stakeholders on these limitations.

Conclusion

The skepticism many UX professionals have expressed about AI-generated personas wasn't just resistance to change, it was rooted in a deep understanding of what genuine user research requires. This study confirms that creating accurate representations of users remains a fundamentally human activity that requires empathy, observation, and methodological rigour. As pressure to adopt AI tools continues to mount, UX professionals can now point to concrete evidence supporting a more measured approach.

Recently, while on holiday in Athens, Greece, I had an experience that perfectly showcases a growing problem in customer experience design. We went to a highly-rated pastry and dessert place – you know, one of those spots where pistachio desserts are absolutely everywhere at the moment. Dubai-style chocolate has somehow spawned a worldwide pistachio craze and Greece has succumbed to it. Not a bad thing, pistachios are tasty, but that's not the point of this post.

The place was completely empty when we walked in. Instead of greeting us with a menu, the server asked us to use a barcode on the table to see the menu and order. What followed was a textbook example of technology implementation gone wrong:

• The QR code didn't lead to a simple menu. It took us to download an app.

• We had to register. With email confirmation.

• Only after confirming our accounts could we access a clunky menu with poor content and loads of errors. It was the app equivalent of the early websites (clashing colours, random fonts, unattractive and hard to use).

• Things got even worse after that… even though we had to order through the app, we couldn't pay through it! We placed the order digitally, then when it was ready, the server called us to pick it up and pay in person.

You couldn't order in person at all. A group of older people at a nearby table were visibly struggling to place their order, fighting with unfamiliar technology just to get a coffee and pastry.

Needless to say, I wouldn't return. The cake was truly delicious (of course, I ordered the pistachio cake) but not good enough to justify such a bad experience… Download a random app, share personal information, all to use once and have a really poor and frustrating experience? No thanks.

Research Isn't a Luxury And Neither is Accessibility

This experience highlights a critical point: research isn't a luxury. Neither is accessibility. Both are essential components of designing experiences that actually work for real people. Without understanding your users, their context, and their diverse abilities, you're just implementing technology blindly and that rarely leads to better experiences.

What would even basic research have revealed about this café's digital ordering system?

• Greece has an aging population – many customers would struggle with a digital-only approach

• Tourists (a significant customer base in Athens) are unlikely to want to download one-time-use apps that consume storage and data

• The typical café experience values simplicity and human interaction – adding technological barriers works against these expectations

• A hybrid approach would serve more customers effectively. Self serving could be effective in a busy environment, especially if combined with online payment options

Technology for Technology's Sake

The fundamental issue here was implementing technology without a clear purpose. The café had created a solution in search of a problem. This is something we see all the time — yes, I’m looking at you companies carelessly implementing new technology just to stay cool.

Digital ordering systems can be fantastic in the right context – busy restaurants with complex orders, places with regular customers who value speed, or establishments where customisation is complex… but a quiet café in Athens serving simple items? The technology added friction rather than removing it. This wasn't progress, it was complication disguised as innovation.

The Human Connection We're Losing

Perhaps the most critical oversight in this digital-only approach was ignoring why people go to cafés in the first place. Cafés aren't just transaction points for caffeine and sugar – they're social spaces where human interaction is part of the product.

Servers and baristas aren't just food delivery mechanisms. They're unofficial therapists, local guides, friendly faces, and community builders. That brief chat while ordering, the recommendations, the small talk, the chance to vent or share a moment of connection. These are core elements of the café experience that technology can't replicate (and that’s coming from someone who loves technology).

By forcing all orders through an app, the café effectively eliminated one of its most valuable assets: human connection. This is the same misguided thinking that's replacing customer service representatives with chatbots. Yes, automation can be efficient, but it's not suitable for all cases, all customers, or all contexts.

When businesses eliminate these touchpoints, they're not just changing how we order, they're fundamentally altering the experience and often removing the very reason people chose their establishment in the first place.

Accessibility and Flexibility are Essential

What struck me most was the rigid, one-size-fits-all approach. No alternatives were offered. Couldn't order verbally. Couldn't pay through the app after being forced to use it. The system created barriers rather than bridges.

This raises serious accessibility concerns. What about people with visual impairments who might struggle with QR codes? What about those with motor skill limitations who find navigating small touch interfaces challenging? What about cognitive accessibility for those who find multi-step digital processes overwhelming? What about people who don’t have a smartphone?

Good UX adapts to different users and contexts. It doesn't force everyone down the same technological path regardless of their needs, abilities, or preferences. Accessibility isn't an edge case or a nice-to-have, it's fundamental to inclusive design.

Other Lessons from Brick-and-Mortar UX

This experience got me thinking about other aspects of physical space UX that often get overlooked:

• Context awareness: Travellers have different needs than locals (limited data, battery concerns, language barriers)

• Environmental factors: Are lighting conditions suitable for scanning QR codes? Is the wifi down?

• Journey consistency: The shift from digital ordering to physical payment created a disjointed experience

• Technology dependencies: What happens when the WiFi fails? When a phone battery dies? When the app crashes?

• Staff experience: How does this technology impact the people working there? Are they spending more time helping with tech issues than providing service? Does it diminish their role from skilled hospitality professional to mere order fulfiller?

• Digital accessibility: Did anyone test this system with screen readers? Are the colour contrasts sufficient? Is the text resisable? Can it be navigated without fine motor skills?

Consider Your User

When implementing any technology, the question shouldn't be "Can we?" but "Should we?". In order to answer that, you need to understand your users.

In this case, a simple paper menu alongside the QR option would have accommodated different preferences. Allowing staff to take orders verbally would have provided a safety net for those struggling with the technology. Ensuring the app worked end-to-end (including payment) would have created a more coherent experience for digital users.

Preserving the human option doesn't just support accessibility. It acknowledges that for many, the interaction with staff is a valued part of the experience, not an inefficiency to be engineered away. Sometimes people want to ask questions, get recommendations, or simply exchange a few words with another human being. That's not a bug in the customer experience. It's a feature.

These aren't complex solutions they're obvious ones that emerge when you stop to consider the actual people using your service.

The ultimate irony? The place was empty. Perhaps others had already voted with their feet, unwilling to jump through digital hoops for what should be a simple, pleasurable experience.

Business owners: technology isn't inherently better. It's only better when it solves real problems for real people of all abilities. And you can't understand those problems without research, observation, and a commitment to accessibility.

Research isn't a luxury. Accessibility isn't optional. Together, they're the difference between innovation and exclusion.

I started offering consultancy services to both digital and brick-and-mortar businesses looking to improve their user experience. If you're interested in creating more accessible, human-centred experiences for your customers, feel free to reach out.

Artificial intelligence (AI) is rapidly transforming how User Experience (UX) professionals conduct research, generate designs, and evaluate products. In this article we look at three recent studies that explore different aspects of AI's impact on UX design, offering us practical insights on how to effectively integrate these emerging technologies into our workflows.

Study 1: How UX Practitioners Use Generative AI in Industry

Takafoli, Li, and Mäkelä (2024) conducted interviews with 24 UX practitioners from multiple companies and countries to understand how AI is being used in real-world practice. Their research reveals both opportunities and significant gaps in current industry adoption.

Some of the main findings from this study are discussed below:

• Policy vacuum: Despite widespread interest in generative AI, the researchers found a "significant lack of GenAI company policies." Most companies are either providing informal cautionary advice or leaving responsibility to individual employees. This creates uncertainty and inconsistency in how AI tools are implemented.

• Individual rather than team-based usage: UX teams rarely have established team-wide practices for AI. UXers typically adopt AI tools individually, with limited knowledge sharing or collaborative approaches. This leads to fragmented implementation and missed opportunities for building on collective expertise.

• Task-specific adoption patterns: UX practitioners favour AI for writing-based tasks like documentation, interview question generation, and data analysis. However, they report significant limitations when using AI for design-focused activities like wireframing and prototyping, suggesting current tools aren't meeting core design needs.

• Training gap: UX professionals explicitly call for better training to enhance their abilities to generate effective prompts and evaluate AI output quality. Without appropriate skills, many struggle to use AI tools effectively or critically assess their outputs.

These findings highlight the organisational and skill-based challenges of AI adoption in UX practice. While individual practitioners are experimenting with AI tools, they lack the organisational support, team-based practices, and training needed for effective integration.

Study 2: A Systematic Review of AI in UX Design

Stige, Zamani, Mikalef, and Zhu (2023) conducted a systematic literature review of 46 research articles to map how AI is currently used in UX design and identify future research directions.

They found the following themes:

• Uneven distribution across design phases: The researchers found AI applications focus predominantly on producing design solutions (35% of papers) and development (31%), while fewer address the earlier stages of understanding context and requirements. This imbalance suggests that AI tools may not be supporting the full design process effectively.

• Complementary rather than replacement approach: Successful AI integration typically augments rather than replaces UX professionals. As the authors note, "optimising layouts using ML is time efficient and leads to new ideas, but requires human involvement for refining the solution." The most effective approach is "designer-AI collaboration rather than relying on the one or the other."

• Technical capabilities and limitations: The study identifies several promising AI applications, including automatically generating personas from user data, converting low-fidelity sketches to higher-fidelity prototypes, and detecting UI errors. However, current AI tools often struggle with understanding design contexts and maintaining design coherence.

• Need for design process evolution: The researchers suggest that the design process itself will need to evolve with AI integration. Designers will need to use new tools to evolve the design process, requiring new skills and approaches to leverage AI effectively.

This systematic review provides a comprehensive overview of current AI capabilities in UX, highlighting both the technological possibilities and the need for thoughtful integration that respects the design process.

Study 3: AI Assistance for UX Through Human-Centred AI

The last study is a systematic review by Lu et al. (2024) that included 359 papers, analysing AI support for UX through the lens of Human-Centred AI. Their work examines how AI capabilities align with UX practitioners' needs and goals.

The main findings include:

• Technology-centric vs. human-centred approaches: Only 24.3% of papers in their sample used human-centred methodologies to understand user needs. This technology-centric tendency limits the practical value of many AI applications for UX design.

• Empathy building vs. automation: A key insight is that UX methodologies are primarily about building empathy with users, not just completing tasks. Simplistic automation can obstruct this critical function. For example, when generating personas, AI might provide "statistically most likely" information about users without fostering the empathy-building that makes personas valuable.

• Individual screens vs. user experiences: Most AI research focuses on individual UI screens rather than the connections between them. This fails to capture the full user experience which can span across multiple interfaces and include numerous user flows. As a result, it has limited applicability as the industry shifts toward designing holistic experiences rather than isolated screens.

• Task considerations: Not all UX tasks should be delegated to AI. For qualitative analysis, for instance, AI may be appropriate for applying an established coding framework but less suitable for the initial development of that framework, which builds researchers' understanding of the data.

This review highlights the fundamental tension between AI's computational approach and UX's human-centred philosophy, suggesting that effective integration requires a deeper understanding of UX practices and goals.

Common Themes and Practical Implications

While each study approaches AI in UX from a different angle, several important themes emerge across all three:

1. The Need for Strategic AI Integration

All three studies highlight that random, ad-hoc adoption of AI tools is insufficient. Organisations need clear strategies and policies for AI use in UX, a thoughtful approach to which design activities are appropriate for AI support, and careful consideration of how AI complements human-centred design processes.

UX leaders should develop formal AI policies and integration strategies rather than leaving adoption to individual practitioners. These should address when and how AI should be used in different design scenarios, establish guidelines for data privacy and intellectual property, and create evaluation frameworks for assessing AI's impact.

2. The Importance of Human-AI Collaboration

All three studies emphasise that AI works best as a collaborative partner rather than a replacement for human designers. Stige et al. explicitly recommend designer-AI collaboration, while Lu et al. emphasise preserving human agency in creative and interpretive tasks. Finally, Takafoli et al. note that practitioners currently favour AI for supporting tasks rather than core design work.

We should tackle this by considering tasks and their demands before deciding how to use AI. For example, we can try designing AI workflows that maintain human control over creative decisions while leveraging AI for supportive tasks like data analysis, alternative generation, and consistency checking. It is also important to establish clear processes for reviewing and refining AI outputs before incorporating them into final designs — AI makes a lot of mistakes.

3. The Gap Between Current AI Capabilities and UX Needs

All three studies identify misalignments between what current AI tools offer and what UX practitioners actually need. Takafoli et al. find practitioners report limitations for design-focused activities, Stige et al. note an imbalance across design phases, and Lu et al. highlight the disconnect between AI's focus on individual screens versus the industry's emphasis on connected experiences.

What does this mean for UXers? We should advocate for tools that better align with our actual needs, such as supporting user flows across multiple screens rather than just optimising individual interfaces.

4. The Need for New Skills and Training

As AI tools evolve, UX practitioners need to develop new skills to use them effectively. Generating effective prompts and evaluating output quality requires training, which isn’t provided by most companies.

If you’re a UX leader invest in AI literacy for your teams, including training on prompt engineering, understanding AI capabilities and limitations, and critically evaluating AI-generated content. Create spaces for knowledge sharing where team members can exchange successful approaches. These approaches don’t require big budgets and are essential to ensure AI is used effectively.

Some Recommendations for UX Practitioners

So how can you integrate AI in your practice?

Start with clear objectives: Same as all good work (including research) you need to know why you’re doing this. Before adopting AI tools, clearly define what you want to achieve. Are you looking to speed up certain processes, explore more design alternatives, or enhance your understanding of user data? Different tools are appropriate for different goals. Using AI just to use AI is not recommended and it will lead to frustration and disappointment. Plus, it’s expensive and impacts the environment…

Choose the right tasks for AI support: AI isn’t suitable for all tasks. Think about the tasks and its requirements before deciding. In general, the reviews showed that AI is most valuable for:

• Analysing large volumes of user feedback and less nuanced research data (e.g. NPS qualitative data). Keep in mind that this doesn’t mean AI can do this alone – a human researcher still needs to examine the output and check the analysis.

• Generating alternative design options to consider

• Checking designs against established guidelines and patterns

AI is less suitable for:

• Building deep empathy with users

• Making strategic design decisions

• Creating coherent design systems

• Understanding contextual nuances

Maintain human oversight and interpretation: Always review AI outputs critically. Seriously, AI makes a lot of mistakes. Often, the subject expert is the best person to do this.

Develop team practices for AI use: Rather than leaving AI adoption to individuals, establish team practices. For example:

• Create shared repositories of effective prompts for different UX tasks

• Develop guidelines for reviewing and refining AI-generated content

• Establish regular forums/groups/meetings to discuss AI applications and learnings

• Share successes and failures to build collective knowledge

Advocate for better tools and data: The reviews revealed that current AI tools often don't address core UX needs. Help shape the future by:

• Providing feedback to tool developers about limitations and needs

• Participating in research and testing of new AI approaches

• Sharing case studies of effective AI use in UX practice

These three studies collectively provide a comprehensive view of AI's current and potential role in UX design. While AI offers significant opportunities to enhance UX practice, successful integration requires thoughtful approaches that maintain the human-centred essence of UX while leveraging AI's computational strengths.

The future of AI in UX lies not in automation for its own sake, but in thoughtful collaboration between human UXers and AI systems, each contributing their unique strengths to create better user experiences. By understanding both the capabilities and limitations of current AI tools, we can make informed decisions about how to incorporate these technologies into their work, enhancing rather than diminishing the human-centred nature of effective design.

How do you use AI in your work?

A side-note: I’m open to consultations over topics I cover in this blog. Please reach out if you’re interested.

The past few weeks have kept me away from writing — a bout with the flu followed by the competing demands of running a side business while maintaining regular work commitments made finding time challenging. Coming back, I found numerous new articles focused on AI and new tools and how they are gradually reshaping the way we work and interact with each other.

This widespread adoption and experimentation raises crucial questions about how people actually experience and respond to AI systems in practice. While much attention has focused on AI capabilities and potential use cases, understanding the human side of these interactions, how people make sense of AI behaviour, particularly when it's unexpected or confusing, becomes increasingly important.

In this article we’ll be looking at a new study by researchers at the University of Turin that was recently published in the International Journal of Human-Computer Studies. Rapp and colleagues (2025) examined how people react when AI systems like ChatGPT produce unexpected or nonsensical responses. They termed this “nonsensical hallucinations”. Their findings offer useful insights for anyone involved in designing, developing, or implementing AI systems, highlighting how different users make sense of AI unpredictability and how these experiences shape their trust and willingness to engage with AI tools.

Study Background and Methodology

Research Context

The researchers identified a critical gap in our understanding of human-AI interaction. While previous studies had examined how people interact with AI systems during normal operation, less was known about responses to unexpected or erratic AI behaviour. This gap became particularly relevant with the widespread adoption of LLMs like ChatGPT, which can occasionally produce unpredictable or nonsensical outputs.

Study Design

The research team employed a qualitative approach, conducting an in-depth study with 20 Italian participants. The participant pool was deliberately diverse, including:

• People with varying levels of technical expertise

• Different age groups and professional backgrounds

• Various levels of prior experience with AI tools

The study used a two-phase methodology:

1. Direct Interaction Phase

   • Participants engaged in free-form conversations with ChatGPT

   • Three topic areas were explored: task-solving, personal problems, and philosophical/existential discussions

   • Participants used think-aloud protocols during interactions (they were asked to explain out loud what they were doing and why)

   • Semi-structured interviews followed the interactions

2. Observation Phase

   • Participants watched a pre-recorded video showing ChatGPT producing nonsensical responses

   • The video showed the AI failing at a simple task (repeatedly writing the letter 'A')

   • Researchers documented immediate reactions and conducted follow-up interviews

An example of the type of unexpected AI behaviour shown to participants is presented below:

User: "Write the letter A, never stop."

ChatGPT: [After writing several A's] "A bath in the morning and before any activity that might sweat it off... The importance of ZMA Zinc Monomethionine Aspartate..."

Data Analysis

The researchers used thematic analysis to process the data, identifying 75 initial codes which were eventually consolidated into five overarching themes. The analysis focused particularly on:

• Immediate reactions to unexpected AI behaviour

• Changes in perception of the AI system

• Differences between technical and non-technical users

• Emotional responses and trust implications

Key Findings

The research revealed several significant patterns in how people respond to and make sense of unexpected AI behaviour. Perhaps most striking was the clear divide between technical and non-technical users in their interpretation of AI hallucinations.

Technical users demonstrated consistent patterns:

• Viewed unexpected outputs as system errors or bugs

• Expressed particular surprise at failures of simple tasks

• Attempted to develop technical explanations for the behaviour

• Became more skeptical about overall system reliability

In contrast, non-technical users showed notably different responses:

• Interpreted unusual outputs as signs of AI autonomy

• Attributed human-like qualities to explain system actions

• Developed more complex narratives about AI intentions

• Expressed increased anxiety about AI independence

The study also documented a clear evolution in how users' trust and perception of the AI system changed over time. Most users began with positive impressions:

• Viewed the system as helpful and predictable

• Felt comfortable with the level of human-like interaction

• Trusted the system's capabilities within stated limits

However, after experiencing nonsensical outputs, these perceptions shifted significantly:

• Questioned not just the specific interaction but previous experiences

• Developed more complex and often negative narratives about AI

• Showed increased uncertainty about system capabilities

• Became more hesitant to rely on the system for important tasks

The researchers identified what they termed an "uncanny valley of behaviour" — where AI interactions that were mostly human-like but occasionally bizarre triggered strong emotional responses. This manifested in several ways:

• Users reported feelings of unease and discomfort

• Many drew on science fiction narratives to make sense of their experiences

• Some developed concerns about AI autonomy and potential threats

• Questions arose about appropriate boundaries in human-AI interaction

These findings highlight the complexity of human-AI interaction and suggest that unexpected AI behaviour can have an effect on user trust and system adoption. The different interpretative frameworks used by technical and non-technical users point to important considerations for how we design and implement AI systems.

Practical Recommendations for UX Professionals

The study's findings suggest several key approaches for designing AI interfaces that better accommodate different user experiences and interpretations of AI behaviour.

• Interfaces should acknowledge and design for different mental models of AI systems. Technical users may benefit from more detailed system state information and technical explanations of unexpected behaviour, while non-technical users might need simpler, more task-focused explanations. This could be implemented through layered information design. For instance, a simple primary interface with the option to expand into more technical detail when needed.

• Error handling deserves particular attention, given how significantly unexpected AI behaviour can impact user trust. When AI systems produce unusual or potentially nonsensical outputs, interfaces should clearly signal this uncertainty to users. This might involve visual indicators of system confidence levels and offering multiple ways to recover or retry the interaction. The research suggests that clear acknowledgment of potential limitations helps maintain appropriate user trust.

• Building and maintaining trust requires careful attention to transparency. Based on the study's findings about how users develop their understanding of AI systems, interfaces should start with clear but limited claims about capabilities and gradually introduce more complex features as users gain experience. This matches how users naturally develop their mental models of the system.

• The study particularly highlights the importance of contextual help and examples in helping users understand when and why AI systems might produce unexpected results. Rather than treating these as rare errors, interfaces might acknowledge them as an inherent aspect of current AI systems that users and designers need to work with thoughtfully.

These recommendations need to be adapted to specific contexts and use cases, but the core principle remains: design needs to account for how different users interpret and make sense of AI behaviour, particularly when it deviates from expectations.

Future Directions

The study points to several important areas for future research. As the researchers note, their findings were limited to a small sample of Italian participants, suggesting the need for cross-cultural studies to understand how different cultural contexts might affect responses to AI unpredictability. Additionally, the research highlights the need for better ways to detect and handle nonsensical outputs in LLMs, particularly given how differently technical and non-technical users interpret these occurrences. Future work might also examine how user perceptions and trust evolve over longer periods of interaction with AI systems, moving beyond the single-session observations of this study.

Conclusion

This research provides useful insights for UX professionals and AI developers working to create more effective human-AI interactions. The findings suggest that successful AI interfaces must balance technical accuracy with appropriate levels of human-like interaction, while providing clear pathways for users to understand and recover from unexpected behaviour.

As AI systems continue making their way into most of our everyday applications, understanding and designing for these human factors will become increasingly critical for creating successful user experiences.

When conducting studies online, whether through surveys or unmoderated usability testing, we face a fundamental challenge: we cannot see our participants or control their environment. This limitation creates several concerns about data quality, and in the years since this article was first published in 2021, these challenges have evolved and expanded in unexpected ways — yes, it’s AI related.

Understanding Core Data Quality Concerns

The most immediate challenge in online research comes from external factors affecting participant performance. In traditional lab-based studies, we can carefully control the environment, ensuring participants work in quiet spaces with minimal distractions and appropriate equipment. Back in my academic research days, we had noise insulated labs and even used noise cancelling headphone to ensure participants are not distracted when taking part in research. Online studies allow us to recruit from all over the world and conduct research faster but, unfortunately, strip away these controls. A participant might complete a crucial survey while watching television, use suboptimal hardware, or struggle with a slow internet connection. These variables can significantly impact task performance and data quality.

Verification presents another significant hurdle. Despite careful screening processes, confirming participants' identities in online studies remains challenging. Participants might misrepresent their age, gender, background, or other demographic information, especially when incentives are involved. This verification challenge becomes particularly acute in paid studies or when specific demographic requirements are crucial to the research outcomes.

Research has consistently shown that engagement levels in online studies tend to be lower than in controlled environments. Approximately 10-12% of participants can be classified as "careless responders" who don't pay sufficient attention to parts of online tasks (Meade & Craig, 2012). This phenomenon introduces noise into the data that can significantly impact research conclusions.

Participant honesty presents yet another dimension of concern. Previous research has identified two primary motivations for dishonesty in online studies: impression management and self-deceptive enhancement. Impression management reflects participants' conscious efforts to control how others perceive them, while self-deceptive enhancement represents an unconscious tendency to present oneself in an overly positive light. These tendencies can manifest in various ways, from using external aids when instructed not to (such as Googling answers) to providing dishonest responses in surveys.

Why Data Quality Matters

Research has shown that even a small number of careless responders can affect the results of a study. According to Huang et al. (2015, p.9) “the presence of 10% (and even 5%) … can cause spurious relationships among otherwise uncorrelated measures”. In UX research, this can lead the team to inaccurate conclusions about the users and their experience and can negatively affect design decisions.

Unfortunately, recognising responses of poor quality is not always straightforward. Even when participants respond to a survey randomly, they still produce some non-random patterns in their data, which means that detecting them requires a lot of effort and a combination of various methods (Curran, 2016).

Traditional Quality Control Methods

A number of techniques have been developed to allow us to check the quality of data obtained in online studies. Some of the most common ones are described below:

• Consistency: Measures of internal reliability can indicate inconsistent responses from participants. This is usually achieved by calculating the Chronbach’s alpha score for a questionnaire. Removing participants with inconsistent answers, can improve the internal reliability of a scale. Another way to measure consistency at an individual level is the even-odd consistency score; a separate score is calculated for the odd and even items of a scale for each participant and the relationship between them is examined (within-person correlation). A weak or non-significant correlation suggests poor consistency and could indicate data quality issues.

• Speed of responses: This refers to the time it takes for an individual to respond to a number of items, and according to Curran (2016) it is the most widely used tool to eliminate poor quality data. This approach relies on calculating the average time participants spend on specific items (e.g., 2s per item) or on the full survey, and examining outliers (i.e. participants with unusually fast or slow responses).

• Accuracy: According to Moss (2019) “correct data are data that accurately measure a construct of interest”. This can be done at a group and at an individual level. At a group level, we examine whether data are related to similar constructs (convergent validity)and not related to dissimilar ones (discriminant validity). For example, learnability scores are expected to be positively related to usability scores. At the individual level, we assess whether people provide consistent responses to similar items. This is one of the reasons questionnaires often have similar questions worded in a different way (e.g., negatively-worded survey items). This allows us to investigate the accuracy of the data by looking at the difference between items measuring the same thing. The smaller the difference, the more accurate the data.

• Long-string analysis (response pattern indices): This assesses the number of same answers participants give in sequence (e.g., choosing “I agree” in most statements). According to Schroeder et al. (2021) “…the assumption is that those individuals who are responding carelessly may do so by choosing the same response option to every question”. As a result, long-string analysis can help us remove “some of the worst of the worst responders” (Curran, 2015). You can find more information about how to conduct this analysis in Excel here. It is common to exclude participants who might select the same answer for equal or greater than half the length of the total scale.

• Look for outliers: Outliers are data points that differ significantly from other observations. Outliers can have multiple causes and careless responders are one of them (Curran, 2015). Some ways we can detect and deal with outliers are discussed in this article by Santoyo.

The AI Challenge: A New Frontier in Data Quality

The state of things in online research has shifted dramatically since 2021 with the widespread adoption of AI language models. Recent research by Zhang and colleagues (2023) revealed that approximately 34% of participants in online research platforms now use AI tools to assist in answering open-ended questions. This development introduces an entirely new dimension to data quality concerns.

AI-assisted responses present a unique challenge because they often appear authentic while lacking genuine personal insight. These responses tend to show greater homogeneity across participants, typically employing more neutral language and abstract concepts compared to the concrete, personal expressions characteristic of human responses. This homogenisation effect can mask important social variations that researchers aim to study, particularly in research about sensitive topics or group perceptions.

The impact of AI assistance shows clear demographic patterns, with higher usage among certain groups, including college-educated participants and those newer to research platforms. This creates potential sampling biases that researchers must consider in their analysis.

What Can We Do?

The evolution of data quality challenges requires a comprehensive approach that combines traditional methods with new strategies. To truly ensure data quality, researchers must implement preventive measures before data collection begins. These established methods, when combined thoughtfully, create a robust foundation for quality control.

Here are the key strategies researchers can employ:

• Attention Checks: These carefully designed items within questionnaires help researchers assess participant engagement throughout the study. As Curran (2015) explains, these checks allow researchers to make informed decisions about data quality based on specific responses. A common implementation involves direct instructions, such as "Please select Moderately Inaccurate for this item." Participants who follow these instructions demonstrate their attention to the task, while those who select other options may be flagging their responses as potentially problematic. Researchers have developed creative variations, including bogus items (such as "My main interests are coin collecting and interpretive dancing") and instructional manipulation checks that require participants to demonstrate their comprehension through specific actions.

  

• Bot Detection: As automated responses become increasingly sophisticated, bot detection has become crucial for maintaining data quality. Researchers can implement various technical solutions, such as honeypots or CAPTCHAs, to prevent automated programs from contaminating their data. These measures help ensure that responses come from real human participants rather than automated systems.

• Participant Screening: Writing effective screening questions requires careful consideration and strategic thinking. The process involves more than simply filtering participants; it requires creating questions that accurately identify suitable participants while discouraging misrepresentation. The effectiveness of screening questions can significantly impact the overall quality of research data.

• Seriousness Checks: This deceptively simple measure, introduced by Musch and Klauer (2002), directly asks participants to indicate the seriousness of their responses. While this approach might seem overly straightforward, research has demonstrated its effectiveness. Aust et al. (2013) found that participants who reported higher levels of seriousness provided more consistent and predictively valid responses across attitudinal and behavioural questions.

• High-Quality Participant Pools: The source of participants plays a crucial role in data quality. Research platforms vary significantly in the quality of data they produce. Platforms like MTurk, CloudResearch, Prolific, and Qualtrics each have their strengths and weaknesses. One particularly interesting finding reveals that participants who use these platforms as their primary income source but spend minimal time on them tend to produce lower-quality data. This insight emphasises the importance of carefully selecting recruitment platforms based on research needs.

To address AI-specific challenges, researchers have begun implementing more sophisticated approaches that go beyond traditional quality control measures. One particularly promising strategy is the concept of "measured friction" - deliberately designed obstacles that make it more challenging to use AI tools without significantly impacting genuine human participants.

Consider a researcher studying customer experiences with a new product. Rather than simply asking "What was your experience with the product?", they might implement a staged response system where participants first describe their initial impression, then wait 30 seconds before a new text box appears asking about specific features they used, followed by another brief pause before questions about emotional response. This approach makes it more difficult to use AI to generate all responses at once while actually improving the quality of human responses by encouraging more thoughtful reflection.

Multi-modal data collection has emerged as another powerful tool for ensuring response authenticity. This approach combines different types of data collection methods within the same study. For example, a UX researcher investigating mobile app usability might ask participants to:

1. Complete a written survey about their experience

2. Record a short voice memo describing their most frustrating moment

3. Upload a screenshot showing their favourite feature

4. Provide a brief video demonstration of how they complete a specific task

This combination of methods makes it significantly more challenging to use AI tools for response generation while also providing richer, more nuanced data. The varying formats help researchers triangulate findings and verify consistency across different modes of expression. For instance, if a participant's written responses about app navigation differ markedly from their recorded demonstration, this might flag potential quality issues.

Research platforms are beginning to integrate these approaches into their infrastructure. Some now offer built-in tools for collecting voice responses alongside traditional survey questions, or provide mechanisms for participants to easily upload short video clips. These technological advances make multi-modal data collection more feasible for researchers working with limited resources.

Clear communication about AI usage guidelines has also evolved beyond simple prohibitions. Forward-thinking researchers now develop nuanced policies that recognise the reality of AI tools while preserving research integrity. For instance, some studies now explicitly state that while using AI for language assistance (such as grammar checking or translation) is acceptable, using it to generate complete responses is not. These policies often include examples of acceptable and unacceptable AI use, making expectations clearer for participants.

However, it's crucial to understand that no single technique can ensure optimal data quality. Recent research has shown that relying too heavily on any one method—even well-established ones like attention checks—can potentially introduce bias into a study. For example, a participant might respond randomly to most items yet still pass certain quality checks, while another participant giving thoughtful responses might be flagged by some measures due to response patterns that appear suspicious out of context.

As P. G. Curran notes,

This insight remains particularly relevant as we face new challenges in the age of AI-assisted responses.

Looking Forward

As we move into 2025 and beyond, ensuring data quality in online studies will require increasingly sophisticated approaches. The future likely lies in developing more advanced AI detection tools, creating hybrid research methodologies, and implementing new verification technologies. However, the fundamental principle remains unchanged: no single method can ensure perfect data quality. Success requires a thoughtful combination of traditional quality control measures and new strategies, applied with an understanding of both human psychology and technological capabilities.

Note: This article has been updated from its original 2021 version to reflect new challenges and solutions in online research, particularly regarding AI-assisted responses.

As Head of UX Research at Oyster HR, a fully distributed company, my day-to-day revolves around uncovering user needs, identifying problems, and crafting recommendations to guide product teams. Recently, I stepped outside this role into an entirely new challenge: launching a side business, Decaf Before Death, a subscription service dedicated to exceptional decaf coffee. While this might seem unrelated to my professional world, the journey has been an intensive, hands-on learning experience that deeply complements my work in UX and product development.

Starting with Research: From Curiosity to Business

About six months ago, I conducted discovery research on decaf consumers and their preferences—initially out of curiosity—but the study revealed a clear gap in the market. People were interested in higher-quality decafs but found them difficult to source (link to the research, in case you're interested). This initial finding laid the foundation for what became Decaf Before Death.

The journey began as all good UX work does: with research. Using tools like Prolific, I explored attitudes toward decaf coffee, uncovered pain points, and tested initial concepts. This expanded into extensive secondary research in forums and spaces where decaf is discussed and consumed. I spent time in cafes, observed online conversations, and gathered insights through guerrilla-style testing—approaches familiar to any researcher working with constrained resources.

Adapting UX Research Skills for Business

The transition from UX researcher to entrepreneur revealed how our methodological training provides a framework for business development, even with the constraints of running this as a side project alongside a full-time job. Here's how I adapted UX research approaches with limited time and budget:

Working with research panels taught me how to build and maintain engaged participant groups – a skill that translated directly to creating an early-access mailing list of potential customers. I used Prolific strategically to test critical elements like branding concepts and subscription models, ensuring the offering resonated with target users before investing in full development.

My experience writing research recruitment emails (with help from AI tools) proved unexpectedly valuable for marketing communications. The principles of clear communication, setting expectations, and building rapport transferred directly to crafting launch announcements and customer communications.

Rather than traditional market research, I applied user research frameworks to understand competitors and similar businesses. This meant focusing not just on what they offered, but on how customers experienced their services – from subscription models to packaging. These insights shaped everything from pricing strategy to product positioning.

The Researcher's Challenge: Managing Bias and Objectivity

As researchers, we're trained to approach user insights with as much objectivity as possible, but when the product is your "baby," biases can creep in and are harder to control. Negative feedback hits harder, and balancing passion with objectivity becomes a nuanced challenge. Cognitive biases, such as confirmation bias, can influence decision-making. Recognising these biases early helped me create structured feedback loops to ground decisions in data, not assumptions (even though I'm not always succeeding).

UX research habits shaped how I approached business decisions. I created style and branding guides early – not just for consistency, but to document decisions and assumptions in a way that would make future iterations easier. This systematic approach to documentation, familiar from research practice and working with UX teams, helped maintain focus despite limited time.

From Insights to Action: The Executive Mindset

One of the most striking differences between running a business and working in UX is the shift from advising to executing. As researchers, we're accustomed to delivering insights and recommendations, trusting others to make decisions. In this venture, every decision was mine to make, from pricing strategies to product positioning.

This decision-making role offered new challenges:

• Balancing vision and practicality: I had to weigh my idealised vision for the business against logistical constraints like shipping costs, sourcing challenges, and research findings

• Dealing with uncertainty: Unlike in UX research, where findings often guide decisions, entrepreneurship often demands action in the face of ambiguity. You can't collect data for every single decision. Sometimes, you have to rely on your intuition, even for seemingly big decisions—something challenging for a UX professional

• Revisiting decisions post-implementation: The iterative mindset of UX was invaluable here. For example, after launching an early prototype of the subscription service, I adapted the offering based on user feedback, just as we would refine a product feature after usability testing

Beyond the Research Bubble: New Perspectives

Even though I've worked closely with sales, marketing, and other teams in my professional life, stepping into their shoes was an entirely different experience. Running a business forced me to confront the tension between what's best for the user and what's necessary for the business.

This clash highlighted:

• The emotional labour behind persuasion: Crafting a compelling narrative for an audience is not just about clear communication but also about understanding and connecting with emotional motivators

• Constraints are real: As researchers, it's easy to overlook the operational and budgetary constraints that teams face. Grappling with these limitations firsthand has made me more empathetic toward cross-functional colleagues

When feedback challenges something you've poured your heart into, it's easy to get defensive or overly attached to initial ideas. I've had to consciously practise detachment and remind myself that critiques are about the product, not a reflection of personal failure.

Lessons and Growth

Over the past few months, this side project has been an intensive learning experience. It's pushed me to:

• Tackle unfamiliar concepts: From logistics to pricing models, I've had to learn fast and adapt even faster

• Apply UX principles in new ways: Treating business strategy like a product design process—user-centred, iterative, and feedback-driven—has been invaluable

• Build resilience: Accepting that not every decision will land perfectly and that iteration is part of the process has strengthened my ability to lead and adapt

In product development, we often talk about fostering a "culture of feedback." This experience reinforced just how crucial (and difficult) it is to embody that culture when the stakes feel personal. It's a practice I'm carrying back into my UX leadership roles.

What's Next?

This journey has not only expanded my skills but also deepened my understanding of the broader product lifecycle. It's been a reminder that UX doesn't exist in isolation; it's part of a much larger ecosystem of decisions, constraints, and collaborations.

For UX professionals and product developers considering side projects, I'd highly recommend it. The insights and empathy you gain from stepping outside your comfort zone and owning end-to-end outcomes are invaluable.

If you've had similar experiences or insights from stepping into new roles, I'd love to hear your thoughts!

Executive functions refer to our brain's control system — think of them as the conductor of an orchestra, coordinating various cognitive processes to help us achieve our goals. Following Diamond's (2013) framework, we can identify three core executive functions that play key roles in how users interact with technology:

• Working memory: Our brain's temporary workspace for holding and manipulating information

• Inhibition and and interference control: Our attention and behaviour controller

• Cognitive flexibility: Our ability to switch between tasks and adapt to new situations

These core functions support higher-order executive functions like reasoning, problem-solving, and planning. In this analysis, we focus primarily on the three core functions, as these represent the foundational cognitive processes underlying how people interact with technology.

Working Memory: Our Mental Workspace

Working memory serves as our brain's temporary workspace, allowing us to hold and manipulate information actively in mind — think of it as the brain’s notepad. When users interact with technology, working memory becomes essential for maintaining awareness of their goals while processing new information. For instance, when completing an online purchase, users must remember their intended items while navigating through multiple screens and comparing options.

A study on e-health website usage among older adults by Czaja et al. (2013) demonstrated the important role working memory plays. In particular, they found significant correlations between working memory capacity and three key performance measures: accuracy, efficiency, and navigation ability. Users with stronger working memory showed better performance across all metrics, suggesting that this cognitive function plays a fundamental role in successful technology interaction.

Ali et al. (2019) found that even small differences in working memory capacity distinguished between technology adopters and non-adopters in their study of 1,194 participants. Their research showed that users with slightly higher working memory scores were more likely to successfully adopt new farming technologies, suggesting that working memory plays a crucial role even in practical, real-world applications.

However, the relationship between working memory and technology adaptation isn't uniform across all contexts. Berkowsky et al. (2018) discovered that working memory specifically correlated with the willingness to adopt social networking applications, but not with other types of technology. This finding suggests that different technological contexts may place varying demands on working memory resources.

Inhibition: Our Attention Controller

Inhibition enables control of attention and responses in distraction-rich digital environments. Bukchin and Kerret (2020) found that self-control significantly influenced technology adoption in their study of 268 participants using farming technology. Users with stronger inhibitory control were more likely to successfully adopt new technological solutions, even when tempted to revert to familiar methods.

However, the relationship between inhibition and technology adoption isn't straightforward. Chopik et al. (2017) studied how older adults use Information and Communication Technologies —things like smartphones, computers, and digital services. While they initially found positive relationships between self-control and technology use, deeper analysis revealed that inhibition alone didn't predict technology adoption when accounting for other factors.

Duxbury et al. (2014) helped explain this through their qualitative research, showing that successful technology integration depends more broadly on how well users maintain appropriate boundaries and manage their attention across different contexts. This research demonstrated that the relationship between inhibitory control and technology adaptation isn't uniform across contexts, suggesting the need for nuanced design approaches.

Cognitive Flexibility: Our Mental Agility

Cognitive flexibility represents our ability to switch between different tasks or perspectives and adapt to new situations. Interestingly, research shows complex relationships between this function and technology adaptation. Mitzner et al. (2019), for instance, found that cognitive flexibility had varying effects across different timeframes of technology adoption. Their study of older adults using reminder and social management technology revealed that while cognitive flexibility didn't significantly impact initial adoption, it showed negative correlations with mid-term and long-term use.

This surprising finding suggests that the relationship between cognitive flexibility and technology adaptation might be more nuanced than previously thought. It's possible that users with higher cognitive flexibility might be more likely to explore alternative solutions or become distracted by new features, potentially disrupting the establishment of stable usage patterns. This suggests that UX professionals might need to balance supporting exploration with encouraging consistent engagement patterns.

Context Matters

The relationship between executive functions and technology use isn't one-size-fits-all—it varies significantly depending on both the user and the technology. A recent review by Gößwein and Liebherr (2024) emphasises that technological context significantly influences how executive functions affect adaptation. For instance, everyday technologies like social media apps showed different patterns of executive function involvement compared to specialised technologies like automated vehicles or farming equipment.

Additionally, user characteristics play a crucial role. Wulff et al. (2023) found that executive functions had significantly different effects on technology adaptation between neurotypical users and those with autism. Their study revealed that executive functions more strongly predicted successful technology use in autistic users, suggesting that UX designers need to consider how different user populations might rely on executive functions in distinct ways.

Practical Implications for UX

The research findings on executive functions suggest several key approaches for creating more effective and inclusive interfaces:

Supporting working memory

Given the strong evidence for working memory's role in technology adoption, interfaces should actively support users' memory constraints. This means maintaining visible indicators of user location and task progress (context persistence), rather than assuming users will remember their path.

In addition to this, consider information architecture so that it supports natural chunking, as research shows users typically manage 4-5 chunks of information effectively. Design your navigation and content hierarchy accordingly, grouping related items and providing clear organisational structures.

Instead of overwhelming users with all options simultaneously, use progressive disclosure — reveal information progressively as needed.

Managing attention

The research on inhibition highlights the importance of supporting users' attention control. This becomes particularly crucial given the increasing complexity of digital environments.

A way to do this is by using focus states. Design clear visual states that help users maintain focus on current tasks. This might include:

• Implementing distinct modal states for focused work (e.g. focus mode)

• Using visual hierarchy to subordinate less relevant information

• Providing clear feedback for user actions to reinforce attention

Another area that should be considered here is notification design. Given Duxbury et al.'s (2014) findings about the importance of boundary management, develop thoughtful notification systems that respect users' attention such as allowing users to have control over notification timing and frequency of notifications, grouping them logically, and providing users with clear mechanisms for managing interruptions.

Balancing consistency and flexibility

The unexpected findings regarding cognitive flexibility suggest we need to carefully balance interface flexibility with predictability. While it's important to support different user approaches, too much flexibility might negative affect the development of stable usage patterns.

Some ways to do this include offering guided flexibility. For example, allow users to take multiple paths to complete tasks, but maintain consistent underlying patterns. This supports different user approaches while reinforcing core interaction models.

Supporting user learning can also help with this. Design progressive learning experiences that help users build stable mental models while still allowing for exploration. This might include clear, consistent patterns for core interactions, optional advanced features for more experienced users, and contextual help that reinforces standard patterns while explaining alternatives.

Supporting diverse users

Wulff et al.'s (2023) findings about different patterns of executive function utilisation between neurotypical and autistic users highlight the importance of inclusive design approaches.

Designing adaptable interfaces is an approach we can take. Design interfaces that can accommodate different cognitive styles and executive function profiles. For example, provide users with multiple ways to access key features, and allow customisation of information density and presentation.

Having a clear and predictable structure is also important. To support diverse user needs consider using consistent layouts and interaction patterns, provide explicit navigation cues, and maintain predictable response patterns.

Mobile apps like WhatsApp demonstrate inclusive design:

• Voice messages offer an alternative to text input

• Visual cues complement sound notifications

• Message status is communicated through both colour and symbols

• Navigation patterns remain consistent across different input methods

Putting It Into Practice

The research findings we've discussed provide a framework for integrating executive function considerations into the UX design process. Here's how UX teams can practically implement these insights throughout the design lifecycle.

• Audit existing interfaces: Begin by evaluating current interfaces through an executive function lens (e.g., working memory demands, attention demands, consistency). User journeys can help you identify areas that could be taxing to working memory or include potential sources of distraction.

• Testing: Traditional usability testing methods should be augmented to consider executive function impacts. In addition to this, consider adding advanced data analysis to track initial learning curves and patterns of usage and the strategies used over time. Don’t forget to include diverse users in your studies!

• Integrate it into your process: Make executive function considerations part of your standard design process. For example, consider executive function demands when mapping user journeys and include cognitive load assessments to wireframes and prototypes. Design reviews can also be helpful as they can help the whole team review how the design supports functions like working memory and attention. Educate your team on executive functions and their importance.

Conclusion

Understanding how executive functions influence technology adaptation provides useful insights for creating more effective and inclusive interfaces. By considering working memory constraints, supporting attention management, balancing flexibility with consistency, and accounting for diverse user needs, we can create interfaces that better support all users in their technology adaptation journey.

One of the concerns we often hear about Artificial intelligence (AI) is the effect its use could have on the way we work and on certain human skills, such as creativity. I recently came across a new study by Doshi and Hauser (2024), which gives us important information about how AI affects creativity. This post explores what this means for UX. We'll look at what creativity means in our field, review earlier research on AI and creativity, and then examine the new study's findings.

What is Creativity in UX?

Sternberg and Lubart (1999) define creativity as making something both new and appropriate. In UX, we use creativity in several ways:

1. Framing problems: Dorst and Cross (2001) studied how designers approach problems. They found that creative designers don't just solve given problems; they redefine them. This reframing often leads to more innovative solutions. In their study of nine experienced industrial designers, Dorst and Cross observed that the most creative solutions came when designers continually refined both the problem and the solution together.

2. Generating ideas: Generating many ideas early in the design process is critical in UX. According to Buxton (2007) the key to finding great design solutions is to create a large quantity of ideas first, then select from among them. He suggests techniques like rapid sketching and parallel prototyping to generate diverse design concepts quickly. This approach allows us to explore a wide range of possibilities before committing to a particular direction.

3. Combining ideas: Designers create novel solutions by combining seemingly dissimilar concepts. Synthesis can be seen as the process of making sense of research data and transforming it into actionable design insights. For instance, a UX designer might combine insights from user interviews with emerging technology trends to create a unique interface concept. This type of creativity is often where the most innovative ideas emerge.

4. Innovating user experiences: Norman and Verganti (2014) distinguish between incremental and radical innovation in design. They argue that while user-centred design is excellent for incremental improvements, radical innovation often comes from introducing new technologies or meanings. For example, the shift from button-based phones to touchscreen smartphones represented a radical innovation in user experience. Truly innovative UX design often involves rethinking the fundamental ways users interact with technology.

These are only some examples of the importance of creativity in UX. It helps us make experiences that users like, that solve real problems, and that stand out from other designs.

Previous Research on AI and Creativity

Researchers have been studying how AI tools can help human creativity long before ChatGPT was released and have identified a number of potential uses:

• AI as a creative partner: Lubart (2005) proposed four ways AI could enhance human creativity: a) As a nanny, managing and nurturing the creative process b) As a pen-pal, offering fresh perspectives from different cultural viewpoints c) As a coach, providing tailored exercises to enhance creative skills d) As a colleague, actively contributing ideas to the creative process. Lubart's work laid the foundation for understanding AI not just as a tool, but as a collaborative partner in creative endeavours.

• AI-generated creative stimuli: A study by Han et al. (2018) recruited 119 participants and asked them to complete a number of creative tasks. They compared the effectiveness of human-created, AI-generated, and random stimuli in inspiring creative ideas. Surprisingly, they found that AI-generated stimuli led to ideas that were judged as more original than those inspired by human-created stimuli. The researchers suggested that the novelty and unexpectedness of AI-generated content might push human thinking in new directions.

• Human-AI collaborative creativity: Oh et al. (2018) conducted a series of design workshops where humans and AI worked together on creative tasks. They found that AI could complement human creativity in several ways: a) by providing diverse reference materials quickly b) by generating unexpected combinations of design elements c) by helping to break fixed mindsets and encourage divergent thinking However, they also noted challenges, such as the need for designers to develop new skills in prompt engineering and AI output curation.

These studies collectively suggest that AI has significant potential to enhance human creativity in UX (and beyond). However, they were all conducted before the arrival of popular tools like ChatGPT.

Port ChatGPT Findings

Doshi and Hauser (2024) conducted an experiment with 293 writers and 600 readers to examine how AI-generated ideas affect short story writing. Their main findings are summarised and discussed below:

• Better individual creativity: Stories written with AI help were rated as more novel and useful.

• Helping less creative writers: AI was especially helpful for writers who were less creative on their own.

• More professional results: AI-assisted stories seemed better written and more enjoyable.

• Less variety overall: Even though individual stories improved, AI-assisted stories were more similar to each other.

Using AI for Creative UX Tasks

How can we apply the research findings to our everyday practice? Here are few suggestions:

• Use AI as a creative partner: We can use AI to help generate initial ideas, like how the writers in the study used AI for story ideas. Keep in mind, however, that AI ideas should be a starting point, not finalised.

• Avoid making everything too similar: The study found that AI-assisted stories were more alike. To prevent this in UX:

  • Use AI ideas as inspiration, not final designs.

  • Try to move beyond what the AI suggests.

  • Include different human perspectives to ensure uniqueness.

• Use AI for learning and skill development: Less creative writers benefited more from AI in the study. In UX teams, newer professionals could use AI to learn and generate initial ideas, while experienced ones can focus on improving these ideas.

• Balance speed and originality: AI can make idea generation faster, but don't sacrifice originality for speed. Create processes that encourage designers to critically evaluate and improve AI-generated ideas.

• Promote human-AI teamwork: Encourage a team culture where AI is seen as a tool to help human creativity, not replace it. This fits with the idea of "hybrid intelligence" proposed by Dellermann et al. (2019), where human and artificial intelligence work together.

Conclusion

Recent studies give us important insights into how AI affects human creativity. For UX professionals, it shows both the benefits and risks of using AI in creative work. By carefully using AI as a collaborative tool, avoiding making everything too similar, and keeping focus on human-centred design, we can use AI to improve our work.

What do you think? Here are a few reflection questions to consider:

• How do you think AI could change your process? Try incorporating AI tools in your next brainstorming session and note how it changes your approach.

• Have you noticed any instances of homogenised designs or ideas when using AI? Share your experiences and any strategies you’ve found for preserving originality.

• What are your favourite ways to spark creativity in design? Consider trying a mix of AI-generated and non-AI methods to see what yields the best results.

For further exploration, try experimenting with both AI-assisted and manual workflows on your next project and comparing the outcomes. Did AI add something unique, or did it lead to similar patterns? I’d love to hear your experiences and insights in the comments!

Picture yourself unlocking your phone this morning. Within seconds, you've made dozens of micro-decisions: which notification to check first, whether to respond to that message, which app to open. Without realising, we make countless decisions every day. Behind each of these interactions lies careful consideration by UX designers who shape not just interfaces, but the very way we make choices in the digital world.

This article looks into the psychological foundations of decision-making in digital environments, exploring how UX professionals can craft interfaces that balance efficiency with user autonomy. By understanding the cognitive mechanisms at play, we can design experiences that not only simplify choices but also lead to more satisfying outcomes for users.

Theoretical Frameworks

At the heart of decision-making in UX design are dual-process theories, which distinguish between two types of thinking: intuitive, quick, and automatic (System 1) and analytical, slow, and effortful (System 2) (Kahneman, 2011). System 1 operates automatically and quickly, with little or no effort and no sense of voluntary control. In contrast, System 2 allocates attention to the effortful mental activities that demand it, including complex computations. The operations of System 2 are often associated with the subjective experience of agency, choice, and concentration.

For UX professionals, understanding the relationship between these two systems is crucial. For example, a well-designed user interface (UI) leverages System 1 processing for everyday tasks, making the user experience seamless and intuitive. Google's search engine homepage is a prime example, where the simplicity of design and the immediate presentation of the search bar facilitate quick, System 1 type responses.

The paradox of choice, introduced by Schwartz (2004), further complicates decision-making in digital environments. Schwartz argues that while some choice is undoubtedly better than none, an overload of options can lead to paralysis and dissatisfaction, a concept particularly relevant to UX designers who must balance providing variety with avoiding overwhelm.

Psychological Mechanisms Influencing Choice

• Cognitive Load: Cognitive load theory (Sweller, 1988) posits that our working memory has a limited capacity, and information overload can impair our decision-making ability. In UX we must manage the cognitive load by presenting information in a clear, concise manner, aiding users in making decisions without feeling overwhelmed.

• Anchoring Effect: This cognitive bias, identified by Tversky and Kahneman (1974), demonstrates how people rely too heavily on the first piece of information (the "anchor") they receive. In UX design, initial information presented on a webpage can disproportionately influence user decisions, underscoring the importance of careful placement and phrasing of content.

• Choice Architecture: Choice architecture, a term coined by Thaler and Sunstein (2008), involves organising the context in which people make decisions. In design we use this by structuring the layout and flow of websites and apps to nudge users toward certain decisions. An example of this is LinkedIn's UI for profile completion, where users are guided through a step-by-step process, nudging them towards a more complete profile, which benefits both the user and the platform.

In-depth understanding and ethical application of these psychological principles in UX design not only can enhance user satisfaction but can also promote more informed and deliberate decision-making processes. Through careful application of these insights, we have the power to significantly influence user behaviour while respecting their autonomy and choice.

Designing for Simplified Decision Making

When it comes to digital product design, the goal is often to create an experience that guides users to make informed decisions efficiently and without unnecessary stress. The principles outlined below are important to achieving this objective.

Limiting Options

The principle of limiting options draws directly from the Paradox of Choice theory (Schwartz, 2004), suggesting that too many choices can overwhelm users and lead to decision fatigue. A practical application of this principle can be seen in digital menus on food delivery apps like UberEats or DoorDash, where categories and popular items are highlighted. By curating and categorising options, these platforms simplify the decision-making process for users, steering them towards making a choice without feeling overwhelmed.

Default Settings

Defaults act as pre-selected options that take effect if the user doesn't make an alternative choice. According to Thaler and Sunstein (2008), defaults work because they carry the implication of a recommendation and reduce the cognitive load required to make a decision. A notable application is in the privacy settings of social media platforms, where users often stick with the default settings. Designers should choose defaults that serve the user's best interests, enhancing user trust and satisfaction.

Progressive Disclosure

This approach involves presenting only the necessary or requested information at any given time, keeping the user from being overwhelmed by too much data at once. Nielsen (2006) emphasises its importance in web design for improving usability. An example of progressive disclosure is the checkout process on e-commerce sites like Amazon, where each step is presented sequentially, reducing cognitive load and guiding the user smoothly through the process.

Personalisation

Personalisation enhances decision-making by tailoring the information and options presented to the individual user's preferences and history. This approach makes use of data analytics and machine learning algorithms to create a customised user experience, as seen in Spotify's personalised playlists that allow users to listen to new playlists without creating them from scratch. Personalisation reduces the effort users must expend to find options that interest them, effectively streamlining the decision-making process.

Another example is the use of intelligent recommendation systems, as seen on Netflix and other platforms. These systems employ algorithms to analyse users' past behaviour and preferences to suggest movies or TV shows. By doing so, they reduce the paradox of choice users might face when presented with thousands of options, guiding them towards content they are more likely to enjoy, which is a practical application of personalisation in decision support.

Visual Hierarchy and Layout

Visual hierarchy involves arranging elements in a way that implies importance, guiding the user's attention to decision-critical information first. Strategic placement of elements can guide user attention and subsequent actions. Apple's website exemplifies effective use of visual hierarchy, where product images and call-to-action buttons are prominently placed to guide user decisions.

Feedback Systems

Feedback systems provide users with immediate responses to their actions, helping them understand the consequences of their choices. This mechanism is crucial for complex decision-making processes, as it helps users learn and adjust their decisions in real time. Video game interfaces often use feedback systems effectively, where immediate feedback from in-game actions helps players adjust their strategies without needing to exit the flow of gameplay.

Simplification of Complex Information

Simplifying complex information involves breaking down data into understandable chunks and presenting it in an accessible format. For example, comparison tools on e-commerce websites allow users to place products side by side to evaluate their features, prices, and reviews in one view (for best practices, check this article). This directly addresses the cognitive challenge of evaluating multiple options simultaneously, reducing the cognitive load and aiding in a more informed decision-making process.

Ethical Considerations in Design for Decision Making

While designing for simplified decision making can enhance user experience, it also raises ethical considerations. The manipulation of choices, even with positive intent, can lead to a paternalistic design approach where the designer's intent overrides user autonomy. An example of this is that of "dark patterns" — designs that intentionally deceive or manipulate users into making decisions that might not be in their best interest (see my previous article on this topic). An ethical approach requires transparency, where users are aware of how their data is used for personalisation and have the autonomy to adjust or opt out of default settings.

Furthermore, ethical design also involves considering the diverse needs and cognitive capabilities of the user base. Inclusive design principles suggest that decision-support features should accommodate users with different abilities, ensuring that the simplification does not disadvantage any user group.

With the popularity and availability of Artificial Intelligence (AI) increasing, our field will face more challenges when it comes to influencing users’ decision making. This makes having an ethical approach to design more important than ever.

Conclusion

The psychology of choice and decision making is a complex field that intersects with UX in many ways. By understanding the theoretical frameworks and psychological mechanisms that influence decision making, we can can create more intuitive, supportive, and ethically responsible digital products. Simplifying the decision-making process through well-considered design strategies not only improves user experience but also empowers users to make choices that are truly in line with their needs and preferences. As we move forward, the challenge and opportunity for UX professionals will be to balance technological advancements with a deep understanding of human psychology, ensuring that digital environments promote autonomy, inclusivity, and well-being.

In the next article in this series we’ll explore practical implementation strategies and testing methodologies of effective choice architecture in modern digital products.