Don’t Fire Your UX Researchers Yet

Let’s dive into what he had to say. 

Let me start with a confession. When AI-assisted research tools began flooding our industry, I was quietly thrilled. A career of long nights of transcribing interviews and colour coding sticky notes until I started hallucinating, the promise of automated synthesis felt like someone had invented the dishwasher. I was ready to embrace it wholeheartedly.

Then I actually used it. And therein lies the story.

AI in user research is genuinely useful - I want to be clear about that. It can process survey data at speed, identify surface-level patterns, and generate a passable thematic summary before you've finished your second coffee. But "genuinely useful" is doing a lot of heavy lifting in that sentence, because the moment you ask it to do the real work of research - understanding the messy, contradictory, emotionally layered reality of human behaviour - it wobbles alarmingly.

Users, bless them, lie constantly. Not maliciously, not always knowingly. They lie because their ego gets in the way. They want to seem more rational than they are, more expert, more consistent, more the kind of person that reads the small print. They are the type of participant that say they never impulse buy, but their screen recording shows three abandoned carts open in other tabs. That gap between stated and actual behaviour is precisely where great research lives. It is also, at present, precisely where AI falls flat.

What the evidence actually says

Unhelpful Responses

Nielsen Norman Group (NN/G) ran a direct head-to-head test of AI synthetic users against three of their own existing studies with real participants. They found that AI-generated responses were "one-dimensional" and "too shallow to be useful" for most research activities. Real users cared about some things more than others; synthetic users seemed to care about everything equally, fatal for feature prioritisation.

A separate NN/G evaluation found that there are currently "few design-specific AI tools that meaningfully enhance UX design workflows", concluding that AI is not yet ready for designers to rely on as a primary tool.

The Black Box Problem

A widely-cited industry review identifies the "black box problem" as a core ethical issue: AI systems are opaque enough that research participants cannot clearly understand how their data is used, and therefore cannot meaningfully consent. 

Sycophancy

A peer-reviewed study from researchers at Cornell University demonstrated that five state-of-the-art AI assistants consistently exhibit "sycophancy", producing responses that match what users seem to believe rather than truthful ones. 

An AI trained on human feedback learns, over time, to tell people what they want to hear. In a research setting, this means that if your team has already formed a hypothesis about your users, an AI analysis tool is likely to find evidence that supports it, and quietly down-weight the contradictions. A skilled human researcher does the opposite. 

A Human Element

Beyond collecting data, a vital part of a researcher’s job is to build trust with participants. People open up to human interviewers in ways they simply don't with automated systems. Silence, hesitation, the way someone's voice drops when they describe a frustrating experience are all data too, and they require a human in the room to catch them. Empathy, as I'm fond of telling anyone who'll listen, is not a feature that ships in a project update.

As Steve Portigal, one of the field's most respected voices, puts it in the updated edition of his essential guide Interviewing Users (Rosenfeld Media, 2nd ed., 2023): interviewing is not simply talking to people. It is a discipline. Poor interviews produce inaccurate information that takes businesses in the wrong direction. AI doesn't know it's doing a poor interview. 

AI’s Place in Research

None of this means AI has no place in your research practice. It means it has a specific place: handling the grunt work so researchers can do more of the interesting work. Let it transcribe. Let it do first-pass coding. Let it crunch your Likert scale data while you go for a walk. But keep a human in the loop for interpretation, relationship-building, and the creative leap from insight to design implication.

Even NN/G concludes that AI is currently most helpful in the planning and analysis stages of research, functioning as a structuring mechanism rather than a source of original insight. That's a useful tool. It's not a replacement.

Think of it less as a replacement and more as an extremely fast, slightly overconfident intern: 
Useful? Certainly. 
Ready to lead the project unsupervised? Absolutely not.

The organisations getting this right aren't the ones who replaced their research teams with AI. They're the ones who used AI to give their researchers more time to be researchers. More time in the field. More time thinking. More time being genuinely curious about the humans they're designing for.

Which, when you think about it, is rather the whole point.

List of References

1. Rosala, M. and Moran, K. Synthetic Users: If, When, and How to Use AI-Generated Research At: https://www.nngroup.com/articles/synthetic-users/
2. Sponheim, C. (2024) AI isn’t ready for UX Design At: https://www.nngroup.com/videos/ai-not-ready-for-ux/
3. Sharma, M., Tong, M., et al (2023) Towards Understanding Sycophancy in Language Models At: https://arxiv.org/abs/2310.13548
4. Whitman, L. (2025) AI in user research: The battle of ethics and efficiency At: https://maze.co/collections/ai/ethics-user-research/
5. Rosala, M (2026) Use AI Responsibly in Analysis At: https://www.nngroup.com/videos/use-ai-responsibly-in-analysis/
6. Moran, K., Gibbons, S. et al (2025) The UX Reckoning: Prepare for 2025 and Beyond