# I think, therefore I use AI. But only for the right things **Author:** Vera Tans **Date:** 2026-06-05T00:00:00.000+02:00 **Tags:** UX, Artificial Intelligence, Research AI hasn't changed what good user research is. It has made it easier to skip the parts that really matter, if you're not careful. Research itself hasn't changed. The pressure on researchers to work faster, cheaper, and more scalable has. AI takes some of that pressure off, but only when the person using it knows exactly what they're holding. ![UX AI Blog Image](https://images.ctfassets.net/q9t0s93ozyxw/5VZva038OqQhw0kpU5wILp/03b257a9da6d6a3afa3e573ae47c2538/UXAIBlogImage.png) The hype around AI and user research has produced two camps: those who claim you never have to talk to a human again, and those who refuse to touch the stuff on principle. As far as I'm concerned, both have it wrong. The real question is more nuanced, and honestly more interesting: "where in the research cycle does AI actually help, and where does it quietly make things worse?" Nielsen Norman Group recently looked into how AI is being used across the full research cycle, from planning to synthesis. The findings didn't surprise me. AI is mostly used as a thinking partner, as a second pair of eyes, as a note-taker, and as a shortcut. ## What AI does reasonably well There are specific, well-defined tasks in the research process where AI genuinely earns its place. Think: - Transcription and note-taking. - Drafting interview guides with a custom GPT trained on your research method. Think living prompts that you keep refining and making more specific. Also specifying Skills in Claude, or writing instructions in Gemini. - Structuring research plans, especially useful for less experienced researchers who need a content skeleton to start from. - Meta-analysis of large amounts of existing research that would otherwise sit gathering dust. - Building quick prototypes in Figma Make or with Claude Code for usability tests or as a proof of concept. These are tasks that are fundamentally mechanical: structured input produces structured output. AI handles them well precisely because the benchmark for quality is readability, not judgement. The pattern here is that these are tasks where someone with professional expertise immediately sees when the output is off. AI drafts an interview guide, you check it. AI summarises ten interviews, you verify the final synthesis. The human stays in the loop, and as far as I'm concerned, that's essential. ## Where AI falls short, and why that matters The most dangerous application of AI in research is also the most tempting: analysis and synthesis. There's an entire category of "five simple prompts that will transform your research" content floating around that fundamentally misunderstands what an insight actually is. AI can cluster themes, but it doesn't understand what those themes mean for the lives of your users. It can recognise patterns in transcripts, but it doesn't feel the hesitation in someone's voice just before they said what they said. It has no concept of insight, which we as humans (often) do. Nor can AI scope a study. It doesn't notice when a participant is visibly uncomfortable. It misses the telling silence, the nervous laugh, the moment when someone almost says something but then holds back. A human moderator asking contextual follow-up questions in a live interview isn't just gathering information, they create the conditions under which the truth can surface. AI doesn't do that. Not yet. A recent usability test makes this concrete. In the closing questions, a participant gave a high NPS score; he was "actually really happy" with the webshop. On paper, a satisfied customer, done. But while I watched him place an order live, something else caught my eye. He had invented his own workarounds for all kinds of small bugs and errors in the flow. He navigated around them so smoothly that it almost looked planned. When I probed on it, it turned out he would actually prefer those features to work differently. The workarounds cost him extra time every single time, he had just gotten used to them. An AI tool that only processes the score and the transcript would have concluded: happy customer, next. The real insight was in what he was doing, not in what he was saying. Something similar applies to the interview itself. My background in journalism taught me never to ask leading questions, and to dig deeper at the exact moment an answer opens something up. That kind of follow-up is precisely what AI can't do. AI can perfectly well prepare the structure of an interview, deliver a solid skeleton, maintain living prompts. But AI isn't in the room. If you stick strictly to an AI-generated question list and don't recognise the moments when you should deviate from it, you end up with an interview that's tidy but never goes deep. ## The paradox of recognition AI delivers the most value in the hands of someone who has already developed the capacity to use it. Not because a certain number of years of experience is a prerequisite, everyone can and should learn to work with AI, but because you need that capacity for recognition to see when the output is wrong. The ability to say: this interview guide contains leading questions, this synthesis misses the core, this "insight" is really just a rephrasing of the data. So the point is whether the craft-level insight is already in your fingers. And you build that insight in exactly the same way all the generations of researchers before us did: by doing the work yourself. Hand AI the wheel too early and you risk skipping the very parts where that insight is formed. It's like someone who drives with GPS from day one and never builds a mental map of the city. You get there, but you don't know what to do when the GPS drops out. My concern, therefore, is that we as researchers risk skipping the foundation. That we become researchers who are fluent in AI tooling, but who haven't built up the underlying skill that makes this tooling safe to use. And you can fall into that trap at any level, junior, medior, or senior. ## AI tools without a researcher in the loop One of the more worrying trends is the spread of AI research tools being placed directly in the hands of product designers and product owners, without a researcher involved. The risk here is that misinformation gets used. A concrete example: when an AI tool is asked to generate tasks for a usability test, it typically produces leading tasks. These are tasks that tell the participant what should be easy. A usability test built on leading tasks yields nothing, the same goes for an interview with leading questions. It confirms the interface or the assumption instead of challenging it. If no one with expertise has looked at the output, that test just gets run anyway, produces data, and drives decisions. Wrong decisions, if you ask me. The same goes for desk research. AI is a useful starting point, it can help you quickly orient yourself on a topic, cluster existing literature, and surface blind spots in your own knowledge. Use it for that, but stop there. AI has no concept of truth. It generates plausible text, text that sounds like a reliable summary, but which is in reality a static reconstruction of what has been written before. It doesn't distinguish fact from fiction, an outdated source from current insight, consensus from fringe opinion. It presents everything in the same confident tone. That's exactly what makes it dangerous for desk research. Not because AI is always wrong, but because you can't see when it's wrong unless you already know the subject. And if you already know the subject, you didn't really need AI. ## So what does good AI use in research actually look like? Awareness, that's the key. Before you reach for an AI tool, ask yourself: what am I actually trying to achieve here, and what am I willing to trade off? If you need a participant who really doesn't want to talk to a human, AI-moderated research can be appropriate. If you need nuanced follow-up questions in the moment, you need a human. If you want to know what your existing research said, AI is useful. If you want to know what it means, do (part of) the work yourself. Both researchers and AI tools are getting better. That's a reason to keep calibrating sharply, not a reason to hand over the wheel. The best researchers keep getting better at knowing which specific tasks they can delegate, and they verify the output every single time. Feel free to use AI to summarise or cluster information, but make sure you're the one adding the nuance. Personally, I regularly let AI do the groundwork for my syntheses, the rough sorting, recognising the first patterns, while I then add the deeper meaning and context myself. That didn't work straight away, by the way. The first syntheses I had it produce were flat: every insight carried equal weight, there was no hierarchy, no sense of what was actually hitting the core of the problem. I could only notice that because I had done the interviews myself and already had the key insights in my head. I knew which insight weighed most heavily, and I could see that the AI didn't see it. For me, that was the signal to rewrite the prompts, make them more specific, and teach the AI what "important" meant in this particular study. But without that capacity for recognition, I would never have known the first synthesis was unreliable. It would simply have been the truth. The reason it works now is precisely because I collected all the underlying input myself and I understand it. ## I think, therefore I use AI It's the discipline: be conscious. Reflect before you delegate. AI is a tool, not a replacement for thinking. Research itself hasn't changed. The pressure on researchers to work faster, cheaper, and more scalable has. AI takes some of that pressure off, but only when the person using it knows exactly what they're holding. With my colleague Stephan and myself, I see that we now have more time for higher-value work. We can shift our focus and spend less time on repetitive tasks. A good example: where we used to have to split our attention between moderating and taking notes, during an interview or usability test we can now focus entirely on the interaction with the participant. My advice: use it consciously and always verify it. Make sure someone with expertise takes a look before anything it has generated gets anywhere near a decision.