AI and user research

This page is for observations/ideas/concerns etc about AI in the context of user research. There are two obvious versions of this: Using user research to study topics related to AI, and using AI as a tool to conduct user research.

AI-related research topics

First, AI is an interesting topic for user research, as it strongly intersects with psychological factors - attitudes towards it, and effects it has on users. Plenty of academic literature exists that provide models of attitudes towards AI, in which components could be theoretically relevant predictors of variables of interest, or be mapped out over an organisation. General models such as the three-component trusting beliefs models can be translated to AI, and there are dedicated models (with associated scales) for AI, such as Attitudes to AI at Work.

Measuring how AI affects users is perhaps primarily an issue of finding theoretically rich, ideally validated measured of the effects of interest, e.g., around job satisfaction or job stress. The question on the AI side of things may be - what do you mean by AI, and how will users understand technical terms of interest - how granular do you need to be?

AI as a research tool

The second angle on AI and user research is, of course, its use as a tool. My experience has been mainly using LLM-based models as a way to do thematic analysis on transcripts from semi-structured interviews. With suitable prompts, honestly to my surprise, I couldn't find much fault with the "bottom-line" output. That is: A slide with the results of the analysis wouldn't look much different to a human doing it, and if it did I'd not automatically assume the human had done something more useful.

One problem I nevertheless have is about the next step down the road. By not doing your own analysis, the information has flowed past rather than through your head. There's no "side-effect" of you, in your own brain, having built an understanding of the context you're studying. Maybe the immediately, bullet point-level information on the slide deck doesn't change, but ideally there should be more than that. The bullet points shouldn't be all there is - they should be the tip of an iceberg. You'd want to use the gained insight beyond just making that slide, in a strategic way, in a way that makes further decisions deeply informed by human needs and influences. That capability-building I felt got almost completely lost. I knew as much as anyone else looking at a slide with bullet points.

Another problem is from a broader societal, strategic, and ethical perspective. Using an AI model for your analysis means being dependent on the company owning and providing access to that models; and therefore also on the decisions and interests make by people connected to that company. AI models are never neutral, obviously, and they're not generally transparent. Who knows what, perhaps subtle, perhaps even insidious, biases have been built in? Maybe not even intentionally! And even if there's no obvious case of that, your (relative) independence is completely compromised. That is: One thing I like about user research is that you can be a "helping outsider" - ideally, the value you're bringing is about finding things out that are important even if they're not necessarily convenient to corporate processes in the short term. Using AI seems to me to almost unavoidably means losing that slight but important distance from the overall corporate world - it's in the heart of your analysis.

That said, those are subtle philosophical points and from a pragmatic perspective there might not be much of an option sometimes. AI-based analyses are a way to get to a reasonable mediocre output very quickly if you need that. Where I think a better workflow lies is in using AI specifically not to make it easier to get past the metaphorical blank page, but as an independent critic after you've done the work. The only problem I could in any case see myself having if having that final step make me less likely to try to refine analyses and interpretation to the same standard, since it's being checked anyway. But that's perhaps a discipline that can be maintained to get the best of both worlds.