Why thinking that AI can do all your user research is not a peachy idea

Proper user researchers do proper research reporting valid, evidence-based, relevant and actionable insights to inform digital program decisions. What’s wrong with that?

Well, the trouble is, there’s an annoyingly persistent viewpoint that anyone can do useful user research without any specialist training or experience — especially if they can use AI. After all, research is only about asking a few punters some questions and then reporting what they say. Or watching them try to make sense of some digital prototype and noting their struggles. And, in fact, AI can do most if not all of that, reducing human effort and project costs. Peachy.

Well, it’s only peachy if you subscribe to a highly diluted idea of what research is and what it does. This, of course, relies on the assumption that automating research will reduce human effort and costs involved.

This article focuses on the methodological limits of substituting AI for a researcher — or of relying on a non-researcher equipped with AI. I do not attempt to assess every AI-enabled research tool or use case. What I do focus on is where human contextual judgement and accountability lie and should remain decisive.

The argument I make here is not that AI has no place in user research. Quite simply, AI cannot replace the situated judgement and experience of a professional user researcher. In short, AI cannot replace an experienced researcher’s ability to recognise, make sense of, interpret and apply knowledge of the research contexts to deliver research that has reliability, validity, generalisability and credibility/objectivity.

So how does that work?

User research — like any other kind of professional research — is done in context, the values and attributes of which directly influence human perception, actions, responses and outcomes. That maxim applies to both research participants and the researcher, incidentally.

Two interdependent contexts shape all research events. One is the context that research participants and the researcher bring with them: emotional state, learned experience, beliefs and attitudes and so on, all of which influence behaviour in often obscured ways. The other is the context in which the research itself is being done; the organisational ethos or politic for instance, the stakeholder’s understanding of their world, the project imperatives and so on. It is research design, facilitation, analysis, and evaluation — even down to the words used to write up research findings — that connect these contexts, determining the value of evidence to the organisation’s project.

What is explicitly understood by the researcher is that all these contexts can and will have an influence on research outputs. So, any research event — a user testing session, a discovery interview, card sort, whatever — is consequential to its context.

A proper researcher builds this into the research design (e.g., participant profiling and recruitment, accessibility planning, and instrument development), analytic methods and reporting strategy. How? By layering contextual understanding.

These layers of context are visualised in this ‘Research Onion’ diagram, based on an original concept created by Mark Saunders et al, (2015), and adapted to User Research. Each layer represents an incremental step in evolving the optimal research event to deliver what’s of most value to the organisation.

Each layer asks a contextually-related question with the answer steering the researcher in a certain direction towards the next, ultimately resulting in a research strategy that aligns with known contexts. For instance, ‘research paradigm’ asks: what is the organisation’s view of the world — its accepted ideas, methods, rules and priorities, ethics and principles, and so on. ‘Approach to generating new knowledge’ asks what kind of new knowledge does the organisation prize most. The answers to both drive a consideration of the most appropriate research method — generative (discovery) or formative (prototype testing) and so on. And that leads into ‘strategy’ and a consideration of the users — our research participants. Who are they? Where are they? What’s their everyday context? What are we hoping to learn from them — and why — what will this do? And then onto selection of data type — quantitative, qualitative, mixed etc.

The researcher may conclude a need for two or more complimentary lines of research — discovery interviews and quantitative date from an online self-report survey. There’s much value to be discovered by comparing and contrasting two different research methods applied to the same set of research questions. A far richer and more robust basis for interpretation, for example, reflecting a carefully designed and integrated study to address a complexity of contextual phenomena.

It’s a game of chess. You move the pieces based on understanding of the moment’s context and all the actors involved, you work with the consequences.

Research done well yields formidable knowledge that is reliable, valid, generalisable and credible/objective — the Four Pillars of Quality Research. The researcher has applied due diligence to the contexts in which the research is being done, and — particularly important when working in social or interpretive research (as 99.9% of user research is) — the researcher takes a reflexive approach* to their work.

The reverse is equally true: research done without due consideration of contextual factors and impact is just painting by numbers. (I’m not saying there’s anything wrong with painting by numbers. I mean, if I want to paint a picture fit for hanging on a wall, painting little numbered spaces is a lot easier and more achievable than trying to create a picture from scratch, even copying from a photo. Or, even easier, I could ask an AI agent to create me one. Peachy.)

But ‘research by numbers’ is not proper research.

So, am I suggesting that using AI in user research is akin to doing research by numbers? Yes and no. There’s a difference between a researcher using AI to support their research, and any Tom, Dick or Harry using AI to do research.

I’m quite a fan of using AI to support some parts of my research work. I use AI to accelerate working with quantitative data, exploring possible data visualisations, locating participant quotes to support a particular point (which, of course, must be verified by the researcher), and so on. My situated understanding of the organisational, stakeholder, user and project contexts shapes the research instrument, and how data is analysed and presented for maximum impact/value. With due diligence and governance, large scale survey datasets can be cleaned and analysed more quickly. But — and it’s a bit ‘but’ — accountability for anonymisation, validity, reliability and interpretive credibility/objectivity remains with the researcher.

AI can be used to speed the identification of common phrases, words, sentiments in qualitative datasets, and to groups strings into categories that the researcher wants to test. AI can be used to help focus clear and critical thinking; it can be your bouncing board to challenge interpretations and ideas.

Another example of the utility of AI in user research is its application in mass online self-reporting surveys or asynchronous user testing. For instance, AI can be cleverly applied to create pseudo-synchronous follow-up probes in individual testing instances so the researcher can acquire greater contextual insight into individual user behaviour and attitudes.

So yes, AI, used carefully, critically, and with purpose can bring added value to the work of the researcher.

But I draw the line at using AI (solely or in the hands of an amateur) to analyse — that is, interpret — findings from any form of research.

Bottom line:

1. Context is consequentially important to any user research and its deliverables.

2. The researcher’s job is to discover and understand, be acute to, and use this knowledge to inform the research that we do and the deliverables we produce.

3. AI brings a context to the research (as a ‘research participant’).

4. There is absolutely no way of understanding any context that the AI is ‘speaking from’, other than ‘sources’ reported by AI.

5. AI is not reliable — it makes mistakes. As do we all.

6. To rely wholly or mostly on AI to ‘do research’ is to invite, at best, a number-cruncher to generate some numbers or, at worst, a wholly misleading fairy-tale account of humans, what we do, how we do it, and why.

• A reflexive approach to research is where the researcher adopts an explicitly critical approach to how their own personal context of biases, knowledge, emotional state, personal motivations and so on could affect the research data and deliverables.

Lesley Crane PhD is a professional researcher specialising in user experience in digital contexts.

Why thinking that AI can do all your user research is not a peachy idea was originally published in Bootcamp on Medium, where people are continuing the conversation by highlighting and responding to this story.