July 3, 2025
I’m firmly curious and exploratory when it comes to AI, but as a researcher of nearly 20 years, I’m a healthy sceptic about entirely outsourcing this necessarily human-centred discipline to GenAI – as we all should be. User research is a craft, a deeply human discipline demanding expertise in ethics, data management, bias and human behaviour that AI cannot yet lead on. But this doesn’t mean that we can’t use LLMs as our research assistant.
AI tools can synthesise vast amounts of data, and even help automate routine tasks. However, what AI can’t (yet…) replicate is the richness of context-human experiences that happen when people try to engage with the products and services we design. This is best understood through contextual, observational research.
I reflect on the client research I’ve done over the past few years and how valuable it’s been to understand the context the users of our solutions are operating in, such as the use of technology in prisons and the dynamics when digital solutions meet physical space (which we explored when designing Click and Collect at Pret), In most situations, user research will still rely on empathetic engagement, using our human eyes to observe what people do and translate that into insight.
I can’t count the number of times a user has told me they like to do something in a certain way during the preamble of a research session, only to then do it in a totally different way in practice. Researchers are experts at interpreting human behaviour and recording it as insight. An LLM is a large language model after all. It can do an adequate job of summarising the words that people say in an interview, but there’s no chance right now of using it successfully to run or interpret the outputs of a usability test where the words might be accompanied by nuances in body language. An uncomfortable shifting in the seat, a pause, or a hesitation can indicate something else about the user’s experience of the product, and that also needs to be observed and used in decision making.
When I first heard about synthetic user data at South by South West (SXSW) conference in 2023, I admit I rolled my eyes. But now, after working on niche government services where users can be hard to recruit for research, I’ve realised synthetic data has potential – but only with rigorous foundations. Synthetic data models first require quality seed data, derived from thoughtful, inclusive research. So you will still need to do some actual research with users for this to be a successful exercise.
Relying solely on AI-generated or LLM-derived data also risks perpetuating biases and marginalising minority voices who are not well-represented in language models. Good synthetic data starts with good user research.
There’s a lot of focus on the use of AI to improve speed and productivity, but one essential contribution user-centred design folks make is building empathy within delivery teams. Nothing creates change like sitting a sceptical stakeholder or engineer with a user and watching them struggle through a feature. What will be lost if we only test with synthetic personas? Ensuring user-centricity is part of the DNA of the team still demands a human touch.
Research ethics aren’t optional; they’re foundational. When we collect any data from users through interviews or observations, we need to obtain informed consent to collect and process that data under the General Data Protection Regulation (GDPR). As user research becomes more democratised, I’ve grown increasingly concerned about the accidental sharing of Personally Identifiable Information (PII) or company confidential information with AI systems. Simply anonymising data as “Participant A” isn’t enough – details like a job role or company affiliation can inadvertently identify individuals. This matters even if you don’t think you’re researching a sensitive topic. Helpfully, the Market Research Society, which some of our clients partner with, recently issued clear guidelines including:
It’s tempting to quickly feed transcripts into an AI or rely solely on desk research generated by an LLM. But caution is essential. We all know LLM analysis is prone to overstating insights, omitting things and making things up. AI hallucinations – those convincingly incorrect outputs – underscore the importance of human oversight and critical judgment when running research analysis. If you don’t know your material, how can you validate that it’s actually reliable insight?
If you choose to analyse with an LLM, you can mitigate this by manually analysing some of your transcripts and, at a minimum, reading your transcripts before you analyse them. Treat AI-generated insights like feedback from a junior researcher – probe, question, and validate carefully.
In the past, our design service focused on providing clients with deep specialists in various design roles – content designers, UX designers, service designers and researchers, but I can see moves now towards AI-powered generalist UX roles. In order to stay relevant, user-centred design people may need to apply our skills as a curious generalist, blending analytical rigour with adaptability, great prompting skills and most importantly, knowing what we don’t know, and when to lean on a human expert for help.
AI will augment rather than eliminate research and design roles, enabling researchers to focus more on observation, interpretation, strategy, and storytelling. AI in user research isn’t about replacement – it’s about enhancement. It’s our research assistant and collaborator, not our guide.
All business models have their pros and cons. But, when you consider the type of problems we help our clients to solve at Equal Experts, it’s worth thinking about the level of experience and the best consultancy approach to solve them.
If you’d like to find out more about working with us – get in touch. We’d love to hear from you.
