There is a version of AI-assisted UX research that is genuinely useful: faster transcription, smarter tagging, synthesis that surfaces patterns a human analyst would miss. And there is a version that quietly hollows out the discipline -- synthetic users replacing real ones, auto-generated personas nobody vetted, research decks produced without a single conversation ever taking place.
The industry is drifting toward the second version.
Where AI earns its place
Transcription, thematic clustering, sentiment analysis across large datasets -- these are legitimate wins. Running a 60-minute interview through a model and getting a structured summary in minutes is a genuine productivity gain. The model is doing what it does well: processing volume and surfacing signal.
Use it here. Use it liberally.
Where the risk accumulates
Synthetic users are the biggest current risk. Models trained on internet text have no lived experience of your product or your market. They produce plausible-sounding answers that confirm whatever assumptions were baked into the prompt. The insight that matters most in UX research is almost always the surprising thing -- the workaround that reveals a deeper need. A model cannot surprise you.
AI-generated personas carry the same problem. A persona built from 'typical 35-year-old working mother in Auckland' is a stereotype dressed as research. It has no roots in observation.
The rule we apply
AI assists with everything that comes after the conversation -- transcription, tagging, synthesis, draft write-ups. The conversation itself stays human. That is where the actual learning happens.
Speed is not a substitute for validity. A research deck produced in two hours by a model is not research. It is a confident set of assumptions. Use AI to run more research, not to avoid it.