In April Nielsen Norman Group asked design and technology professionals what they call the work they do. 604 people answered. Two months earlier, Rachel Banawa had put the same open question to 120 respondents who didn't exist, generated on a platform built for the purpose, with quotas set so the fake sample would cover more job roles than the real one.
Both groups gave the same headline answer. "UX" is still the word used to describe what they do. Nothing has replaced it. If that's all you happened to need to know, the synthetic audience would have satisfied, in hardly any time, for almost nothing.
But then look at the spread. The humans used more than fifty different labels, most of them once or twice. The synthetic respondents used just twelve. None of the odd ones turned up, and nor did some obvious ones: nobody said "experience design", nobody said "consumer research". And the roles behaved differently. 93 per cent of the synthetic product managers reached for UX terms, against 44 per cent of the real ones. The machine had decided what a product manager would say and then said it, every time.
The machine does the averaging before you ask
NN/g's name for this is mode collapse: a model that keeps producing a narrow set of outputs instead of the full range it was trained on. In plain terms, it gives you the most common answer, then gives it to you again. The report's own conclusion is that synthetic data may understate the differences between groups "by pulling responses toward the most common framing", and that if the tail of the responses is what you care about, you really need actual humans.
But the tail is usually where the interesting stuff sits. Nobody commissions research to confirm that most customers are fine. You do it to find the one who isn't, the one whose problem nobody in the room has thought of, the one who uses the product in a way the brief never imagined. Research exists to find exactly that person, and that is the person a synthetic sample is built not to contain.
This isn't a fringe worry. User Interviews surveyed 150 researchers in May and 41 per cent already use synthetic users as a preliminary step before talking to anyone real. At SAP, Eric Mahlstedt's research team re-ran several of their old studies with synthetic participants and compared the answers to what the real ones had said. His verdict was "overly generic". It felt, he said, "less like conducting research and more like a sophisticated internet search". Arin Bhowmick, SAP's chief design officer, wrote that up this month with the reassurance that "nobody is using synthetic users to make the final calls". I believe him. I also think that's how it starts.
Getting the average, now faster
I wrote in July about Alex Murrell's Age of Average: fifty coffee shops from fifty cities, side by side, and you couldn't tell one from another. The same pale interiors, the same friendly blob illustrations, the same skinny typeface on every rebrand. That was people copying each other, globally, over some years.
This is the same thing with the lag removed, because the averaging happens before anyone gets a chance to ask. Cornell and Michigan researchers studied 1.5 million Change.org petitions around the platform's "write with AI" tool in late 2023. Petitions got longer and their titles converged on the same handful of verbs ("implement", "mandate", "urge") that almost nobody had used before. They did no better. The share getting even one comment fell by about five per cent. The outputs changed and the outcomes didn't, which is the title of the paper.
And at the individual level, the trade is measurable. Anil Doshi and Oliver Hauser gave 300 writers a short story task, some with AI-generated ideas to start from. The stories with AI help scored higher on novelty and were better written, especially from the less practised writers. They were also 10.7 per cent more similar to each other. Each writer got better. The set got narrower. Hauser called it an emerging social dilemma, and I think that's the right name for it.
The likely answer is the obvious one
I think what underpins this is simple. A model is built to give you the likely answer, and the likely answer is the safe one. It won't hand you the outlier, or the view that argues with your brief, or the customer who says the thing nobody in the room wants to hear. It will hand you something plausible, fluent and agreeable, which is the most dangerous kind of wrong because it looks like 'job done'.
Design has always had a version of this problem. Every organisation runs the same "best practice", tests its way to the same patterns, and ends up smooth, competent and interchangeable. But at least the old convergence left a trail. Somebody chose the pale interior. Somebody could be argued with. A synthetic sample that agrees with your assumptions leaves nothing to argue with, because there was never anyone there.
The risk is good enough work. And let's face it, good enough work is where brave decisions go to die.
What to do with the synthetic study you've already got
Keep using them for what NN/g and SAP both say they're for: hypotheses, a first pass, a check on the broad pattern. Just don't let one be the last word on anything that matters.
And the next time a synthetic study comes back, before reading the summary, ask what the strange answers were. The ones that appeared once. The ones that don't fit. If there aren't any, you haven't been given the research. You've been given the average, and you already knew that.
Sources
Rachel Banawa, "Synthetic Users vs. Human Respondents: Same Conclusion, Less Variation", Nielsen Norman Group, 16 October 2026.
Rachel Banawa, "No New Name Has Replaced 'UX'", Nielsen Norman Group, 31 July 2026.
Arin Bhowmick, "Are synthetic users really the future of UX research?", UX Collective, 8 October 2026. Source for Eric Mahlstedt's comments and for User Interviews, The State of Synthetic Users (May 2026).
Alex Murrell, "The Age of Average", 2023.
Stephen Barber, "Brand-led customer experience, in practice", July 2026.
Isabel Corpus, Mor Naaman, Allison Koenecke and Eric Gilbert, "Introducing AI to an Online Petition Platform Changed Outputs but Not Outcomes", Nature Human Behaviour, 30 September 2026. Reported in the Cornell Chronicle.
Anil R. Doshi and Oliver P. Hauser, "Generative AI enhances individual creativity but reduces the collective diversity of novel content", Science Advances 10(28), 2024. Reported in ScienceDaily.