Content Benefit

Content Strategy | UX Writing | Content Design | Globalization | Communications

AI is not your real user

AI is not your real user

Synthetic users in UX research: what they are and why the debate is heating up

Synthetic users are participants simulated by a language model. You describe a profile, say a 45-year-old purchasing manager based in Turin, and the AI answers interview or survey questions the way that person supposedly would. There is no one to recruit, no calendars to juggle and no incentives to pay: a full round of interviews wraps up in an afternoon. For teams with thin budgets and tight deadlines the pitch is tempting, so much so that one startup in the space, Synthetic Users, picked “User research. Without the users.” as its tagline.

The debate flared up again in early October with a UX Collective piece by Arin Bhowmick, Chief Design Officer at SAP. The headline asks whether synthetic users are really the future of UX research, and his answer is a flat no. I agree with him, and I will try to explain why, starting from the data.

Researchers embrace AI, but not simulated participants

In May 2026 User Interviews surveyed 150 research professionals for its report The State of Synthetic Users. 81% regularly use AI tools in their work, yet only 8% turn to simulated participants. 47% describe themselves as skeptical and want more evidence before trusting them, 17% are opposed, and enthusiasts stop at 3.3%. 21% say there is no context in which they make sense at all.

Concerns are widely shared: 88% worry about the quality of insights, and 79% about two separate risks, stakeholders putting too much faith in generated findings and bias against underrepresented groups getting amplified. The uses considered legitimate stay at the margins, such as drafting surveys and screeners (46%) or running usability tests for directional signals (34%).

What NN/g found comparing simulated and real participants

Nielsen Norman Group ran the simplest and most convincing check: it repeated three studies already carried out with real participants, this time using Synthetic Users and ChatGPT, and compared the answers. Asked to describe a typical day for a medical sales rep in Colombia, the AI did well, with an account very close to what fieldwork had captured. Once the questions moved to behavior, things changed.

In a study on online learning, real people admitted dropping courses halfway: one participant had finished three out of seven because a new job ate up her time. The synthetic participants claimed they had completed every single one. The same went for discussion forums: those who had actually tried them found them contrived and stayed away, while the simulated users described joining in eagerly and valuing the sense of community. Kate Moran and Maria Rosala sum it up neatly: real people care about some things more than others, while synthetic users seem to care about everything.

SAP’s test: closer to an internet search than an interview

Bhowmick’s article also describes an internal experiment. Eric Mahlstedt, VP of User Research & Insights at SAP, re-ran some company studies with synthetic users and got generic, plausible answers that resembled anecdotes already floating around online. In his words, it felt less like conducting research and more like a sophisticated internet search, presented in a way that creates the illusion of empathy.

I know that illusion well. When a friend recommended Grok to me as a confidant, I came away uneasy: everything was too synthetic and too fast for any real exchange to happen. I wrote about it in Loneliness, AI cashes in, and I find the same feeling in these transcripts of interviews that never took place.

Too smart and too nice: sycophancy and hyper-accuracy

The first flaw is sycophancy: language models tend to side with whoever is asking, and a 2023 Anthropic study describes it as widespread across leading AI assistants. Bhowmick cites a 2025 study in which eight designers worked with an AI persona named “Alice”: she endorsed proposals uncritically and never questioned cost or feasibility. The more senior designers called those conversations one-way and self-affirming. NN/g saw the same pattern: a drone delivery app pitched to a synthetic sales rep instantly became “very useful,” practically a game changer.

The second flaw is being too good. MeasuringU put ChatGPT through a tree test on the navigation of the IRS website: in every run it found the right path in at least nine tasks out of ten, while human participants landed at about 51% success. A tool that finds its way better than nearly every user reveals nothing about where users struggle. Back in 2023 Aher, Arriaga and Kalai named this phenomenon hyper-accuracy distortion, after observing it in recent models asked to simulate classic psychology experiments.

Everyone the same: when differences between people disappear

Then there is a subtler problem, and to me the most serious one. In 2024 a team of political scientists led by James Bisbee asked ChatGPT to answer as thousands of American voters, then compared the results with a benchmark survey, the American National Election Studies. The averages matched strikingly well, but the answers varied far less than human ones, relationships between variables often failed to hold, and the same prompt produced different results three months apart.

A study by Angelina Wang and colleagues, published in Nature Machine Intelligence, goes further. When they impersonate a demographic group, models tend to portray it the way outsiders imagine it, stereotypes included, and to flatten its internal variety. For UX research this means that minorities, people who use technology in unexpected ways and those with accessibility needs, exactly the people a good study should bring to light, get erased without anyone noticing. An average never complains.

Why a person is not a probability calculation

This is where my own conviction comes in. People do not reason by following statistical patterns, similarities and probabilities, whereas a language model works exactly that way: it produces the most plausible answer based on what it has already read. User research exists to uncover what is not plausible: the participant who drops her courses because she changed jobs, the customer who says one thing and does another, the contradiction that dismantles an assumption taken for granted.

That is why I have always argued that interfaces stay manageable when you design for personas, as I wrote in Personalization is tiring, provided those personas come from flesh-and-blood individuals and not from an average. A profile invented by AI and then interviewed by the same AI is a closed loop that confirms what was already known. The appeal of the quick method resembles the appeal of common sense, and there too caution is in order: common sense is not a method. So no, AI is not a user.

Where synthetic users can still help

None of this means throwing everything out. Both NN/g and Bhowmick acknowledge some sensible uses: desk research on an unfamiliar field before interviews, piloting a discussion guide to catch ambiguous questions, forming hypotheses and draft personas to validate later in the field. The boundary is clear: simulated answers should never be presented as findings gathered from real people, which NN/g considers unethical, and they should be avoided with niche populations that are poorly represented in training data.

The biggest risk is organizational. Where research culture is fragile, a synthetic report can become the excuse to never talk to customers, and the decision ends up with whoever is in the biggest hurry. It ties into the question of who calls the shots in design, which I covered in Designers and AI: who decides.

Time will tell if I am right

Models will improve, and someone will argue that the gap with real people is narrowing. Maybe so. Still, I remain convinced that the value of research lies in the unpredictable, and that no statistic, however refined, can capture the hesitation of someone looking for a button and not finding it. We design products for human beings, and to understand them we have to listen to them. Time will tell if I am right.

Related sources

Chiara Bonifazi