Ninety-seven percent of UX researchers use AI somewhere in their workflow. Only 8% regularly use tools that generate synthetic participants. That gap, from User Interviews' 2026 "State of Synthetic Users" report surveying 150 respondents, tells you what the profession actually thinks of the category.
I think about this through one lens: distance to first proof. Distance to first proof measures how many days until a real person uses a real version and forms an opinion about it. Synthetic users promise to collapse that distance to near zero. Simulate an interview, get feedback before lunch.
The speed is genuine. The feedback is sycophantic.
A study published in Science this year by Cheng et al. observed approximately 2,400 people interacting with an AI system. They found that AI chatbots affirmed user actions 49% more frequently than humans did. Participants exposed to affirming AI were less willing to repair relationships and became more convinced they were right.
Synthetic users produce the same effect on product teams. They confirm what you already believe, and they do it quickly enough that the confirmation feels like evidence.
Nielsen Norman Group found that a synthetic user evaluating a drone delivery concept responded enthusiastically, while real users typically offered more critical, nuanced feedback. In a separate observation, synthetic users listed seven generic factors without prioritization. Real users distinguish between essential and nice-to-have features. Products ship on that distinction.
Lewis and Sauro reviewed 12 peer-reviewed papers on synthetic users and found 14 discouraging results against 9 encouraging ones. Park et al. attempted to replicate 14 classic studies using LLMs. Only 21% succeeded. Bisbee et al. found that high-level means matched but the details collapsed: inaccurate subgroup means, small standard deviations, inaccurate regression coefficients.
The lens applies here directly. Synthetic users produce false signal between question and proof. Your team builds confidence on it, then real users arrive and the signal inverts. Time spent interpreting synthetic feedback gets spent again on real feedback. The distance doubled.
Of 150 respondents in the User Interviews survey, not a single one reported zero significant concerns about synthetic users. Only five people, 3.3%, showed genuine enthusiasm. Eighty-nine percent worry about the quality and accuracy of insights. Eighty percent worry that stakeholders will over-trust AI findings.
Meanwhile, 62.7% of organizations have no guidance on synthetic user use at all. The concern is high and the governance is absent. User Evaluation found that synthetic users tended to agree rather than push back, and endorsed concepts that actual participants later questioned or rejected. Real participants, they concluded, remain essential for consequential decisions.
I agree with the 64% of researchers who hold negative views of synthetic participants. I diverge on the framing. The debate treats sycophancy as a quality problem, something to fix with better models. Cheng's research says otherwise: users rated sycophantic AI responses as more helpful and trustworthy, despite the distorted judgment those responses produced. The sycophancy feels good because it is a structural property of models trained on human preference, built into the training loop itself.
Synthetic users can generate hypotheses and stress-test discussion guides. Those are legitimate uses. But the moment they replace the real participant in the real chair using the real product, the team has measured the distance to a fiction and called it progress. Can product teams tell the difference between confirmation and evidence when the confirmation arrives in four minutes instead of four weeks?
.webp)
.webp)
.webp)
