Synthetic users are the most seductive idea in research right now, and also the most misunderstood. The pitch is hard to resist. Skip the recruiting, the incentives, the scheduling, and ask an AI to role-play your customers instead. Around half of researchers expect synthetic users to be one of the biggest trends of 2026. The way they are being sold, though, has run well ahead of what the evidence supports.
So let's look at the evidence rather than the pitch. This piece covers what synthetic users are, what the 2026 research found when people tested them honestly, the narrow set of jobs they do well, and how to use them without quietly poisoning your decisions.
What synthetic users are, and what they are not
A synthetic user is an AI-generated persona, with no real person involved. You describe a customer type, and a language model invents a plausible character and answers your questions as that character. The output looks like an interview transcript. Underneath, it is fiction the model wrote to match your prompt.
People miss this distinction, and it changes how much you can trust the result. Synthetic users are a different animal from AI-moderated interviews, where an AI runs the conversation but the participant is a verified human. With AI moderation, the moderator is automated and the data is real. With synthetic users, the data itself is generated. One scales how you collect real human insight; the other does away with the human. When a vendor blurs the two, read carefully, because the reliability gap between them is enormous.
What the 2026 research found
When researchers stopped speculating and tested synthetic users against reality, the results came back consistent and unflattering.
Nielsen Norman Group ran the most cited test, comparing synthetic-user responses against three of their own existing studies with real participants. Their conclusion was blunt: synthetic-user responses for many research activities are too shallow to be useful. The answers felt one-dimensional next to the rich, contextual, sometimes contradictory insight that real people gave.
The deeper problem was not just shallowness. Synthetic users predicted real human behavior poorly, and they did it in a dangerous direction: they were sycophantic. They wanted to please. They praised concepts that real users went on to question or reject. A January 2026 issue of ACM Interactions made the structural argument, calling AI-generated participants a category mistake that undermines the purpose of research, and a separate review of synthetic-user experiments landed in the same place from the data.
It helps to understand why they fail, because the failures are baked in:
- Training-data drift. A model reflects the web text it was trained on, which is the past, not your current customer. Ask it about your 2026 product and you get a 2023 internet's idea of an answer.
- Sycophancy. Modern models are tuned to agree with the framing of the question you ask. Lead them even slightly and they will happily confirm whatever you hoped to hear.
- No capacity for genuine surprise. The value of research is the thing you did not expect: the workaround you never imagined, the objection that reframes the product. A model generating plausible answers cannot hand you that, because it can only produce what is plausible to it.
That last one is the heart of it. Real research pays off because real people surprise you, and synthetic users, by construction, cannot.
Where synthetic users genuinely help
This does not make synthetic users worthless. It gives them a narrow, real role, and that role sits before the research, not instead of it. Think of a synthetic user as a brainstorming partner rather than a data source.
Used that way, they pull their weight:
- Sharpening your questions. Run your discussion guide past a synthetic persona to catch confusing wording, dead-end questions, and obvious gaps before a real participant ever sees it.
- Generating hypotheses. Use them to rough out what you might hear, so you walk into real sessions with sharper things to listen for.
- Stimulus pre-tests and pre-mortems. Pressure-test a concept or a prototype for glaring problems before you spend real recruiting budget on it.
- Desk research and edge cases. Check your logic and surface scenarios worth investigating, especially for hard-to-reach audiences you want to prepare for.
The rule underneath all of these: synthetic users help you get ready to learn from real people. They do not do the learning for you, and they are no basis for a real decision about pricing, strategy, or what to build. As one summary put it, they are useful before the research, not in place of it.
The honest bottom line on the debate
For now, the skeptics have the better case, and it is worth saying plainly. You cannot replace real participants with AI-generated ones and trust the result. The 2026 evidence agrees on that point.
It is also worth holding the position loosely. The same researchers who warn against synthetic users today caution against outdated skepticism, because what is true this year may not hold in eighteen months as models improve. So neither dismiss them nor put faith in them. Use synthetic users for the narrow jobs they do well, demand evidence before you trust them with anything more, and keep your real decisions anchored to real people.
If you want the scale that made synthetic users tempting without losing the truth, the answer is not fake participants. It is real participants at scale, which is what AI-moderated interviews deliver: hundreds of genuine conversations, AI-moderated for depth, with actual humans on the other end.
How to use AI in research without fooling yourself
A simple test keeps you honest. Before you trust any AI-generated insight, ask: is there a real person behind this answer?
- If yes, as in an AI-moderated or voice interview with a verified participant, you are collecting real data with an automated method. Trust it the way you would trust any well-run study.
- If no, as in a synthetic user, you are looking at a model's guess. Use it to prepare, to brainstorm, to pre-test, but not to decide.
For the fundamentals of running research with real people, our guides on how to conduct effective user research and the types of user research cover the ground that no model can shortcut.
Where this fits at User Evaluation
User Evaluation is built around real participants. Its AI-moderated interviews use AI to run and analyze conversations with actual people at scale, so you get the speed and reach that make synthetic users tempting, without trading away the lived experience that makes research worth doing.
What to take away
Synthetic users are a real tool with a small, honest job: helping you prepare to learn from real people. The 2026 evidence says they are too shallow, too agreeable, and too incapable of surprise to stand in for actual participants when something is at stake. Use them to sharpen questions and test hypotheses, keep them far from real decisions, and remember the question that settles every case. Is there a real person behind this answer?