Synthetic users: when they help and when they mislead

What are synthetic users?

A synthetic user is a large language model prompted to role-play a research participant — "you are a 42-year-old freelance accountant who invoices eight clients a month" — and then interviewed or surveyed in character. Tools built on this idea let you define a segment, generate a panel of hundreds of personas, and get transcripts back in minutes. No recruitment, no scheduling, no incentives.

The pitch is obvious: research without the humans. And the transcripts look remarkably real.

That surface realism is what makes synthetic users worth taking seriously — and what makes them dangerous.

Where do synthetic answers actually come from?

A synthetic user's answer is a prediction of what text would plausibly follow your question, given everything the model absorbed in training — reviews, forum threads, research reports, marketing copy. Three properties fall out of that mechanism.

Synthetic users regress to the plausible median. The model produces the most representative-sounding answer for the persona you described. Real markets make money at the edges — the segment that behaves unlike the stereotype. The median is what you already believed.

Synthetic users have no lived friction. A real participant can tell you the invoice export failed last Tuesday and what they did about it. A synthetic accountant has never exported an invoice. Ask anyway and the model won't decline — it will confabulate a fluent, specific-sounding answer. That's worse than silence, because it looks like data.

Synthetic users are agreeable. Language models are tuned to be helpful, and helpfulness leaks into role-play. Synthetic participants accept your framing, find your concept "really interesting," and rarely deliver the flat "I would never pay for this" a blunt human gives you in minute four.

What does this look like on a real decision?

Maya is a product manager at a B2B invoicing tool. She has to decide whether to replace per-seat pricing with usage-based pricing — the kind of question teams are most tempted to hand to a synthetic panel, because it's about opinions rather than behavior.

A synthetic panel of "freelance accountants and small-agency owners" returns exactly what the internet believes about usage-based pricing: predictability concerns, fairness framing, a preference for caps. Coherent, well-written, and identical to what Maya's team guessed on a whiteboard in ten minutes. Nothing in it is checkably about her product or her customers.

Live interviews with recruited users surface the finding that decides the question. Several agency owners resell invoicing to their own clients and quietly mark it up — so "usage" isn't their usage, and the proposed model breaks a revenue stream Maya didn't know existed.

That's an unknown unknown: a fact nobody thought to ask about, surfaced because a real person with a real workflow was in the room. Synthetic users cannot produce unknown unknowns about your market. Everything they "know" was already written down somewhere.

A synthetic user can tell you what the internet would say about your product. Only a real user can tell you what happened last Tuesday when they tried to use it.

When do synthetic users genuinely help?

The legitimate jobs share one property: the AI stands in for a pilot participant, not a data source. You're testing your instrument, not your market.

Dry-running a discussion guide. Run your guide against a few synthetic participants before any real session. You'll catch double-barreled questions, jargon your audience won't share, leading phrasings, and a guide fifteen minutes too long — at zero cost to your recruit pool. This is the single best use of synthetic users.

Rehearsing analysis. If you're building a synthesis workflow — tagging, affinity mapping, AI-assisted analysis — synthetic transcripts are free, safe test data. No consent issues, no burned participants.

Brainstorming hypotheses. A synthetic panel quickly enumerates the conventional objections to a concept — a checklist of what to probe with real people. A list of questions, never a list of findings.

Training interviewers. Practicing probes and follow-ups against a synthetic participant is a flight simulator: useful practice, nobody confuses it with a real flight.

Synthetic users vs. real participants: what changes?

The punchline: synthetic users win every operational row and lose every evidential one. The right question was never "synthetic or real?" It's "which step of research was actually slow?"

Why is "AI participants" the wrong place to put the AI?

The demand behind synthetic users is legitimate. Traditional research is slow — recruit for a week, schedule across three more, moderate one session at a time, synthesize by hand.

Synthetic users fix that by replacing the participants. But the participants were never the bottleneck. They were the point — the only element of a study that contains information you don't already have. The bottleneck was everything around them: authoring, scheduling, moderation, synthesis.

Those steps are where AI works without corrupting the evidence. This is how Sera is built: AI drafts the study — goals, discussion guide, screener — from a URL or Figma link in about two minutes; AI moderators run voice interviews with real recruited participants in parallel rather than one per calendar slot; synthesis runs automatically over full transcripts you can audit. Real-human research in under 24 hours — roughly the turnaround synthetic panels promise, with actual people on the other end.

AI moderation has its own failure modes — emotional subtext, the brilliant off-script tangent — and we've written about them honestly in Can you trust AI-moderated research?. But the asymmetry matters. A moderation error costs one interview's depth, and the transcript shows you where. A synthetic panel's error is invisible: fluent, confident findings about a market that was never consulted, with nothing to audit.

A simple rule for deciding

Before using synthetic output, ask one question: will this change a decision?

• No — it's rehearsal. Piloting a guide, testing an analysis pipeline, training a teammate, listing hypotheses. Use synthetic users freely.
• Yes — it's evidence. Pricing, positioning, build-or-kill, redesigns. Use real participants, and label anything synthetic so it can't launder itself into a readout as a finding.

The teams that get burned aren't the ones who use synthetic users. They're the ones who let a pilot tool drift into an evidence tool because the transcripts looked real and the deadline was close.

Hold that one line and synthetic users earn their place in the kit: a flight simulator, clearly labeled, parked next to the actual plane.

Frequently asked questions

What are synthetic users in UX research?

Synthetic users are AI-generated research participants: large language models prompted to role-play a target customer and answer interview or survey questions in character. Vendors position them as a faster, cheaper substitute for recruiting real people. The output is plausible text about what such a person might say, not a record of what anyone actually did.

Are synthetic users accurate?

Synthetic users are accurate about generic, well-documented behavior and systematically wrong about the specifics that decide product questions. They reproduce the median opinion in their training data, over-agree with the premise of your questions, and cannot report real friction with your product, because they have never used it.

Can synthetic users replace real research participants?

No — not for decisions. Synthetic users can replace pilot participants when you are testing whether your questions make sense, and they can generate hypotheses worth checking. Any finding that will change what you build, price, or ship needs real participants, because synthetic findings cannot surprise you in the ways real markets do.

What are synthetic users actually good for?

Three jobs: dry-running a discussion guide to catch confusing or leading questions before real sessions, rehearsing your analysis pipeline on realistic-looking transcripts, and brainstorming hypotheses to test properly. In each case the synthetic output is scaffolding for a real study, never the study itself.

How are synthetic users different from AI personas?

An AI persona is a summary document — a description of a customer type, sometimes generated from real interview data. A synthetic user is that persona made interactive: an AI answering questions in character. The persona summarizes evidence you collected; the synthetic user manufactures new "evidence" that nobody ever said.

Do synthetic users save money on research?

Synthetic users remove recruitment and incentive costs, which is real money. But if a synthetic finding steers a decision wrong, the cost is a shipped mistake — usually orders of magnitude more than the recruitment you skipped. AI-moderated research with real participants captures most of the speed savings without giving up ground truth.