A synthetic focus group is a simulated group discussion in which AI personas take the role of participants and an AI moderator follows a discussion guide. Synthetic focus groups are fast and cheap but not reliably accurate. Studies show AI participants converge on the majority view, invent opinions, and answer more politely than real people, so findings need validation with real participants.
Summary:
|
Synthetic focus groups promise a full group discussion in minutes, with no recruiting, scheduling, or incentives. The appeal is obvious, but a group discussion is not a stack of individual answers. When AI participants talk to each other, new failure modes appear.
This guide explains what a synthetic focus group is, how it works, and what research says about its accuracy. It also shows where AI group discussions break down and how to validate their output. Teams working in Consumer Insights can then decide where a simulated group fits and where only real people will do.
What Is a Synthetic Focus Group?
A synthetic focus group is a simulated group discussion in which AI personas take the role of participants and an AI moderator follows a discussion guide. Synthetic focus groups are fast and cheap but not reliably accurate. Studies show AI participants converge on the majority view, invent opinions, and answer more politely than real people, so findings need validation with real participants.
The output looks like a real session: transcripts, themes, and summaries. The difference is that no real participant took part. Every opinion in an LLM focus group comes from a language model playing a role.
The method belongs to the wider family of synthetic audience testing, where AI stands in for people to explore ideas early. Treat it as a way to generate hypotheses, not as consumer evidence.
Synthetic vs AI-Moderated vs Virtual Focus Groups
Three formats are often confused:
Synthetic: AI participants, led by an AI moderator. No real people take part.
AI-moderated: real participants, led by an AI moderator instead of a human.
Virtual: real participants in an online session, the format covered in this guide to the virtual focus group.
Because "AI focus group" can mean either of the first two, check whether real people took part before you read any result. The label alone tells you nothing about evidence quality.
Real-participant sessions are also changing as AI takes over tasks such as theme coding. If you want AI help without losing real voices, see how AI-assisted focus groups with real people work.
The traditional format still matters as the baseline for comparison. Well-run focus group discussions with real people are what any synthetic result should eventually be checked against.
How Synthetic Focus Groups Work
Most platforms follow the same four steps:
Define the personas. Set who is in the room, from demographics to attitudes and brand use.
Load the discussion guide. The guide holds the questions, probes, and stimuli.
Run the moderated discussion. The moderator agent poses questions and manages turns.
Analyze the transcript. The system codes themes and writes a summary.
Each persona answers the moderator and also responds to the other personas' turns. That feedback loop is what separates a group from a set of separate interviews, and it is where most of the risks begin.
The analysis step deserves caution too. A summary can read confidently even when the transcript behind it was shaped by the first speaker or by the majority. Always read the transcript, not only the summary.
Building the Participant Personas
Personas come in two kinds. Demographic personas are built from a short prompt, such as a 34-year-old parent who buys organic snacks. Grounded personas are built from real customer research, such as interview transcripts or survey data.
Grounding helps. In a Stanford study of 1,052 people, agents built from two-hour interviews matched participants' survey answers about 85% as accurately as the participants matched their own answers two weeks later. They also beat agents built only from demographics. The catch is that grounding requires real people first.
The group's mix matters too: segments, attitudes, and brand users all shape the debate. Personas are only as good as the data behind them, a point explored in this guide to synthetic data vs real data.
Moderating the Discussion
The moderator agent handles turn-taking, probing, and showing stimuli. It works from a discussion guide built on strong focus group questions that move from broad to specific.
Stimuli are a weak spot. A persona reads a text description of a pack or concept, while real participants see and handle the thing itself. Good stimulus design matters most when real people are in the room.
Three settings change results more than most teams expect:
Speaking order: who talks first sets the tone for everyone after.
Group size: more voices can mean more pressure to agree.
Number of rounds: longer discussions give more room for drift.
Synthetic Focus Groups vs Synthetic Interviews
Synthetic interviews collect each persona's answer independently. Synthetic focus groups let personas read and respond to one another. That single difference changes the risk profile.
An independent answer can be wrong, but it cannot be bent by another AI participant. In a group, one persona's output becomes the next persona's input. Errors compound, majority views gain weight, and the group format adds risks that one-to-one simulation does not have.
Are Synthetic Focus Groups Accurate? What the Studies Show
The direct evidence is thin. Zhang and colleagues presented a tool called Focus Agent at the ACM IVA conference in 2024. They compared five real focus groups, with 23 participants in total, against simulated discussions on the same topic. Humans produced 39 unique codes and the simulations 47, and most human opinions also appeared in the AI discussions. The paper reports no similarity score, so "similar" is a qualitative judgment from one small study.
Research on single synthetic respondents also applies to group members. In a study of synthetic survey data in Political Analysis, synthetic averages landed close to real ones, but responses varied far less than human data. Across 22 regressions, 48% of coefficients differed significantly from the real survey, and the sign flipped in 32% of those cases.
The pattern is consistent: averages look reasonable, but the spread is too narrow and some groups are caricatured. A group of such participants inherits those weaknesses before any group effects begin.
For a researcher, this matters because focus groups exist to surface the range of views, including the awkward outliers. A method that compresses that range removes the very thing the session was meant to find.
Simulated Participants Behave Differently From People
A 2026 ACL paper by researchers at UC Irvine and Cohere compared simulated users with real ones in conversations with AI agents. The setting was agent evaluation, not focus groups, but the behavior gaps carry over to any LLM focus group.
Politeness: simulated users used politeness markers in 39.2% of turns, against 19.9% for real people.
Questions: simulated users asked questions in 18.8% of turns, about twice the 9.8% rate of real users.
Representation: simulated users were least accurate stand-ins for African American Vernacular English speakers and Indian participants.
The takeaway for synthetic interviews and groups alike is that transcripts will read smoother and more cooperative than real ones. The audiences represented worst are often the niche audiences researchers most need to hear.
Where Synthetic Group Discussions Break Down
Group discussions add risks that come from AI participants influencing each other. These matter more in a group than in single synthetic interviews because every flaw gets amplified by the next speaker.
Conformity and Majority Pull
A study of group conformity in Findings of ACL 2025 ran more than 2,500 simulated debates. When the proponent side held a 2:1 majority, neutral agents sided with it 63.53% of the time. When that side was the minority, the figure fell to 39.40%.
More capable models also swayed others more than larger groups did. Majority agents grew more extreme as discussions went on, a form of group polarization. Minority agents often declared full agreement, a pattern that resembles the spiral of silence in human groups.
Confabulation and Persona Drift
Research from the University of Illinois on persona inconstancy tracked multi-agent debates. In reflections after the debate, 15.59% of the opinions agents gave appeared nowhere in their earlier answers or in the discussion. That is confabulation: an opinion invented with no source.
Agents also took on another participant's persona in 3.12% of debate messages, despite instructions to keep their own. The authors found that prompts urging agents to defend their own views made instability worse.
Agreeable Participants and No Real Reaction to Stimuli
Nielsen Norman Group tested AI participants against real users in 2024 and found the feedback often favorable and vague. In one example, a real learner had finished three of seven courses, while the synthetic user claimed to have finished them all. As the authors put it, AI loves to please, so nearly every idea sounds like a good one.
There is also a hard limit. AI personas cannot see, hold, taste, or use a product, so their reactions to concepts and packs are guesses. The same logic runs through the risks of synthetic users in UX work.
Synthetic vs Real Focus Groups: Key Differences
Dimension | Synthetic focus group | Real focus group |
Participants | AI personas | Real people recruited to a brief |
Time to run | Minutes to hours | Days to weeks, including recruitment |
Cost | Low per run | Higher: recruitment, incentives, moderation |
Group dynamics | Simulated, prone to conformity | Natural social influence |
Range of opinions | Narrower, pulled to the majority | Wider, including outliers |
Reaction to stimuli | Described from text | Real reaction to the actual stimulus |
Emotional and nonverbal data | None, words only | Expressions, hesitation, attention |
Reproducibility | Shifts with model version and prompt | Shifts with sample, but the sample is real |
Use as evidence | Hypotheses only | Consumer evidence |
The speed and cost advantage is real. The evidence value is not the same, and the table shows why the two should not be mixed up in a research readout.
When to Use a Synthetic Focus Group
Synthetic groups work best as a thinking aid before real research:
Piloting discussion guides: find confusing questions before fieldwork.
Generating hypotheses and probes: surface angles you may not have considered.
Anticipating objections: draft likely pushback to test later.
Narrowing concepts: cut a long list to a shortlist for real concept testing.
Avoid them for final concept or claim decisions, sensitive topics, niche audiences, and anything presented as consumer evidence. Any AI focus group tool you use should label its outputs as synthetic.
A practical rule helps: if a wrong answer would cost real money or reach a customer, validate it first. The cost of a pilot is small, and the cost of a launch built on an invented opinion is not.
For a wider view of where simulated respondents fit in a research program, see when to use synthetic audiences across methods.
If you are comparing tools, our guide to consumer research platforms covers the main options.
How to Run a Synthetic Focus Group Responsibly
Good practice limits the conformity and invention described above:
Collect private answers first. Record each persona's view before the discussion and again in private afterward, to separate real views from conformity.
Rotate speaking order and vary group composition. Run several groups instead of one.
Ground personas in real research. Use customer data, not only demographic prompts.
Record the model version. Results shift between versions, so you need the log to reproduce them.
Label every output as synthetic. Readers should never mistake it for field data.
Validating Synthetic Focus Group Findings With Real People
The safest framework has two stages:
Stage 1, generate. Run several synthetic groups to produce candidate themes, objections, and probes.
Stage 2, validate. Test those themes with real participants, then compare the two sets of results.
Sort the comparison into three buckets: themes that appear in both, themes that appear only in the synthetic groups, and themes only real people raised. The first bucket is promising, the second needs scrutiny, and the third shows what the simulation missed.
Over several projects, the third bucket becomes a useful record. If real people keep raising the same kinds of themes that the simulation misses, you learn where your synthetic setup is weakest and can adjust personas or guides accordingly.
This is where the choice of real method matters. The trade-offs are laid out in this comparison of AI-moderated interviews vs focus groups.
Teams also need a repeatable process for validation with real participants so that no synthetic theme reaches a decision untested.
What Real Participants Show That Synthetic Groups Cannot
Real participants show emotional reactions, hesitation, and attention to a stimulus. Synthetic personas can only describe these in words, and a described reaction is not an observed one.
Facial coding reads expressions as they happen, so a pause or a flicker of confusion becomes data. Reading these signals alongside what people say catches the gaps that polite answers hide.
Attention matters just as much. Eye tracking shows what people looked at before they spoke, so you can see which part of a concept or pack earned attention and which was ignored.
Both signals come from the same session, which lets researchers compare what a participant said with what they did. That is the check a simulated group cannot offer, and it is the second stage of the validation framework above.
Decode supports validation studies with real participants, with 90%+ facial coding accuracy across 62 facial expressions, 96% eye tracking accuracy, and 70+ languages supported.
Multi-language support matters when you test the same themes across markets, because politeness and expression norms differ by culture.
A consumer insights platform that combines these signals lets you run validation and reporting in one place.
Frequently Asked Questions
1. What is a synthetic focus group?
A synthetic focus group is a simulated group discussion in which AI personas take the role of participants and an AI moderator follows a discussion guide. It produces transcripts, themes, and summaries without any real participants.
2. Are synthetic focus groups accurate?
Not reliably. The direct evidence is thin, and research on single synthetic respondents shows lower variance and stronger stereotyping than real data. Multi-agent studies add conformity, invented opinions, and excess politeness.
3. How do synthetic focus groups work?
You define personas, load a discussion guide, run a moderated discussion in which personas respond to each other, then analyze the transcript. Settings such as speaking order and group size change the outcome.
4. What is the difference between a synthetic focus group and an AI-moderated focus group?
A synthetic group uses AI participants. An AI-moderated group uses real participants led by an AI moderator. Only the second produces evidence from real people.
5. Why do AI participants agree with each other in synthetic focus groups?
Language models lean toward agreeable answers, and in group settings they tend to side with the majority or with more capable agents. Studies of multi-agent debates measure this conformity directly.
6. Can synthetic focus groups replace real focus groups?
No. They can speed up early thinking, but they cannot show real reactions to stimuli or provide emotional and nonverbal data, and their findings need validation with real participants.
7. What is the difference between synthetic focus groups and synthetic interviews?
Synthetic interviews collect each persona's answer independently. Synthetic focus groups let personas influence each other, which adds group-specific risks such as conformity.
8. When should you use a synthetic focus group?
Use one to pilot discussion guides, generate hypotheses and probes, anticipate objections, and shortlist concepts. Avoid them for final decisions, sensitive topics, and niche audiences.
Test What Synthetic Groups Suggest With Real People
Synthetic groups can point you toward good questions. Real participants tell you whether the answers hold, including the emotional and attention responses that AI personas cannot produce.
Decode is trusted by 150+ global brands and backed by 17 patents, with 90%+ facial coding accuracy, 96% eye tracking accuracy, and 70+ languages supported.
