Synthetic users are AI-generated simulations of research participants, typically built on large language models and conditioned with demographic, behavioral, or first-party data. Researchers interview, survey, or test them as stand-ins for real people to explore hypotheses, refine study design, and gather directional signals quickly. Because they predict plausible responses rather than lived experience, findings should be validated with real participants.
Summary:
|
UX researchers, product teams, and market researchers are all under pressure to deliver answers faster, and AI-generated participants look like a shortcut. Type a question, get a thoughtful reply, skip the recruiting. The appeal is real, and so are the risks.
This guide explains what synthetic users are, how they are built, where they fit in traditional user experience testing and in consumer research, how accurate they are, and how to validate what they tell you.
What Are Synthetic Users?
Synthetic users are AI-generated simulations of research participants, typically built on large language models and conditioned with demographic, behavioral, or first-party data. Researchers interview, survey, or test them as stand-ins for real people to explore hypotheses, refine study design, and gather directional signals quickly. Because they predict plausible responses rather than lived experience, findings should be validated with real participants.
Three things set them apart from tools you may already know:
Scripted bots follow fixed branches and cannot respond to a question nobody wrote in advance.
General chatbots answer as a helpful assistant, not as a specific type of customer.
Traditional personas are static documents. A synthetic user can be asked a new question, which is the main difference.
One clarification avoids confusion. "Synthetic user monitoring" in software performance testing uses scripted bots to check that a website loads and a checkout works. It has nothing to do with simulating research participants.
Synthetic Users vs Synthetic Personas vs Synthetic Respondents vs Digital Twins
The vocabulary overlaps, and vendors use the terms loosely. This is a practical way to separate them:
Term | What it is |
Synthetic user | An individual simulated participant |
Synthetic persona | An archetype representing a segment |
Synthetic respondent | A simulated survey answerer |
Synthetic panel | A group of synthetic users run through the same study |
Digital twin | A model of a specific real customer, built from that person's data |
AI personas and AI-generated personas are the broad, informal labels. They can describe anything from a one-line prompt to a carefully conditioned model, so ask what data sits behind the name.
What Is Synthetic Research?
Synthetic research is the practice of running studies on simulated participants instead of, or alongside, recruited humans. It spans UX research and market research, including the work teams do to understand consumer insights. In academic work, the approach is often called silicon sampling: conditioning a language model on real survey backstories to generate a virtual population.
Synthetic research is also different from synthetic data used to train AI models. Training data is generated so a model can learn. Synthetic research generates responses so a researcher can learn. Our guide to synthetic data versus real data explains the overlap and where the two diverge.
How Synthetic Users Are Created
Building a synthetic user follows four core steps:
Define the audience. Specify who the participant represents and what they should know.
Condition the model. Provide demographic, behavioral, and attitudinal context so the model answers as that person.
Run the study. Interview, survey, or show stimuli to the synthetic user or panel.
Analyze the outputs. Read responses for patterns, gaps, and hypotheses, as you would with human data.
The quality of step two decides almost everything. A model's general training data reflects the internet. First-party data reflects your customers. The more the responses draw on the second, the less generic they become.
Data Sources Behind Synthetic Users
Teams generally choose from three levels of grounding:
Generic LLM knowledge: Fast and free, but it returns averages and common stereotypes.
Demographic and psychographic profiles: Add structure such as age, region, values, and habits, though labels alone say little about reasoning.
Proprietary customer data: CRM records, past interviews, survey results, and analytics. This is the richest source, and a research repository makes it far easier to use.
Grounding in real research data improves relevance for a simple reason. Without it, the model fills gaps from training data that may not represent your audience. Stanford researchers built a dataset covering 60 US demographic groups and found a major gap between what language models say and what those groups think. Groups such as adults aged 65 and older and widowed people were poorly reflected.
Persona Conditioning and Multi-Agent Simulation
Persona conditioning is the craft of shaping how a synthetic user behaves. Designers often use a personality model such as the Big Five (also called OCEAN), then add memory of earlier answers and continuity of context so the persona stays consistent across a long interview.
Multi-agent setups go further. Several conditioned personas answer the same prompt, or respond to each other, which is how synthetic panels and simulated group discussions are produced.
The strongest evidence for conditioning on real interviews comes from Stanford. A team interviewed 1,052 people for two hours each and built an agent for every participant from the transcript. On the General Social Survey, the agents reproduced participants' answers at 85% of the accuracy people achieved when repeating their own answers two weeks later.
That figure is normalized against each person's own consistency, not raw accuracy. It also measures survey attitudes, so it should not be read as proof that synthetic users predict behavior.
Types of Synthetic Research Methods
Each traditional method has a synthetic equivalent, and each has a ceiling:
Traditional method | Synthetic equivalent | What it can produce | What it cannot produce |
Interviews | Synthetic interviews | Hypotheses, vocabulary, rehearsal | Real stories and surprise |
Surveys and panels | Synthetic respondents and panels | Pretesting, directional splits | True population distributions |
Focus groups | Synthetic focus groups | Early reaction contrasts | Genuine social influence |
Usability and concept testing | Synthetic testing | Opinions, heuristic-style reviews | Observed behavior |
Synthetic Interviews
Synthetic interviews cover problem exploration and discovery. A researcher asks a synthetic user about a task, a frustration, or a purchase and reads the reasoning that comes back. That reasoning is a hypothesis, not a finding.
Their best use is rehearsal. Running a discussion guide against a synthetic user exposes leading questions and awkward flow before a human session. This makes real interviews better, and user interviews remain where discovery actually happens.
Synthetic Surveys and Panels
Synthetic respondents are useful for survey pretesting, screener design, and simulated quantitative panels. Running a draft questionnaire through a synthetic audience testing workflow can reveal ambiguous wording and broken logic before launch.
The main risk is low response variance. Real populations disagree in messy ways, while simulated ones tend to cluster around the most plausible answer. A panel that agrees with itself too neatly should raise suspicion.
Synthetic Focus Groups
Synthetic focus groups simulate a discussion by having multiple personas respond to a moderator and to one another. They are fast for contrasting how segments might react to the same idea.
The limit is group dynamics. In a real room, dominant voices, hesitation, and social influence shape what people say. Simulation reproduces the words of a discussion but not the pressure behind them. When group dynamics matter, run real focus group discussions.
Synthetic Usability and Concept Testing
Synthetic users can give early concept reactions, test messaging, and perform heuristic-style reviews of a design. For early filtering, a concept testing for UX approach with synthetic users can narrow a long list quickly.
Heuristic-style review is the closest fit, since it mirrors an expert applying a set of usability rules to a design.
What synthetic users cannot do is show real interaction behavior. They report opinions about a design. They do not hesitate, misclick, or abandon a task, so real unmoderated user testing is still needed to see what people do.
Benefits of Synthetic Users
Used within their limits, synthetic users offer real advantages:
Speed: Answers arrive in minutes, not weeks.
Cost: Early rounds need no incentives, recruiting, or scheduling.
Scale: One researcher can run dozens of simulated sessions.
Access: Hard-to-recruit or sensitive audiences can be explored first, with care about what the simulation can claim.
Reproducibility: The same scenario can be rerun under identical conditions, which helps when comparing study designs.
The deepest benefit is front-loading. Teams can map the problem space, sharpen hypotheses, and remove weak questions before they spend a human-study budget. The result is fewer wasted sessions and better questions.
Synthetic Users vs Real Users
The two are not interchangeable. This comparison shows where they differ:
Dimension | Synthetic users | Real users |
Data source | Model training data plus conditioning | Lived experience |
Depth | Fluent but derived | Original and unexpected |
Variability | Often low | Naturally high |
Emotional authenticity | Described | Felt and visible |
Context | Imagined | Actual environment |
Cost and speed | Low cost, minutes | Higher cost, days or weeks |
The most important row is surprise. Real users can say something the researcher never imagined. Synthetic users mostly return patterns that already exist in their training data.
What Synthetic Users Cannot Capture
The core limit is the say-do gap. Synthetic users only state. They do not behave, look, hesitate, or react. Even real participants often say one thing and do another, which our guide to the say-do gap explains, and a simulation inherits that gap in full.
Several other things stay out of reach:
Lived experience, including what it feels like to use a product while distracted or stressed
Environmental context such as noise, lighting, or a slow connection
Non-conscious emotional and attention signals
Net-new market shifts that appeared after the model's training data ended
Common Use Cases for Synthetic Users
Map use cases to the research stage. In discovery, synthetic users help build vocabulary. In study design, they test guides and screeners. In concept screening, they shortlist ideas. Between research phases, they fill small gaps while the next human wave is organized.
Risk varies by use. Rehearsing a discussion guide is low risk, because a wrong answer costs little. Using synthetic output to set a price or approve a launch is high risk, because a wrong answer costs a great deal.
UX and Product Research
UX teams use synthetic users to rehearse discussion guides, generate hypotheses, gather early prototype feedback, and learn the vocabulary of an unfamiliar domain. For a deeper look at risks and practices in this area, see our guide to synthetic users in UX research.
Teams assembling a UX research stack can compare the best user experience testing platforms to see which human methods sit alongside simulation.
Market Research and Consumer Insights
In market research, synthetic panels support concept screening, message and claims testing, survey pretesting, and audience exploration. CPG, media, and retail teams use them to shortlist ideas before human testing. Our guide to synthetic audiences in consumer research covers when this approach fits.
Teams that run these studies on a consumer research platform can pair simulated screening with real-participant studies in one workflow.
Marketing and Advertising
Marketers use synthetic customers and AI-generated personas for creative pre-screening, segment-level messaging, and early positioning checks. Treat the result as a filter. A shortlist of creative ideas still needs real creative testing with people who react to the work.
How Accurate Are Synthetic Users?
Accuracy depends on three things: the audience, the topic, and how well the persona is grounded in real data. Studies report encouraging results in narrow settings and weak results elsewhere.
In an academic experiment, researchers conditioned a language model on thousands of real survey backstories and compared its vote choices with American National Election Studies responses. The tetrachoric correlations were 0.90, 0.92, and 0.94 across three election years, with pure independents the notable exception. That first experiment cost $29 to run.
Read this carefully. The match was at population level, in US political opinion, with real survey data supplying the conditioning. It does not mean any synthetic user is accurate on any topic.
This is the difference between directional signals and decision-grade evidence. A directional signal tells you where to look. Decision-grade evidence is strong enough to stake a launch or a budget on. Synthetic output usually belongs in the first category.
How to Validate Synthetic Findings
Validation turns a plausible answer into a trusted one. The core method is parity testing: run the same questions with a human sample and measure where the results diverge.
Several habits make validation routine:
Label every synthetic output clearly, so nobody mistakes it for human data.
Recalibrate personas with fresh human data on a regular schedule.
Track where synthetic and human answers diverge, by topic and segment.
Our guide on how to validate AI moderated research findings shows the same discipline applied to interview data, and the principle carries over directly.
Limitations and Risks of Synthetic Users
The failure modes are well documented, and several are easy to miss:
Sycophancy: Models drift toward agreement. In a Stanford study of 11 language models, the models endorsed users' positions 49% more often than humans did on interpersonal dilemmas. That study did not test synthetic users, but concept reactions that are too positive deserve suspicion.
Low variance: Simulated panels cluster, so minority views disappear.
Misleading believability: Fluent, confident prose feels like evidence even when it is not.
Training-data bias: Underrepresented segments are reflected poorly, and knowledge can be out of date.
Overtrust and leading questions: Stakeholders believe the output, and a leading prompt gets the answer it asks for.
Researchers also carry their own cognitive biases into prompts, so a leading question gets the answer it asks for.
Privacy and consent matter whenever real customer data conditions a model. IBM's 2026 breach research found that roughly one in five organizations reported an attack on an AI model or application, and 92% of those lacked proper access controls.
Our team guide to data security in AI moderated research lists the controls to ask vendors about.
When to Use Synthetic Users and When to Use Real Participants
A simple principle guides the choice: synthetic generates questions, humans generate answers. Use this framework to apply it:
Factor | Synthetic users fit when | Real participants are needed when |
Decision risk | A wrong answer is cheap to catch | A wrong answer is costly |
Research stage | Discovery, design, early screening | Confirmation and final decisions |
Audience novelty | Well-documented, familiar audience | New or poorly understood audience |
Evidence needed | Opinions and language | Emotion, attention, or behavior |
Some scenarios make real participants non-negotiable:
Product launches and go or no-go decisions
Pricing
New product lines and unfamiliar categories
Sensitive audiences and topics
Best Practices and Governance for Synthetic Research
Governance keeps a useful tool from becoming a risky habit. Start with guardrails:
Define approved use cases, such as guide rehearsal and survey pretesting.
Require human follow-up for any high-risk decision.
Label synthetic outputs clearly in every report and deck.
Ground personas in real research data, and document prompts and methodology.
Assign researchers ownership of bias checks and of the policy on synthetic use.
Ownership matters most. If nobody is responsible for the policy, it will erode one convenient shortcut at a time. Our guide on how to evaluate AI research tools offers a checklist for assessing any vendor.
Vendor claims vary widely, so compare options side by side. Our roundup of the best consumer research platforms shows what market and consumer teams should expect from a research stack.
Grounding Synthetic Insights in Real Human Emotion and Attention
The strongest approach is hybrid. Use synthetic users to move fast, then use real participants to confirm. In practice:
Use synthetic users to shortlist concepts and refine hypotheses.
Validate the shortlist with real participants, using facial coding, eye tracking, and AI-moderated interviews.
Real emotion and attention data closes the say-do gap that synthetic responses leave open. A synthetic user may say it likes a concept, while real participants show confusion on their faces and gaze that skips the message.
Decode by Entropik is the human-validation layer that captures what synthetic users cannot, for UX and consumer research alike. The methods work as follows:
Facial coding reads emotional reactions as people watch or use something.
Eye tracking records where attention goes and what is ignored.
AI-moderated interviews capture spoken reasoning at scale, with follow-up questions.
The platform in numbers:
90%+ facial coding accuracy
96% eye tracking accuracy
62 facial expressions
70+ languages supported
17 patents
150+ global brands
The Future of Synthetic Users and Synthetic Research
Three shifts are likely. Synthetic users will move from generic personas toward data-grounded digital twins built from a customer's own history. Agentic research workflows will chain interviewing, analysis, and reporting. Continuous synthetic panels will stay available between human waves.
Adoption will not wait for the methods to mature. McKinsey's 2025 survey of 1,993 participants found that 88% of organizations used AI in at least one function, yet only 7% had fully scaled it. Wide use with little maturity is the setting in which weak practices spread, so standards and validation benchmarks matter.
The researcher's role shifts accordingly. Researchers become methodological guides: choosing when simulation is appropriate, checking bias, and deciding which findings need humans.
Frequently Asked Questions
1. What is a synthetic user?
A synthetic user is an AI-generated simulation of a research participant. It is typically built on a large language model and conditioned with demographic, behavioral, or first-party data, then interviewed, surveyed, or tested.
2. What is the difference between synthetic users and synthetic personas?
A synthetic user is an individual simulated participant. A synthetic persona is an archetype that represents a segment. In practice, teams often use personas to create users, and the terms overlap.
3. Are synthetic users accurate?
It depends on the audience, topic, and grounding. Results can align closely with real survey data in narrow settings, as in political opinion studies, but a match on one topic does not transfer to others. Validate before relying on them.
4. Can synthetic users replace real users in research?
No. They help you explore, rehearse, and prioritize, but they predict plausible answers rather than report lived experience. Launches, pricing, and sensitive topics need real participants.
5. What is synthetic research in market research?
It is the practice of running studies on simulated respondents instead of, or alongside, recruited people. Common uses are concept screening, survey pretesting, and message testing before human studies.
6. How are synthetic users created?
Teams define the audience, condition a language model with demographic, behavioral, or proprietary data, run the study, and analyze the outputs. Grounding in real research data makes the results more relevant and less generic.
7. What is a digital twin in consumer research?
A digital twin is a model of one specific real customer, built from that person's own data. It differs from a persona, which represents a segment, and it raises stronger privacy and consent requirements.
8. Is it ethical to use synthetic users in research?
It can be, with guardrails: label synthetic outputs, never present them as real quotes, get consent when real data is used, protect that data, and confirm high-stakes findings with real participants.
Ready to Validate Your Synthetic Shortlist With Real People?
Synthetic users speed up exploration, while decision-grade confidence comes from real human emotion, attention, and voice data. Decode helps teams validate synthetic shortlists with real participants through facial coding, eye tracking, and AI-moderated research, trusted by 150+ global brands.
