They are three names for AI stand-ins that simulate different things. A synthetic user mimics a user group. A synthetic respondent answers a survey as one member of a sample. A digital twin simulates one specific, real person, built from that person’s own data. The difference matters because the evidence behind each is different, and so is the question each one can answer.
The market uses the three words loosely, and vendors use them interchangeably. The research does not. Nielsen Norman Group’s 2025 review of the evidence separates them by what is simulated: digital twins are “generative AI models designed to simulate individual users”, synthetic users are “models that mimic broader user groups”. That one distinction, a person or a group, decides what data the system needs, what it can be tested against, and what it should never be used for.
What each one simulates
| Term | Simulates | Usually built from | What you get back | Who uses the word |
|---|---|---|---|---|
| Synthetic user | A user group | A description of the group, sometimes research material about it | Interview-style answers from a representative member | UX and product research |
| Synthetic respondent | A sample member | Demographic profiles, often calibrated on real survey data | Survey answers, aggregated into distributions | Survey platforms, agencies, marketing research |
| Digital twin | One real person | That person’s own interviews, surveys or records | Predictions of what that person would say or do | Social science, a growing set of vendors |
The table is the clean version. In practice the edges blur: a synthetic respondent can be interviewed, a synthetic user can fill in a survey, and a vendor can call either a twin. The useful test is not the label on the product page. It is the question underneath: whose data is this system built from, one person’s, a sample’s, or a paragraph about a segment?
Where the words come from
Synthetic users is the UX term. Nielsen Norman Group’s 2024 definition is still the plainest: “A synthetic user is an AI-generated profile that attempts to mimic a user group, providing artificial research findings produced without studying real users.” The unit is the group. The persona speaks for it.
Synthetic respondents is the survey term. When Qualtrics launched its synthetic offer in March 2025, it described a tool that “enables researchers to ask questions of synthetic respondents designed to represent their target customers”, sold as “AI-powered synthetic panels”. Academic marketing uses the same word: Arora, Chakraborty and Nishimura’s Journal of Marketing paper on human and AI hybrids in marketing research has language models “generate synthetic respondents that match those characteristics”, the characteristics of the sample a study needs.
The survey lineage has an older academic name. In 2022, Argyle and colleagues wrote: “We create ‘silicon samples’ by conditioning the model on thousands of socio-demographic backstories from real human participants in multiple large surveys conducted in the United States.” The unit there is the sample, and the test is whether the silicon distribution matches the human one.
Digital twins is the newest of the three in this sense. The Columbia team behind the Twin-2K-500 dataset defines the approach as “LLM-based digital twin simulation, where large language models are used to emulate individual human behavior”, and built its dataset from 2,058 people answering 500 questions across four waves. The best-known study of the kind, from Stanford, built agents of 1,052 real individuals from a two-hour interview each. The unit is the person, and the test is whether the twin answers the way that person does.
What the evidence says about each
- Digital twins. The Stanford study’s first version reported that its agents “replicate participants’ responses on the General Social Survey 85% as accurately as participants replicate their own answers two weeks later”. The revised version of June 2026 reports 83% for agents built from interviews alone, 82% from surveys alone and 86% from both, against 74% for agents given demographics only. Nielsen Norman Group’s summary: “Digital twins work fairly well to replicate both individual-level and group-level human responses.”
- Synthetic users. The same review found that their performance “tends to capture general trends in human behavior but fails at capturing the magnitude of the effects or the variability in the human data.” Direction yes, size and spread no.
- Synthetic respondents. A June 2026 study of LLM consumer panels by Tigre and Souto found that good aggregate numbers can hide “variance compression, coefficient sign-flips” and subgroup errors of 10 to 30 percentage points. The averages match. The segments you would act on may not.
None of this is settled. The ExploraTwin team, from the same Columbia group, put the caution plainly in August 2026: “current empirical evidence suggests that the approach should be tested before being deployed in any particular context.” That holds for all three words. The fuller account of what holds up is in the note on whether synthetic respondents are accurate.
One continuum, not three products
The review that separates the terms also joins them: its findings “reinforce the idea that digital twins and synthetic users exist on a continuum”. What moves a system along it is how much real data about real people it rests on. At one end, a demographic line in a prompt. At the other, a person’s own two-hour interview. The Stanford numbers show the gradient: demographics alone reach 74% of people’s own consistency, interviews 83%.
Between the ends sits the grounded persona: a segment built from real research material, interview transcripts and documents, rather than from a demographic line. It is not a person, so it cannot be tested the way a twin is. It is also not a paragraph of assumptions. The difference in what that grounding does to a reply is the subject of grounded versus prompted personas.
Which one answers which question
| Your question | Closest fit | Check before you trust it |
|---|---|---|
| How might this segment react to a concept, a message or a price? | Synthetic users, ideally grounded in research about the segment | Direction, not size: the review found effects and spread are where they miss |
| What share of a market would choose option A? | Synthetic respondents calibrated on real survey data | The subgroups: that is where the errors grow |
| What would this specific customer say? | A digital twin, built with that person’s consent and data | Consent, and whether the twin is being used outside the context it was built for |
The third row carries an obligation the other two do not. A twin models a real person, so the review asks “whether participants have given informed consent for their data to be repurposed in this way”, and concludes that “digital twins should complement, not replace, human-centered research”. For every row, the same 2024 advice applies: “Treat the data you acquire from synthetic users as hypotheses that need testing.”
Where Perplica sits
Perplica builds grounded personas: composites of a segment, built from your interview transcripts, briefs and documents, and attached to that material with a declared fluency. It does not build digital twins. Its terms forbid modelling real, identifiable individuals, and forbid presenting simulated output as responses from real participants. Every reply stores the reasoning behind it, and every finding in a report links to the turn that supports it, so a hypothesis arrives with its evidence.
What is the difference between a synthetic user and a digital twin?
What is being simulated. A digital twin simulates one specific, real person and is built from that person’s own data. A synthetic user mimics a group. Nielsen Norman Group’s 2025 review puts it in one line each: digital twins are generative AI models designed to simulate individual users, synthetic users are models that mimic broader user groups.
Are synthetic respondents the same as synthetic users?
Close cousins from different rooms. Synthetic respondent is survey and market research vocabulary: a simulated member of a sample whose answers are aggregated into distributions. Synthetic user is UX vocabulary: a simulated member of a user group you question in an interview. Qualtrics sells synthetic respondents as synthetic panels.
Are digital twins more accurate than synthetic users?
In the studies Nielsen Norman Group reviewed, yes: twins built on individual data captured population-level behaviour better than synthetic users built on generic attributes. The best-known study, from Stanford, reports agents built from two-hour interviews reaching 83% of people’s own two-week test-retest consistency in its latest version, against 74% for agents given demographics only.
Does Perplica build digital twins?
No. Perplica’s terms forbid modelling real, identifiable individuals. Its personas are composites of a segment, built from your research material: interview transcripts, briefs and documents. That places them between a demographic prompt and a twin, with the grounding the evidence says matters.