AI for UX · Step 4 · 5 min read
Synthetic users and digital twins
What AI-generated respondents are, what the research says about them (including a 2026 study with 604 people), where they help, where they mislead, and how to talk about them with stakeholders who want to skip real users.
"Ask the AI what users think" is now a product you can buy. Tools offer synthetic users, AI personas and "digital twins" that answer questions instantly, with no recruiting. They can be useful. But a growing body of research shows the same pattern: they sound right, agree with the headline, and lose the variety that research exists to find. NN/g's verdict is short: "Research with synthetic users is still not user research."
What you'll be able to do after reading:
- Explain what synthetic users and digital twins are, and how they differ
- Summarise what studies found when synthetic answers were compared with real people
- Use them for the few jobs they do well, and refuse them for the jobs they do badly
- Answer a stakeholder who says "why recruit, if the AI can tell us?"
1. What they are
Synthetic users are AI-generated respondents: a language model asked to answer as if it were a certain kind of person ("a 45-year-old nurse who shops online"). Digital twins are a variant built from your own data: Dovetail, for example, builds them only from a workspace's calls, tickets, surveys and research, and its answers cite the source conversation (Dovetail).
Both are still a model predicting likely text (see What AI is (and isn't)). A digital twin is grounded in real data, which helps; it is still not a person answering.
2. What the research found
- People: kinds of answer50+
- Synthetic: kinds of answer≈ 12
- Real product managers using UX terms44%
- Synthetic product managers using UX terms93%
| Study | What it found |
|---|---|
| NN/g, Oct 2026: 604 professionals vs 120 synthetic respondents, one open question | Same headline conclusion, much less variation: people used more than 50 different labels, synthetic respondents about a dozen. Synthetic product managers used UX vocabulary 93% of the time; real ones 44% (NN/g) |
| NN/g, 2024: three earlier studies re-run with synthetic users | Sycophantic about concepts (a drone-delivery idea was a "game-changer"); overstated past behaviour (a real learner finished 3 of 7 courses, the synthetic one "all 7"); "cared about everything" when asked to prioritise (NN/g) |
| Argyle et al., 2023, Political Analysis | Conditioned on real survey backstories, GPT-3 reproduced US voting patterns closely at group level ("silicon samples"), but only for distributions, not individuals (paper) |
| Bisbee et al., 2024, Political Analysis | Averages matched, but answers had less variation, relationships between variables often differed, and results changed with small wording changes and over time (paper) |
| Wang, Morgenstern & Dickerson, 2025, Nature Machine Intelligence | With 3,200 human participants across 16 identities: models that replace participants can misportray and flatten identity groups; fixes reduce the harm but don't remove it (paper) |
The academic debate is real: some studies are optimistic about group-level averages, others show what's lost. For design work, where the edge case and the surprising reason matter most, the losses are exactly the problem.
3. Where they help, where they mislead
Pilot a discussion guide, brainstorm hypotheses, draft proto-personas, scan desk research, pick topics.
Before you talk to peopleValidate concepts, report past behaviour, prioritise needs, speak for small or under-represented groups.
Instead of talking to people4. Talking to stakeholders
When someone asks "why recruit if the AI can tell us?", three short answers:
- "It tells us what's likely, not what's true." It can't report what our users actually did last week.
- "It flattens the interesting part." The 2026 study found a dozen kinds of answer where people gave fifty. The surprise we need is in the fifty.
- "Let's use it to prepare, then talk to people." Pilot the guide with a synthetic user, then run five real interviews.
5. Vocabulary
| Term | Meaning |
|---|---|
| Synthetic users | AI-generated respondents that imitate a kind of person |
| Digital twin | A synthetic respondent built from your own customer or research data |
| Silicon sample | A set of AI respondents used in place of a survey sample |
| Mode collapse | When a model's answers bunch around a few typical options |
| Sycophancy | The tendency of a model to agree with and flatter the person asking |
6. Try it (no code needed)
- Write one open question you'd ask real users ("What's the most annoying part of booking a court?").
- Ask an approved assistant to answer it as 10 different users, one line each.
- Count how many truly different answers you got. Note how many sound like a product review.
- Ask three real people the same question. Compare the variety and the surprises.
Reflection questions: Where in your process could a synthetic user save time without replacing a person? What would you lose if your team stopped talking to users?
7. If you lead a team
- Write down what synthetic users may and may not be used for in your research guidelines.
- Label any synthetic output clearly as "not user research" in documents and decks.
- Ask vendors offering digital twins what data they're built from and how they show sources.
Checklist
Learn more
- NN/g: Synthetic Users: If, When, and How to Use AI-Generated "Research" · Synthetic Users vs. Human Respondents: Same Conclusion, Less Variation (Oct 2026)
- Research: Argyle et al., Out of One, Many (2023) · Bisbee et al., Synthetic Replacements for Human Survey Data? (2024) · Wang et al., LLMs that replace human participants can harmfully misportray and flatten identity groups (2025)
Books, courses and videos to go deeper
Related
AI in UX research · AI-assisted synthesis, step by step · Synthetic users · Ethics, bias and inclusion in AI design · Algorithmic bias