Framework · AI & design · 3 min read
Synthetic users
AI-simulated “users” or personas that answer research questions. Use them to brainstorm hypotheses and sharpen your research plan, never as a substitute for real users or as evidence.
- 1Generate
Ask synthetic users for possible problems, objections and edge cases.
"What could go wrong here?" - 2Label
Mark every output as a hypothesis, never as evidence.
"Hypothesis, not evidence: synthetic" - 3Validate
Test each hypothesis with real users, and look for what was missed.
"What did nobody predict?" - 4Report
Cite only real participants. Mention synthetic use in the method.
"All findings come from real people"
Use it when
- You are preparing a study and want to brainstorm hypotheses, objections or edge cases before you talk to real users.
- You want to test a discussion guide or survey for unclear questions before a pilot session.
- A stakeholder proposes "AI personas" or a tool's "digital twins" to skip research, and you need a clear position.
How to use it
- Decide the job before you start. Use synthetic users only for early, low-stakes work: generating hypotheses, finding gaps in a discussion guide, or listing possible objections. Write the purpose at the top of the session. Output: one sentence on what you will use the output for.
- Use approved tools and no real data. Run it in an approved tool and do not feed it real transcripts, customer data or personal data to make it "more realistic". See AI usage guidelines. Output: a session that stays within your data rules.
- Label every output as a hypothesis. Put "Hypothesis, not evidence: synthetic" on every note, slide or board item that comes from it. Output: a labelled list of hypotheses.
- Validate with real users. Turn the hypotheses into questions and tasks for real research, such as interviews, Contextual inquiry or usability tests. Look especially for what the synthetic answers did not predict. Output: a research plan that tests the hypotheses with real people.
- Report only real evidence. In findings, cite only what real participants said or did. If synthetic output shaped the plan, say so in the method section. Never quote a synthetic user as if a person said it. Output: findings traced to real sources, with an honest method note.
Worked example
A designer on an internal warehouse app prepares interviews about a new returns flow.
| Step | What happens |
|---|---|
| Synthetic session | The designer asks an approved AI tool to role-play three warehouse operators and list what could go wrong with the new flow. It suggests damaged labels, missing order numbers and slow scanners. |
| Labelled | All three go on the research board as "Hypothesis, not evidence: synthetic". |
| Real research | Five contextual inquiries on the warehouse floor test the three hypotheses. |
| What real users showed | Damaged labels were real. Slow scanners were not an issue. The biggest problem, which no synthetic answer mentioned, was that returns arrive in mixed boxes from several orders. |
| Report | Findings cite only the five sessions. The method note says synthetic users were used to draft hypotheses. |
The most important finding came from the variation that the synthetic users did not produce.
Watch out for
- Citing synthetic answers as evidence. A quote from a synthetic user in a findings deck looks like a real user said it. Never do it, even with a label.
- Using it to skip research. "We already asked the AI personas" is not research. For internal and expert users, domain depth is what AI lacks most.
- Trusting the average. Synthetic users give plausible, typical answers. The rare expert workaround that changes your design will not appear.
- Feeding it real data to make it better. Uploading transcripts or customer data to "train" a persona breaks your data rules and still does not make it a real user.
Template
Copy this onto your research board or plan when you use synthetic users.
Synthetic session Date: Tool (approved):
Purpose (hypotheses only):
Hypothesis, not evidence: synthetic
1.
2.
3.
How we will test each with real users
1. Method: Participants:
2. Method: Participants:
3. Method: Participants:
Result from real research (fill after)
1. Confirmed / Not confirmed / Changed:
2.
3.
Surprises nobody predicted:
Method note for the report:
"Synthetic users were used to draft hypotheses. All findings come from real participants."
Saved only on this device.