AI-Generated Personas: The Risks of Inventing Your Users
AI-generated personas look complete but repeat the model's stereotypes. What the studies show and how to use them only as a draft of hypotheses to test.
What is a persona and what is it for?
A persona represents a group of users, with goals, context, frustrations, and behaviors. Its job is to remind the team who it designs for and to help prioritize. A good persona is not a pretty portrait: it is a summary of real data that changes decisions.
The problem with generating one with AI is that the text looks complete. It has a name, an age, a job, a quote, and a list of frustrations. It looks like research, but it is a blend of what the model learned about similar profiles. None of it comes from your users.
What do the studies show?
Four recent papers look at the problem from different angles. None of them studies exactly the UX personas for your product, so read them as signals of what can happen, not as a measurement of your case.
A dominant group per occupation. Van der Linden and colleagues (arXiv, 2025) asked four models (GPT-4, Gemini 2.5, DeepSeek V3.1, and Mistral-medium) for more than 1.5 million profiles across 41 US occupations, with the prompt “Generate a profile for: occupation in the United States,” and compared them with US Bureau of Labor Statistics data. The models compressed each occupation toward a dominant profile. Pooled, White workers appeared 31 percentage points below reality and Black workers 9 below, while Hispanic workers appeared 17 points above and Asian workers 12. Gemini generated housekeepers as Hispanic 99.9% of the time. The model matters: the underrepresentation of White workers ranges from about 18 points (DeepSeek) to about 44 (Gemini). The authors note that they tested a single prompt and that some model versions have since been retired.
Stereotypes versus real people. A Penn State study presented at AAAI 2026 (“A Tale of Two Identities”) compared more than 1,500 personas generated by GPT-4o, Gemini 1.5, and DeepSeek v2.5 with answers from 141 real people with similar characteristics. When the chatbot represented a 50-year-old African American woman, it talked about gospel music, tough love, social justice, and natural hair care. The real people talked about work, parenting, volunteering, and health. The researchers describe four representational harms: stereotyping, exoticism, erasure, and benevolent bias.
Stereotypes when you condition. Bombieri and Rospocher (2025) simulated 200 US residents answering sociodemographic questions. They conclude that models often produce homogenized profiles that underrepresent demographic diversity, and that conditioning on gender, ethnicity, or disability can trigger stereotyped shifts.
A counterpoint. Sethi and colleagues (CHI EA 2025) analyzed 600 persona descriptions from five models with eleven lexical diversity metrics. They found no significant demographic bias in that diversity (age, gender, and country), but significant differences between models. That metric measures the variety of vocabulary, which is not the same as measuring whether the content is stereotyped. Still, the result is a reminder that “bias” depends on what you measure and which model you use.
Nielsen Norman Group adds a point about use. Its tests with synthetic users (reviewed in synthetic users) indicate that synthetic data can help develop proto personas, which you must revisit after researching with real users, but that the flat, overly broad answers of synthetic users are not useful for prioritizing features or creating personas.
How do you use them as drafts?
- Use them as proto personas: a starting point to list hypotheses about the group.
- Label each attribute with its source: does it come from your research or from the model?
- Discard anything without a source, especially verbatim quotes, precise ages, and demographic details.
- Check against interviews: each motivation should appear in some real conversations. If you need help organizing what you heard, see AI interview synthesis.
- Review representation: ask which groups are missing and which stereotypes show up. Model bias explains why it happens.
- Version the persona with a date and its sources, so you know what changed when new data arrived.
When should you NOT use them?
- When the persona will circulate as evidence to stakeholders without stating its origin.
- When it is used for populations that are underrepresented in the model’s data, such as people with disabilities or specific communities.
- When it replaces the interviews the team did not have time to run.
- When the result will drive accessibility, health, or financial decisions.
How do you document a persona with sources?
A useful persona forces you to declare the origin of each data point. This is a minimal template:
| Field | Content | Source |
|---|---|---|
| Context | Where and when they use the product | Interviews P01 to P06 |
| Main goal | What they want to achieve, in their words | Quote from P03 |
| Most frequent frustration | What stops them, and how often | Four of six interviews |
| Decision criterion | What makes them choose one option | Session observation |
| What we do not know | Open questions about this profile | List for the next round |
If a field has no source, leave it empty or mark it as a hypothesis. A persona with honest gaps is more useful than a complete, invented one. Review it from time to time: if the product has changed its audience, if the team no longer consults it when deciding, or if its goals describe someone who does not appear in the new data, replace it rather than patch it.
Example
Illustrative example: a team wants a persona for a savings app for families. A colleague generates “Laura, 42, mother of two, works as an administrative assistant.” Before using it, the team reviews: the age and job come from the prompt, the line “I’m afraid of failing” does not come from any interview, and the goals resemble those of any finance app. They replace the persona with one built from six real interviews and keep the generated one as a list of hypotheses to test.
What the team learned: the AI profile helped start the conversation, but real data changed the main goals.
Common mistakes
- Presenting a generated persona as if it came from research.
- Using verbatim quotes that the model invented.
- Not checking which groups are missing from the representation.
- Keeping the first version without testing it.
- Forgetting the date and source of each attribute.
Sources
- Generating the Modal Worker: A Cross-Model Audit of Race and Gender in LLM-Generated Personas Across 41 Occupations (arXiv 2510.21011)
- Penn State: chatbots overemphasize sociodemographic stereotypes
- Mining Impersonification Bias in LLMs via Survey Filling (Bombieri and Rospocher, 2025)
- “When AI Writes Personas”: Analyzing Lexical Diversity in LLM-Generated Persona Descriptions (CHI EA 2025)
- Synthetic Users: If, When, and How to Use AI-Generated “Research” (Nielsen Norman Group)