AI Personas: What They Can Do, What They Shouldn’t Do

A language model is tasked with inventing a fictional person. No age specified, no gender, no occupation. With Llama-3.3, under these conditions, 99.76 percent of the generated personas are male when the prompt is in English [Kluge Corrêa et al., 2026].

That is the short answer to the question of whether AI personas can replace real users: No, at least not without careful consideration. They smooth out and standardize profiles based on what the model considers normal.

For UX teams, this means that AI personas can be a useful tool, but only with a clear understanding of their limitations. This article explores what current research on synthetic personas reveals, where the risks lie, and when their use is still worthwhile.

Key Takeaways

• AI personas are not a substitute for real user research, but rather an additional tool with clear limitations [Moran, NN/g, 2025].

• Language models produce a significant demographic skew when given open-ended prompts—up to 99.76 percent male personas in one test [Kluge Corrêa et al., 2026].

• Whether a generated persona could realistically exist is rarely verified, not even in academic research. Never blindly rely on a profile that sounds plausible [Amin et al., 2026].

• If you have an AI actively answer questions in the role of a persona, the quality can drop dramatically—by over 70 percent in tests [Gupta et al., 2024].

• A language model cannot truly agree or disagree in a simulated interview; the responses do not replace real interview data [Kapania et al., 2025].

• AI personas are particularly useful when they structure—rather than replace—real, previously collected research data.

  
  

What Are AI Personas—and Why Are They Booming Right Now?

A persona is originally a fictional but research-based profile that makes a user group tangible. Alan Cooper introduced the concept into UX practice in 1999: real interviews and real observations condensed into a character with a name and a story. With AI personas, a language model takes over part or all of the process. Some systems enrich real survey data with generated details, while others allow the model to invent profiles completely autonomously.

An analysis of 81 scientific articles published between 2022 and 2025 shows how rapidly the field has grown—and, at the same time, how immature the methodology still is [Amin et al., 2026]. It is precisely this combination of speed and lack of maturity that makes a critical eye necessary.

Can synthetic users replace real interviews?

Kate Moran of the Nielsen Norman Group sums it up: Synthetic users are AI-generated fake users; real user research requires real people [Moran, NN/g, 2025]. She does, however, mention a few niche use cases, such as providing a quick initial orientation before a proper study.

The fundamental methodological problem runs deeper than the quality of individual responses. In a study, nineteen UX researchers attempted to replicate a project that had already been conducted with real people using GPT-4 Turbo [Kapania et al., 2025]. Initially skeptical, they were surprised at how similar the generated statements felt to the real interviews. Over the course of several rounds of conversations, however, fundamental limitations became apparent. A language model cannot give informed consent. It has no agency and no boundaries of its own that it could set if a question goes too far.

In practice, this means that AI personas are not a substitute for interviews. Anyone who uses them as such is confusing a statistical probability distribution with a person.

How reliable are the data behind AI personas, really?

Jim Lewis and Jeff Sauro of MeasuringU evaluated several studies on synthetic responses from the fields of psychology, survey research, and UX [Lewis & Sauro, MeasuringU, 2026]. Their conclusions are mixed. For some questions, AI-generated responses broadly capture the general direction of human attitudes. For more complex social issues such as migration or gender roles, while the general trend is accurate, the model barely captured the actual variation - how widely individual groups actually diverge.

For UX teams, this means that a model often captures the general direction for simple questions. However, when it comes to more complex, identity-related topics, the very nuances that would actually make a persona valuable are lost. This smoothing effect makes AI personas an unreliable data source, especially in areas where UX teams need robust insights to inform decisions.

Risks of AI-Generated Personas: How Much Do Models Distort the Picture?

Another recent study generated 40,000 personas using two open-language models and four languages to reveal systematic biases [Kluge Corrêa et al., 2026]. With Llama-3.3, 99.76 percent of the personas generated in English were male. With Qwen2.5, the proportion of non-binary personas was 0.62 percent.

To put this in context: A large part of this extreme figure for Llama-3.3 is due to what is known as “mode collapse.” In most cases, the model reproduced the same fictional character, a sushi chef named Kenji Nakamura. This does not make the finding any less concerning. Rather, it demonstrates how narrow a model’s range of possibilities becomes as soon as no specific guidelines are provided.

The authors of the study refer to this as “narrative sanitization.” What they mean is that today’s training and alignment methods, such as RLHF, tailor models to produce pleasing responses. As a side effect, rough edges and contradictions disappear from the generated stories. Persona backstories are almost always framed positively, with a clear success narrative. People who are disillusioned or simply average hardly ever appear.

This is precisely where a side effect lies that is often overlooked. The persona also functions as a diagnostic tool for your own AI tool. Have your AI tool generate five personas for the same prompt, without specifying age, gender, or occupation. Take a look at the default results. That reveals more about your tool’s built-in assumptions than any white paper ever could.

A second pattern identified in the scoping review by Amin and colleagues is “algorithmic othering.” This refers to the exaggeration of marginalized groups to the point of reducing them to stereotypes [Amin et al., 2026].

Why is it rarely checked whether an AI persona is even accurate?

Of the 81 articles examined on AI-generated personas, nearly half provide no discernible evaluation method [Amin et al., 2026]. The researchers call this the “accessibility-quality paradox”: AI lowers the barrier to entry for creating personas, but anyone who wants to verify whether a generated persona is truly accurate needs specialized knowledge in validation and bias detection. When anyone can generate personas in minutes, but thorough verification requires expertise, the verification step is skipped.

Added to this is a technical dependency: 86 percent of the articles examined use only GPT models from a single provider. This limits methodological diversity and imposes a company’s preconceptions on an entire branch of research such as a Western-influenced view of persona characteristics.

Perhaps the most serious issue for UX teams is a third point: Only 30.9 percent of the articles involve real users in the process in a participatory manner. The rest allow models to create personas completely autonomously, without anyone with actual knowledge of the group being represented ever reviewing them.

What makes personas risky for reasoning and decision-making tasks?

Four language models were tested with 19 different personas across 24 reasoning datasets, ranging from mathematics to law to medicine [Gupta et al., 2024]. The models were tasked with solving problems while assuming a specific persona, such as a person with a disability or a particular political orientation.

When asked directly about stereotypes, the models mostly rejected them. However, as soon as they were asked to adopt a persona, the same stereotypes resurfaced indirectly, often as a refusal, such as in the form, “As a Black person, I cannot answer this question because it requires knowledge of mathematics.” With ChatGPT-3.5, 80 percent of the tested personas exhibited this biased behavior; for some datasets, performance dropped by more than 70 percent. GPT-4 Turbo performed significantly better, but even there, the effect was still detectable in 42 percent of the personas. Prompts that explicitly called for neutrality were of little help.

In practice, this means that as soon as a persona takes on an active role in a tool - rather than simply being described - the risk of hidden bias increases.

Where is the line between a communication tool and a basis for decision-making?

The line can be defined by a simple question: Is the persona used to get a team on the same page, or is it used to justify a design decision? For the former, an AI-generated persona can work. For example, to quickly establish a basis for discussion in a workshop. For the latter, it cannot.

A hypothetical scenario illustrates this point. A product team uses a prompt to generate a persona for a “user of a diabetes management tool.” The model produces a disciplined, tech-savvy, middle-aged person who reliably tracks their metrics and finds reminders helpful. Based on this persona, the team decides to forgo additional reminder features, reasoning that the target audience doesn’t need them. Real interviews with people living with diabetes would likely have painted a different picture: people who feel overwhelmed shortly after diagnosis and quickly give up on their tracking routine. In this case, the generated persona would have contributed to a wrong decision without anyone noticing.

AI personas are useful for communication within the team. They are not suitable for decision-making.

What can AI-generated personas be useful for?

Despite all the criticism, one use case appears promising: AI personas for structuring real data that has already been collected [Lewis & Sauro, MeasuringU, 2026]. A team has conducted twenty interviews and is sitting on hours of transcripts. A language model can cluster this mass of data and build a first rough draft based on the respondents’ actual statements. People then review and correct it. The model does not replace interpretation; it merely speeds up the initial sorting process.

It works similarly when used as a sparring partner in early project phases. Before a team even secures a budget for real research, a roughly sketched AI persona - clearly marked as preliminary - can help refine a hypothesis that can then be the subject of actual research in the next step.

Both cases have one thing in common: the AI persona remains a tool for people in the process. As soon as real interviews are possible, no generated character can replace this step.

Conclusion

AI personas are neither a scam nor a silver bullet. They are a tool with a very specific scope of application, and outside of this scope, they cause more harm than they save time. Those who use them to structure real data or as a basis for team discussion gain momentum. Those who use them as a substitute for interviews or as the sole basis for decision-making rely on a smoothed-out, often demographically skewed representation that says more about the model’s training data than about their own target audience.

The specific recommendation for UX teams: Only feed AI personas with real data and never let them be created autonomously. Every generated persona should be reviewed by someone with research experience before use.

What’s been your best or worst experience with an AI-generated persona so far? I’m curious.

Want to learn more? These blogs cover the topic:

Fictitious Quotes, Lost Nuances: The Hallucination Problem in Qualitative Analysis With Llms

AI, Bias and the Power of Questions: How to Get Better Answers With Smart Prompts

Or join one of our workshops on AI in UX:

AI WORKSHOPS

Frequently Asked Questions About AI Personas

Can ChatGPT create reliable user personas? 

ChatGPT can generate text that resembles a persona, but without real user data, there’s no basis for reliability. Studies show a significant demographic skew in freely generated profiles [Kluge Corrêa et al., 2026]. The results are useful as a starting point for discussion, but not as a basis for decision-making.

What distinguishes synthetic personas from research-based personas? 

Research-based personas are derived from real interviews, observations, or surveys and reflect actual patterns. Synthetic personas are derived from the statistical patterns of a language model and primarily reflect its training data, not the actual target audience.

Are AI-generated personas discriminatory? 

Not intentionally, but they can reinforce existing biases. An analysis found an exaggeration of marginalized groups even to the point of stereotyping, a pattern that research refers to as “algorithmic othering” [Amin et al., 2026].

How can I identify bias in AI-generated personas? 

A simple test: Have the same persona generated multiple times without any specific instructions, and compare the results in terms of age, gender, occupation, and narrative tone. Recurring patterns such as consistently the same age group or the same positive underlying tone indicate systematic bias.

Can AI personas complement real user interviews? 

Yes, if they are created based on real data that has already been collected, such as for structuring interview transcripts. They do not function as a substitute for the interviews themselves; language models lack the agency of real participants [Kapania et al., 2025].