< 04 > ARTICLES / RESEARCH, UX AND AI

How to Use Synthetic Personas in UX Research Without Losing Sight of Real People

Identical synthetic busts face a faceted glass figure and a lime beam, a conceptual composition about diversity.
AI-generated image · editorial direction: Vini Asevedo + ChatGPT (2026)

How to use synthetic personas in UX research without leaving human diversity behind? AI can arrive at a convincing conclusion and still miss the very people we most needed to hear from.

Generative models now make it possible to create fictional respondents, explore scenarios and discuss product decisions in minutes. But a simulated statement is not an interview; reproducing a common answer does not, by itself, reveal who has been left out.

< 01 >

Getting the Answer Right Does Not Mean Representing People

In October 2026, Nielsen Norman Group published a particularly revealing experiment. The team compared answers from 604 real technology and design professionals with those of 120 synthetic respondents, asking which label best described the work we now associate with UX.

The broad conclusion was similar: UX remained the dominant term. But that similarity faded when the researchers examined the distribution of responses. Around 84% of synthetic respondents used UX-related terms, compared with roughly 70% of humans. The synthetic group produced only about a dozen distinct labels; real people used more than 50.

The study has important limitations: the groups differed in professional composition, and the AI platform sometimes reworded the question. It is therefore not a universal estimate of accuracy. The more interesting finding is that a seemingly correct aggregate result can conceal an enormous loss of variation.

Visual comparison of about 12 distinct terms from 120 synthetic respondents and more than 50 terms from 604 people, according to NN/g (2026).
Data source: Nielsen Norman Group (2026).
< 02 >

The Problem Is Not Just Bias. It Is the Missing Exceptions.

In statistics, the mode is the most frequent value. When a model repeatedly produces the most common answers rather than reflecting the full distribution of possibilities, the result can resemble mode collapse. The consensus looks strong, but the long tail of responses disappears.

That matters in product design. Someone who abandons registration because a term is unclear, a salesperson who depends on an uncommon workflow, or someone who has lost trust in the product rarely fits a description of the average user. These experiences may be exactly what exposes the most important UX failures.

Asking AI to be more creative does not, by itself, solve the problem. Generating twelve different opinions does not prove that those opinions exist in the real population, or that they occur at the right frequency. Simulated diversity is a way to explore possibilities; representativeness requires evidence.

The Gasometer steel structure in Berlin, with crossing beams in repetitive patterns under a blue sky. Photo: Pascal Bullan / Unsplash.
Photo: Pascal Bullan · Unsplash · Unsplash License.
< 03 >

How I Used 12 Synthetic Personas at Turn2C

At Turn2C, a Brazilian platform for consórcio products (a group-based purchasing model), I explored how to use synthetic personas in UX research as a support tool for product analysis. I structured a panel of twelve perspectives around commercial maturity, product familiarity, expectations and pain points, without treating simulated responses as evidence from real users.

I was not trying to create twelve substitutes for interviews. The aim was to avoid accepting an idea simply because it worked for the user profile our team knew best. Simulations became a discussion tool, helping us anticipate questions and formulate hypotheses to investigate with real people and behavioural data.

I organised the panel into three maturity groups. Their names describe useful perspectives—not statistically validated segments or a sample of the user base.

For the operational context behind those hypotheses, see my Turn2C onboarding and activation case study.

The 12 synthetic Turn2C personas, grouped into beginners, hybrids and advanced profiles, with needs and five uses for product analysis. A transcript follows the caption.
Turn2C hypothesis panel · Vini Asevedo (2026). Icons: Lucide Icons (ISC).

Figure 03 transcript: personas and outcomes

Beginners

Uninformed
Guidance to get started
Curious
Find value quickly
Commission-focused
Understand earnings
Pragmatic
Follow clear steps

Hybrids

Integrator
Connect existing tools
Opportunistic
Act on opportunities
Solution builder
Adapt proposals
Disengaged
Rebuild trust

Advanced

Scaler
Scale with automation
Strategist
Compare scenarios
Technical
Understand rules and calculations
Frustrated
Demand reliability

What did they help with?

  1. Broaden perspectives: Look beyond familiar users
  2. Anticipate friction: Identify potential drop-offs
  3. Assess impact: Explore effects of decisions
  4. Surface conflicts: Compare competing needs
  5. Guide research: Create testable hypotheses

Each icon represents a need — not a real person.

< 04 >

The Question Itself Changed

The greatest value of this exercise came from replacing a generic question—“Is this good for the user?”—with a more demanding one: “Who benefits, who is indifferent, and for whom might this become an obstacle?”

A shorter journey could help the Pragmatic Beginner run their first simulation while removing controls that the Technical Advanced persona considers essential. Commission-focused messaging might attract a Commission-Sensitive Beginner without rebuilding trust among people who have experienced recurring problems. CRM integration might be crucial to a Hybrid Integrator but irrelevant to someone who sells only occasionally.

These are examples of tensions the panel helped us examine, not findings from tests with real participants. They helped us formulate research questions and hypotheses about activation, conversion, feature adoption, time to first simulation and support needs—without confusing a simulated reaction with an observed metric.

Railway tracks cross and diverge in multiple directions. Photo: iridial / Unsplash.
Photo: iridial · Unsplash · Unsplash License.
< 05 >

What Other Research Adds

Research from Stanford offers a more promising possibility: agents built from in-depth interviews with 1,052 real people. On a specific battery of questions, the agents reproduced responses at 85% of the consistency participants showed when answering again two weeks later. That is not 85% general accuracy: it is a relative measure tied to that research context.

Meanwhile, a 2026 study covering twelve first-click tests and 3,431 human observations found significantly different AI and human response distributions in 53% of the tasks. Another study, published in the 2026 ACM Web Conference proceedings, found that adding persona attributes to prompts did not consistently improve alignment with human survey responses and increased distortions in some subgroups.

The lesson is not that every use of AI in research is pointless. Fidelity depends on the method, the task, the evidence used to create agents and external validation. A plausible text response is not proof of behaviour.

These questions also connect with my article on design and AI trends for 2027, particularly the role of human judgement in increasingly generative systems.

Three indicators from separate studies: NN/g (604 humans and 120 synthetic respondents), Stanford (1,052 interviewees) and Kuric et al. (53% of tasks with different distributions).
Sources: NN/g (2026), Stanford HAI (2025) and Kuric et al. (2026).
< 06 >

My Criteria for Using Synthetic Personas Responsibly

When deciding how to use synthetic personas in UX research, I start with the signals already available: previous studies, journey analytics, support interactions, recurring pain points and domain knowledge. Where personal data is involved, privacy, purpose and consent must be part of the method.

Next, I state explicitly what each persona represents as a hypothesis, without inventing the proportion of the user base it supposedly represents. I use AI to prompt disagreement, explore edge cases and surface questions we have not yet asked—never to fabricate participant quotes and present them as evidence.

I then document the questions, compare the hypotheses with analytics, support data, qualitative research and usability tests, and record what was supported, challenged or remains unknown. High-impact decisions require validation with real people.

My rule is simple: simulation should improve the questions we ask and the decisions about what to investigate. When it starts closing down the discussion, it becomes a risky shortcut.

Research and documentation also need to stay close to real work. In my Smarters Design System case study, I explain how different levels of familiarity and team needs helped shape research, shared patterns and governance.

Hands sketch interface wireframes in a notebook beside a keyboard and sticky notes. Photo: Kelly Sikkema / Unsplash.
Photo: Kelly Sikkema · Unsplash · Unsplash License.
< 07 >

The Future of Research Should Not Have Fewer People

Turn2C's twelve synthetic personas expanded the range of questions I could bring into product discussions. But none of them felt frustrated by an interface, improvised a workaround during a working day or had anything meaningful at stake when a task failed.

AI can remind us to ask about the people who were left out. Whether those people exist, what they need and how they behave must still be answered by the people themselves.

Ultimately, how to use synthetic personas in UX research is a methodological choice. I do not want AI to replace the voices of research participants. I want it to help me enter real conversations with better questions.

< FIG. 07 > LISTEN TO REAL PEOPLE

Conceptual illustration titled “Simulating is not listening”: a pixelated synthetic bust and a real person listening, with speech bubbles and a Venn diagram comparing AI-generated hypotheses with human context, diversity and lived experiences.
AI-generated conceptual illustration · Vini Asevedo/ChatGPT (2026). Fictional subjects.
Conceptual references: NN/g (2026) and the Turn2C experience (2026).

Sources and reading notes

[1] Nielsen Norman Group (2026) — Synthetic Users vs. Human Respondents. https://www.nngroup.com/articles/synthetic-users-vs-human-respondents/

[2] Nielsen Norman Group (2024) — Synthetic Users: If, When, and How to Use AI-Generated Research. https://www.nngroup.com/articles/synthetic-users/

[3] Stanford HAI (2025) — Simulating Human Behavior with AI Agents. https://hai.stanford.edu/policy/simulating-human-behavior-with-ai-agents

[4] Kuric et al. (2026, preprint) — What Would GPT Click? https://arxiv.org/abs/2605.18302

[5] Morocho et al. (2026, ACM Web Conference Companion) — Assessing the Reliability of Persona-Conditioned LLMs. https://arxiv.org/abs/2602.18462

Individual image credits and sources appear with each figure. Generated illustrations are conceptual, not evidence from research participants.