Why You Shouldn't Use LLMs to Simulate User Personas (And the Ethical Way to Do User Research)

The concept of "synthetic users" is currently the darling of the product management world. On the surface, the pitch is incredibly seductive: Why spend weeks recruiting participants, coordinating schedules, offering Amazon gift cards, and sitting through awkward Zoom interviews when you can just simulate your entire user base?

With platforms like /platforms/gemini boasting massive context windows, it is shockingly easy to upload your product specs, define ten highly detailed personas, and ask the model to "roleplay as a 54-year-old accountant with mild tech anxiety using our new expense tracking app."

Instantly, you get pages of feedback. The simulated accountant tells you that the buttons are clear, the flow makes sense, and they would gladly pay $12 a month for this service.

It is clean, it is painless, and it is almost entirely useless.

Using LLMs to simulate user personas is not user research; it is a sophisticated confirmation bias machine. Here is why synthetic users are an ethical and operational trap, and how you can use AI in your research workflow without losing touch with actual human reality.

The Echo Chamber of the Average Token

To understand why synthetic users fail, we have to look at how these models work. LLMs are trained to predict the most likely next word based on a massive corpus of existing human text. They are, by definition, engines of the average.

When you ask an LLM to play a persona, it does not draw on lived experience. It draws on cultural tropes, stereotypes, and online discussions about that demographic.

Your "54-year-old accountant with tech anxiety" will act exactly how a tech-savvy software developer in San Francisco (who wrote the training texts or designed the prompt) imagines an anxious accountant would act. It will be a caricature.

Real human users are beautifully, maddeningly unpredictable. They do not read your onboarding tooltips. They have ten other tabs open, a screaming toddler in the background, a flaky Wi-Fi connection, and an outdated version of Safari. They click the wrong button five times, get annoyed, and close the tab.

An LLM cannot simulate the physical, emotional, and environmental friction of real life. It will never tell you that your interface is unusable when viewed in direct sunlight on a cracked screen, because the LLM exists in a pristine, text-based vacuum.

The Erasure of Accessibility and Marginalised Voices

From an ethical standpoint, relying on synthetic personas is a massive step backward for inclusive design.

If you want to know how a visually impaired user navigates your interface with a screen reader, you cannot simulate that with an LLM. If you want to understand how your product is received by a community with distinct cultural nuances, asking a model that has been heavily aligned to Western, corporate norms will give you a thoroughly sanitised, homogenized response.

When we replace real research with synthetic research, we systematically erase the voices of those who are already underrepresented in tech. We build products optimized for an imaginary, default user, while completely missing the accessibility hurdles and edge cases that define great product design. Check out our /glossary to dive deeper into how algorithmic bias propagates in large language models.

The Ethical Way to Supercharge Your User Research with AI

This is not a call to throw the baby out with the bathwater. AI can be an incredible force multiplier for user researchers, product managers, and designers—as long as its role is restricted to processing, structuring, and preparing data, rather than inventing it.

Here is how to design an ethical, AI-assisted research workflow:

1. Use LLMs to Stress-Test Your Interview Guides Before you get on a call with a real human, use an LLM to review your questions. You can feed your interview script to a model and ask: * "Identify any leading questions in this script that might bias the user's response." * "Are there terms in this guide that contain internal company jargon that a normal customer won't understand?" * "Suggest three open-ended follow-up questions to help me dig deeper into their pain points."

This uses the model's analytical strength to make your human interactions significantly more effective.

2. Synthesise, Don't Fabricate Never use AI to generate raw data. Instead, use it to help you synthesize the messy, chaotic qualitative data you have gathered from *real* users.

After conducting ten user interviews, upload the clean transcripts. Ask the model to cluster the pain points, identify recurring feature requests, or extract direct quotes related to pricing. The data remains authentic and human; the AI is simply acting as a high-speed highlighter. (Note: Always ensure you have user consent and have scrubbed any personally identifiable information before uploading transcripts to external APIs).

3. Roleplay to Build Empathy, Not to Make Decisions There is one valid way to use a synthetic persona: as a training tool for your own team.

If you want to prep your junior product managers or customer success reps, having them roleplay an interview with an LLM acting as a difficult customer can be a fantastic, low-stakes training exercise. It helps build empathy and teaches them how to handle objections. But the moment you start using those conversations to prioritize your product roadmap, you have crossed the line from empathy-building into delusion.

Nothing replaces the uncomfortable, enlightening, and occasionally frustrating experience of watching a real human use what you built. If you want to design interfaces that actually work, close the chatbot and go talk to a human.