Why You Shouldn't Use LLMs to Simulate User Testing (And the Ethical Way to Gather UX Insights)

The Temptation of the "Synthetic User"

Building digital products is exhausting, expensive, and slow. But of all the stages in the product lifecycle, user testing is notoriously the hardest to scale. Finding real humans, scheduling sessions, compensating them for their time, and listening to them struggle with your navigation menu takes days, if not weeks.

So when large language models (LLMs) arrived with the promise of "synthetic personas," product managers and UX designers everywhere breathed a collective sigh of relief. Why pay fifteen participants to test your new checkout flow when you can prompt an LLM to pretend to be a "busy 45-year-old working mother of two with mild tech anxiety" and get instant feedback for a fraction of a penny?

It sounds brilliant. It is also incredibly lazy, ethically compromised, and technically flawed.

We need to stop using LLMs to simulate human users. Here is why synthetic user testing is a trap, and how you can use AI to make your real-world research workflows better without fabricating humanity.

The Fallacy of the "Average" Predictive Persona

At their core, LLMs are statistical prediction engines. They are trained on vast corpora of existing text to predict the most likely next word. When you ask a model to pretend to be a specific user persona, it does not magically become a human with physical limitations, emotional stressors, or erratic real-world environments. It simply generates the most statistically average representation of how that persona is written about on the internet.

This leads to three massive technical flaws in your feedback loop:

1. The Politeness Bias: LLMs are trained to be helpful assistants. Even when prompted to be critical, they struggle to mimic the raw, frustrated, irrational behaviour of a real human who cannot find the "cancel subscription" button. Synthetic users will almost always understand your UI because the model itself is incredibly smart, even if the persona you asked for is supposed to be tech-illiterate.
2. The Absence of Physical Constraints: A real user testing your app might be holding a crying toddler, standing on a bumpy train, dealing with a glare on their phone screen, or suffering from mild cognitive fatigue at the end of a ten-hour shift. An LLM experiences none of this. It operates in a frictionless vacuum of pure text.
3. Hallucinated Desires: If you ask an LLM if it would buy your new SaaS tool, it will often say yes because that is a pleasant, logical conclusion to the simulated dialogue. It has no actual credit card, no real budget constraints, and no genuine pain point.

Understanding what makes your real users tick is the entire point of UX design. If you replace them with machines, you are essentially building a product designed by an AI, for an AI.

The Ethical Hazard: Erasing Marginalised Voices

Beyond the raw functional failures, there is a deep ethical issue with synthetic user testing.

Real user research is the only phase of product development where the marginalized, the non-technical, and the overlooked get a direct say in how technology is built. When we replace this process with LLM simulations, we rely on training data that is historically biased toward Western, affluent, able-bodied demographics.

If you use an LLM to simulate a user with a visual impairment or someone from a low-income background, you are not getting their lived experience. You are getting a caricature of their experience generated by a model. Relying on these caricatures to make product decisions is not just poor design; it is an active erasure of real human diversity. It allows teams to tick the "user testing" box without ever having to look a real, struggling human in the eye.

How to Ethically Use AI in UX Research

This does not mean you should banish AI from your design stack entirely. LLMs are magnificent productivity multipliers when they are used to process actual human feedback rather than inventing it.

Instead of simulating users, focus on using AI to make your real human research go faster. Here are three ethical ways to deploy LLMs in your UX workflow:

1. Synthesising Real Qualitative Data Real user testing generates hours of transcripts, video recordings, and messy sticky notes. This is where tools like [Claude](/platforms/claude) shine. You can feed transcripts of real user interviews into the model and ask it to cluster themes, extract pain points, and categorise feature requests.

By feeding the model real human words, you ensure the output is grounded in truth, not statistical fantasy. If you run into issues with the model misinterpreting user sentiments, check out our guide on troubleshooting qualitative extraction prompts over at the OpenAI articles hub.

2. Drafting Research Scripts and Scenarios Don't use AI to answer your questions—use it to help you ask better ones. You can use LLMs to draft neutral, non-leading user testing scripts. Prompt the model to identify bias in your own questions, helping you refine your approach before you sit down with real participants. You can find excellent structured templates for this in our [prompt library](/prompts).

3. Finding Patterns in Customer Support Tickets Your customer support backlog is a goldmine of genuine, unprompted user testing. Use an LLM to analyze thousands of real support tickets to identify where users are getting stuck in your existing onboarding flows. This is real-world data, born of real human frustration, surfaced at scale.

Keep Design Human

Designing products is fundamentally an act of empathy. When you swap out real humans for synthetic personas, you trade empathy for convenience. Your product might ship faster, but it will slide into mediocrity just as quickly. Keep your research grounded in reality, talk to real people, and let the AI do the heavy lifting of sorting the notes afterward.