Why You Shouldn't Use AI to Generate User Personas (And the Right Way to Leverage LLMs in UX Research)

The Allure of the Instant User

We have all been there. You are starting a new product cycle, the budget for user research is practically non-existent, and you need to align the team on who, exactly, you are building for. The temptation is staring you right in the face. You open Claude or GPT-4o, type a quick prompt, and within three seconds, you are presented with "Dave, 34, DevOps Engineer from Bristol." Dave has a pet labrador, struggles with context-switching, and apparently spends his weekends optimizing home automation setups.

It feels like magic. You copy-paste Dave into your design tool of choice—perhaps using Figma Weave to quickly wireframe a dashboard tailored to his exact pain points.

But there is a massive, ethical elephant in the room: Dave does not exist. More importantly, Dave is not even a realistic representation of someone who does. He is a statistical average of the internet's collective assumptions about DevOps engineers. When we use LLMs to hallucinate user personas, we are not conducting research; we are designing for an echo chamber of our own biases.

The Danger of Synthetic Empathy

At the heart of good user experience design is empathy. Empathy requires us to step outside our own worldview and understand the genuine struggles, workarounds, and frustrations of real people. LLMs, by their very nature, cannot empathise. They predict the next most likely token based on their training data.

When you ask an LLM to generate a user persona, it does not scour the earth for authentic human struggles. It looks at its internal glossary of associations and spits out a highly polished caricature. This presents three major ethical and practical risks for product builders:

1. The 'Average of Averages' Problem: LLMs smooth out the messy, beautiful edges of human behaviour. Real users are irrational. They use tools in ways the creators never intended. They have bizarre browser extensions that break your Javascript. They use password managers in weird ways. An LLM-generated persona will always be too neat, too logical, and utterly devoid of the friction that defines real life.
2. Erasure of Underrepresented Groups: If your LLM's training data contains implicit biases about who works in tech, who manages finance, or who uses accessibility tools, those biases will be baked directly into your generated personas. You risk building products that exclude real people because your AI-generated "ideal user" is a reflection of historical systemic bias.
3. The Feedback Loop of Doom: If you generate synthetic personas with an LLM, use those personas to write user stories, and then—heaven forbid—use another LLM agent to "test" the prototype, you have completely removed human beings from the loop. You are essentially paying cloud providers to let two instances of matrix multiplication talk to each other while your actual product ticks closer to failure.

To build tools that actually work, we have to understand what makes our real users tick—and that requires genuine human data.

How to Ethically Use LLMs in UX Research

Does this mean LLMs have no place in user research? Absolutely not. But we must shift our perspective. AI should never be the source of user insights; it should be the processor of them.

If you want to use LLMs ethically and effectively in your design pipeline, here is the rule: Garbage in, garbage out. Real data in, structured insights out.

1. Synthesise, Don't Synthesise Instead of asking an LLM to invent a user, do the hard work of interviewing five to ten real people. Record the calls, get clean transcripts, and feed those *real transcripts* into your LLM.

Now, you can ask the model to do what it does best: pull out key themes, categorise recurring pain points, and highlight direct quotes. The resulting persona will be grounded in real-world friction, not synthetic fantasy.

2. Identify the 'Workarounds' Real users rarely follow the golden path. They copy data into Excel, they keep physical sticky notes on their monitors, and they share passwords over Slack. When analysing your research transcripts, ask the LLM specifically to look for "user workarounds." This is where the real product opportunities lie, and it is something an LLM would never invent on its own.

3. Use AI to Challenge Your Biases, Not Confirm Them Once you have drafted a persona based on real human interviews, use the LLM as a devil's advocate. Prompt it with: *"Here is a user persona based on real interviews. Read through this profile and identify any logical leaps, ungrounded assumptions, or potential biases I might have introduced during my synthesis."*

This keeps the designer in control while using the model's analytical capabilities to keep your own human blind spots in check.

The Golden Rule of Product Design

We must resist the urge to automate the parts of our jobs that require genuine human connection. Empathy cannot be outsourced to a GPU cluster.

If you are looking for live examples of how product teams are navigating these design boundaries and building human-centric interfaces, check out the community-submitted workflows on the Figma Weave showcase. Build with real data, design for real friction, and leave the fictional DevOps engineers in the training weights where they belong.