The Ethics of Synthetic User Research: Why Replacing Real Human Interviews with AI Personas is a Dangerous Shortcut

The Seductive Myth of the Instant User

Imagine you are building a new software interface. You want to know how a busy, forty-something operations manager in a regional hospital will navigate your layout. Historically, this meant recruiting five to ten actual hospital workers, scheduling interviews around their chaotic shifts, compensating them for their time, and painstakingly analyzing their feedback. It took weeks, cost thousands of pounds, and required immense logistical patience.

Now, imagine you can open a chat interface on a platform like /platforms/claude, load up a system prompt representing that exact demographic, and run 5,000 simulated user tests in fifteen seconds.

This is "synthetic user research," and it is currently sweeping through product teams like wildfire. Proponents argue it is cheap, infinitely scalable, and incredibly fast. But let’s be entirely honest: replacing real human beings with simulated AI personas is a dangerous, lazy shortcut. It is ethically dubious, methodologically flawed, and almost guaranteed to result in sterile, exclusionary products.

The Echo Chamber of the "Average of Averages"

To understand why synthetic user research is so problematic, we have to look at how large language models actually work. If you consult our /glossary on neural network training, you will remember that LLMs predict the most likely next token based on a massive, historical corpus of human writing.

When you ask an LLM to "pretend to be a 45-year-old nurse who is struggling to use our app," the model does not tap into the lived experience of a nurse. Instead, it accesses the internet's collective caricature of a 45-year-old nurse. It pulls from stock advice columns, forum posts, and generic user personas that already exist online.

You are not conducting research; you are querying a high-tech mirror that reflects back the biases, stereotypes, and average opinions already baked into its training data. This creates an echo chamber. If you build your product based on what the "average" simulated user wants, you will build a product that works beautifully for absolutely nobody. Real humans are quirky, unpredictable, irrational, and wonderfully diverse. LLMs are, by design, highly predictable averages.

The Ethical Hazard of Skipping Marginalised Voices

Beyond the functional failure of synthetic users, there is a profound ethical issue at play here. When we build products, we have a responsibility to design for everyone—including disabled users, neurodivergent users, and individuals from marginalized backgrounds who are routinely ignored by mainstream tech.

These groups are already severely underrepresented in the training data of major foundational models. If your product team decides to save budget by simulating users instead of recruiting them, you are actively writing these people out of the design process.

An AI persona representing a visually impaired user will only simulate what the internet thinks visual impairment is like. It will miss the subtle, real-world physical workarounds, the screen-reader frustrations, and the specific environmental challenges that can only be identified by sitting next to a real person while they try to use your product. Relying on synthetic research is, quite simply, an act of erasure. It allows wealthy tech teams to feel like they have done their diversity due diligence without ever having to actually talk to, compensate, or listen to a marginalized human being.

Where AI Actually Belongs in the UX Workflow

This is not a call to throw the baby out with the bathwater. AI has an incredibly valuable, ethical place in user research—it just belongs in the administrative assistant’s chair, not the user’s chair.

Instead of simulating users, you should use AI to supercharge your analysis of real user interactions. For example:

• Transcription and Synthesis: Use models to rapidly transcribe hours of video interviews, pull out key emotional quotes, and help you categorize pain points.
• Theme Clustering: Feed raw, anonymized notes from human interviews into a model to identify patterns you might have missed.
• Interactive Prototyping: Use collaborative design tools like those on the /platforms/figma-weave platform to quickly generate alternative interface variations based on direct, explicit feedback from your real-world test subjects. You can see live examples of how these collaborative prototyping systems work on the official Figma Support Portal.

In these scenarios, the AI is helping you process the genuine, messy reality of human behavior. It is amplifying human voices rather than replacing them with a cheap synthetic substitute.

Keeping Your Research Honest

If you want to build products that actually make a difference in the real world, you have to do the quiet, unglamorous work of talking to real people. If you find yourself tempted to spin up a fleet of simulated user agents, ask yourself these three critical ethical questions:

1. Whose voice am I ignoring? Who is being left out of this conversation because we chose to run a simulation instead of recruiting real people?
2. Are we validating our own assumptions? Am I prompting this AI in a way that just guarantees it will tell me my design is brilliant?
3. Am I compensating my audience? Real user testing puts money back into the pockets of the communities you are trying to serve. Synthetic testing puts money into the pockets of cloud computing providers.

Technology should be used to bridge the gap between product builders and their users, not to build a permanent wall between them. Turn off the simulation, close the persona generator, and go talk to a human.