Synonyms: AI-generated users, synthetic participants, AI personas, simulated users, AI research participants
Do not index
Definition
Synthetic users are AI-generated profiles that mimic a target user group, built so you can "interview" them instead of recruiting real people. You define the audience and the goal, and a model spins up fictional participants plus interview transcripts in seconds. They look like a persona you can chat with, but the answers come from training data, not lived experience.
Use cases
Lean on them as real research and you'll ship decisions built on plausible fiction. NN/g tested synthetic users against three real studies and found the responses "too shallow to be useful."
The prioritization trap. Ask a synthetic user what matters and it lists seven things, all equally important. Real people prioritize; synthetic ones care about everything, so you can't tell what to build first.
The sycophancy problem. In one NN/g study, synthetic users reported finishing every online course. Real participants reported dropouts, no time, and lost motivation. The AI tells you what sounds good.
The recruiting-is-hard shortcut. You can't reach a niche segment fast, so you simulate them for a first pass, then validate the questions and hypotheses with real users before deciding anything.
How it's used in practice
Use them before real research: sharpen interview questions, learn the domain vocabulary, pressure-test a discussion guide.
Keep them to low-risk calls: lean on synthetic input when speed matters and the cost of being wrong is low, like early hypothesis generation.
Draft, then verify: generate a first cut of personas or journey maps, then correct them against real data and real interviews.
Label the source honestly: never let "synthetic" findings get reported as "validated," and flag every insight's origin so stakeholders don't over-trust it.
🪄
Pro-tip: The dangerous part is that synthetic users never say "I don't know" or "I've never done that."
They answer everything, fluently, with zero lived experience behind it. If a finding feels suspiciously tidy and agreeable, that's the tell. Use synthetic users to find better questions, then spend your real participant budget where nuance actually decides the design.
Challenges & limitations
No unknown unknowns. Synthetic users only reflect what's in their training data, so they can't surprise you, and the biggest product opportunities usually come from surprises.
They erode empathy. Real interviews build a vivid sense of the user in everyone's head. A transcript from a robot doesn't, even when the words look similar.
