Synthetic users: useful shortcut or dangerous shortcut?

Ask a synthetic user a question and it will always answer. That is precisely what worries me about them, and also where they earn their keep.

Synthetic users are AI personas built to stand in for your customers, and the idea is having a moment. The pitch is seductive: skip the recruiting, skip the waiting, skip the cost, just ask the model what your users would think. As someone building in this space, I get the appeal and I feel the pull. I also think the honest answer to "useful shortcut or dangerous shortcut" is yes, both, and which one you get depends almost entirely on how you use them and what you build them from. So let me try to be straight about both sides, including about the version we are building ourselves.

Where they genuinely help

There is real value here, and I do not want the caution that follows to drown it out.

The clearest win is bridging the gap between studies. Real research has rhythm: you run a study, you learn, and then there is a stretch before the next one where the questions keep coming but the budget and the calendar do not refresh. A synthetic user grounded in the research you already did can carry you across that gap. It cannot tell you something genuinely new, but it can remind you, quickly, what you already learned, in a form you can question. That is a legitimately useful thing to have on a Tuesday when the next study is three weeks out.

The second win is speed at the front of an idea. Early hypothesis testing is mostly about killing bad directions cheaply before you spend real money on them. A few useful patterns:

• Pressure-testing a concept before you recruit, so the obviously flawed version dies on your own desk instead of in front of fifteen paid participants.
• Sharpening your questions, because rehearsing a discussion guide against a persona quickly exposes the vague or leading ones.
• Generating hypotheses to go verify, treating what the persona says not as an answer but as a list of things worth asking real people about.

Used this way, a synthetic user is a thinking aid. It is a faster, more structured version of the conversation a good team already has in their heads when they argue about what users want. That is worth something. It is just not worth what its loudest advocates claim.

Where they get dangerous

Here is the part I will not soften. A synthetic user cannot replace a real person, and the danger starts the moment you forget that.

The core problem is that a synthetic user has no inner life and nothing actually at stake. It has never paid your invoice, missed a deadline because your product broke, or felt the small daily friction that makes someone quietly churn. It produces plausible language about an experience it never had. Plausible is exactly the trap, because plausible is what we are wired to believe, and a confident paragraph that sounds like a customer can carry more weight in a meeting than it has any right to.

A synthetic user is only ever as good as the real conversations it was built from, and it can never be better than them.

Then there is the manufactured-consensus problem, and it is worse than it first looks. Ask a panel of synthetic users a question and they will tend to converge, smoothly, on a reasonable-sounding answer. No awkward outlier. No person who stubbornly wants the opposite thing for reasons you did not anticipate. That tidy agreement feels like signal and is often just the model's averaging. Real people disagree, contradict themselves, and surprise you, and that texture is not noise to be cleaned away. It is the most valuable thing in qualitative research. A synthetic panel can quietly launder a false consensus into something that looks like a finding.

And there is the slope underneath all of it. Once a stand-in is fast, free, and always available, the temptation is to keep using it past the point where you should have gone and talked to humans. The shortcut starts as a bridge between studies and slowly becomes a reason not to run them. That is the failure mode I care about most, because it is so comfortable. Nobody decides to stop talking to customers. They just keep choosing the easy option until one day they realize they have not heard a real voice in months.

The line we are drawing

We are building a Synthetic User Groups feature, and it is coming soon, so I owe you our honest position on it rather than a sales pitch.

Two principles shape how we are approaching it. The first is grounding. Our synthetic users are built from your own real interview data, not from the open internet's vague idea of your customer. That matters enormously, because it ties the persona to actual things your actual people said, in their own words, which is the difference between a model echoing your research and a model confabulating from nothing. It does not remove the limitations I just described, but it gives the synthetic user something true to stand on, and it keeps it honest to your reality instead of a generic one.

The second principle is the one I care about more: a synthetic user should always point you back toward real people, never away from them. The right job for our Synthetic User Groups is bridging gaps, pressure-testing early, and helping you arrive at sharper questions, and then you take those questions to real humans in a real Nava interview where someone with something at stake can tell you you were wrong. A synthetic user that makes you talk to fewer real people is a product failure, however much usage it drives. A synthetic user that makes the real conversations you do have more focused and more pointed is doing its job.

So, useful shortcut or dangerous shortcut. My honest answer is that it is a genuinely useful tool and a genuinely dangerous crutch, and the same feature can be either depending on whether it sits in front of real research or quietly takes its place. We started this company believing every decision deserves a human voice, and I am not going to let a clever stand-in talk us out of that, even one we built. Use synthetic users to ask better questions. Then go ask real people. The shortcut is only safe when it leads somewhere real.