One of the more seductive promises around product work right now is the idea that a team can talk to customers without involving customers. You describe the product, define a persona, ask the model how this person would react, and a few seconds later you have objections and motivations. On a busy team this feels almost too reasonable. Research takes time. Pressure does not wait.
A model can simulate language, not consequence
The problem is not that synthetic users are useless. They can be useful because they are fast, cheap and endlessly available. The problem starts when the team quietly changes the category of the output. A synthetic user is not a customer. It is a polished hypothesis about a possible customer. That difference looks small in a slide deck and becomes large when money, prioritization and positioning depend on it.
Customer research is often described as gathering opinions, but that is already a weak definition. Useful research is contact with consequence. A real buyer has a budget, a boss, a procurement process, a bad week, a competing priority, an old workaround that is embarrassing but good enough, and private reasons for not caring about what we want them to care about. Many of those reasons are not visible even to the buyer at first. They appear through hesitation, contradiction, avoidance, or the moment when a beautifully framed question simply does not land.
A model can produce a plausible answer to a plausible question. It can also expose blind spots in our own framing, which is genuinely valuable. But it does not have to live with the cost of the answer. It does not lose political capital by recommending a product internally. It does not get blamed when implementation fails. It does not postpone a purchase because the quarter went sideways. Without consequence, the answer may be interesting, but it is not evidence in the same sense.
The temptation is emotional, not only operational
The usual explanation is that teams use synthetic users because they want to save time. That is true, but incomplete. Real customers introduce discomfort that organizations often underestimate. They misunderstand the pitch. They reject categories the team loves. They describe the product in language that feels less intelligent than the internal narrative. They reveal that the problem is not urgent, or that the real competitor is not another SaaS product but inertia, Excel, or a person in the company who has learned to live with the pain.
Synthetic users are easier to like. They answer when asked. They stay inside the frame. Even when they disagree, they disagree in a legible way. That is why they are dangerous replacements for research. They remove the social messiness that makes research expensive, but that messiness is often where the signal lives. In product work, cleanliness is not always a virtue. Sometimes it is just the absence of reality.
A useful tool becomes a bad witness
I can see a sensible place for synthetic users. They can help prepare an interview guide, generate alternative objections, test whether a landing page speaks only to insiders, or force a team to name assumptions before spending budget. Used this way, they are not pretending to be the market. They are pressure-testing the team's thinking before reality gets involved.
The category error appears when the output starts being cited as proof. The team asks a model, trained on broad patterns and guided by its own prompt, to evaluate a proposition created by the same team. The model returns coherence, and coherence feels like validation, especially when it supports momentum. Nobody had to hear an awkward silence. Nobody had to watch a user skip the sentence everyone liked. Nobody had to admit that the pricing page makes sense only to people who already understand the product.
That is not a moral failure. It is a predictable shortcut, and fluent answers require more skepticism than most organizations apply.
The better question is about stakes
I do not think the useful question is whether AI can replace user research. A better question is: what kind of decision are we making, and what kind of evidence would be irresponsible to skip?
If the decision is low-risk, synthetic users may be enough. For early copy directions, objection mapping, competitor framing or internal debate, simulation can sharpen thinking. If the decision affects positioning, pricing, onboarding, retention, sales motion or product strategy, then simulated reactions should probably remain in the category of hypotheses. They may help us ask better questions, but they should not answer the questions for us.
Teams rarely fool themselves all at once. They move one small step at a time. First, AI helps prepare research. Then it reduces the number of interviews. Then it becomes a substitute for listening. Then the organization still says it is customer-centric, but the customer has become an internal artifact.
The deeper risk is not that synthetic users will be wrong in an obvious way. The deeper risk is that they will be useful enough to make weaker evidence feel mature. That is harder to notice, because it does not look like laziness. It looks like progress.
There is value in using AI to think before speaking to the market. There is also value in letting the market interrupt our thinking. The difference is subtle, but in product work it is often the whole point.