Someone forwarded your boss a demo. Feed the model a persona, ask it a hundred questions, and out comes a tidy chart — no recruiting, no incentives, no three-week timeline. It’s a genuinely seductive pitch, and dismissing it as hype is how you lose the argument. Here’s how to make the case that the sample still has to be made of people.
Name what’s real about the appeal
Start by conceding the parts that are true. Synthetic respondents are fast, they’re cheap, and they never ghost a scheduled call. Open with “AI is dangerous” and you sound like every incumbent who missed the last shift. Agree the constraints are real — then move the conversation to where the risk sits.
A model can only return the consensus
A synthetic respondent is a very good average of things already written down. Ask it about your category and it confidently interpolates the public internet as of its training cut-off — precisely the answer you don’t need. Research earns its budget on the days it tells you something you didn’t already believe.
“A synthetic respondent will never call your roadmap a mistake. It was trained to agree with the premise of your question.”
There’s no one to go back to
Real research leaves a trail: a consenting person, a recording, a recruit you can recontact when a finding raises a new question. Synthetic data has no provenance. When a stakeholder asks “who said that, and what did they mean?” — there’s no one to ask. You can’t audit a hallucination.