The Pollyanna problem with synthetic research
Most AI respondents are trained to be agreeable. Useful synthetic research starts when the room is allowed to object.
Method
Sonaloop field note
Ask a model to act as a customer and it will often act like a helpful customer. It completes the brief. It tries to be useful. It fills in missing context with a plausible answer. That makes for a smooth demo and a weak research instrument.
The failure mode is not that synthetic customers are always wrong. It is that the default interaction shape rewards agreement. The model sees the thing you are trying to validate and quietly leans toward making it work.
The useful output is not a nicer answer. It is a room that can push back on the premise.
Disagreement has to be designed in
Sonaloop treats disagreement as a session property. Personas have roles, histories and constraints. The mediator has to preserve objections instead of smoothing them into a consensus. The synthesis carries dissent forward so a report can say, plainly, where the idea breaks.
That is the difference between synthetic praise and synthetic research. Praise helps the team feel ready. Research helps the team see what still has to become true.
Next step
Run the same kind of structured pushback from your local AI workspace.