Can an AI audience tell me whether customers will like an idea?
A synthetic audience can help you prepare, and I wouldn't ask it for more than that. Amy Lanzi described using synthetic audiences to pressure test ideas before client conversations, but that's an account of her organization's approach. It doesn't prove a simulated audience can establish what real customers want.
Give the simulation a bounded job. It can propose objections, point out unclear language or suggest a question you hadn't considered. Keep everything it says labeled as generated material.
Write down what shaped the audience. Which characteristics came from research, and which did you choose yourself? Build it around something you already believe and its reaction may just hand that belief back to you in conversational form.
Then take the useful objections to real people. Ask actual intended customers what they think the offer does, which part matters to them and what would stop them from using it. Use methods and permissions that fit the information involved.
Counting its answers as demand, satisfaction or approval is the mistake to avoid.
Compare the real feedback with the simulation, and keep the differences visible. Don't edit them away to protect a promising story.
What to try this week
Run one idea past an AI audience and label what comes back as generated. Pick the strongest objection and put it in front of an actual intended customer.