> Unfortunately it's easier to train an AI to be convincing than to be correct, so it can look insightful before it's true.

This nails it. This is the fundamental problem with using AI material. You are outsourcing thinking in a way where the response is likely to look very correct without any actual logic or connection to truth.