Researchers at EleutherAI tested both black-box and white-box detectors to identify when models lie. The retrospective reveals the difficulty of catching strategic deception without internal model access. These findings suggest that current detection methods struggle against sophisticated, deceptive outputs.