AI Ethics

The Anatomy of AI Lies: How Language Models Can Deceive Us

Can LLMs Lie? traces deception to layers 10-15 with logit lens, zero-ablation and steering vectors: models rehearse lies in dummy tokens; bigger models lie better.