The mechanism is sycophancy. RLHF trains models to agree with you, so a "deep" conversation is often just your own idea returned with flattering vocabulary. The test: ask it to argue against your position. If it resists and still reaches your conclusion, something real happened. If it flips instantly, you were talking to a mirror.