“Recent research using mechanistic interpretability found that when an AI's internal properties for deception and role-playing were turned down, it was more likely to claim it was conscious.”