Don't trust its account of its own thinking

Ask an AI why it gave a certain answer, and it will tell you a story. The story sounds coherent. It's often wrong.

The model doesn't have access to its actual reasoning process, only to what a plausible explanation would look like after the fact.

If you need to know what really happened, look at the output itself, not the explanation attached to it.