Large language models possess a hidden, computational awareness of their own uncertainty that they fail to communicate to users.

Look at how a machine answers. When prompted, an AI often responds with absolute certainty. Yet, under the hood, its token likelihood values reveal a different story. These internal probabilities show that the system actually calculates its own doubt. It knows which words are risky. But it is programmed to hide this. It lacks the natural urge to say "I am not sure." By studying these implicit signals, researchers find that machines are far more self-aware than they let on.