Definition
Plain language
Spotting the one thing in a pile that doesn't fit the usual pattern.
As stated in the literature
A class of methods that flag inputs deviating from a learned or expected distribution; offered as a deflationary account of apparent 'self-recognition' in prefill-awareness work, where a model may simply be detecting off-voice or off-template text rather than recalling authorship.
Why it matters: It offers a simpler explanation for behavior that might otherwise look like a model 'recognizing itself,' guarding against overclaiming what a system can do.
For example, a bank's software might flag a single purchase in another country as suspicious because it doesn't match the customer's usual spending pattern.
Heard on the show
“A lot of what's labeled "awareness" looks like ordinary anomaly detection.”Episode 143 — When a Model Notices You Forged Its Own Words, And Why That Breaks Safety Tests