Definition
A structured process some AI companies now use to voluntarily publish cases where their own AI systems behaved in unintended, deceptive, or unsafe ways ('misalignment'), similar to how software companies disclose security bugs. It creates a standing, repeatable record instead of a one-off announcement after an incident.
Why it matters
It gives boards, customers, and regulators an early-warning channel into how often and how seriously an AI vendor's models misbehave — a new due-diligence input when choosing or overseeing AI vendors.