Definition
Plain language
A short public document describing what an AI model is for and what it shouldn't be used for.
As stated in the literature
Structured documentation accompanying a released model: intended use, evaluation results, limitations, and safety caveats.
Also called: model cards
Why it matters: It is the main place a developer states, in public and in advance, what the model should and should not be trusted with.
For example, a released model might come with a page stating it is intended for drafting and summarising, and should not be relied on for medical advice.
Heard on the show
“If it didn't show up, nothing leaked, and that score goes into the model card.”Episode 246 — 160 Perfect Refusals, And The Refusals Were The Leak