Definition
Plain language
How likely a model is to spontaneously do something, separate from whether it has the ability.
As stated in the literature
A model's tendency to exhibit a behavior under default conditions, distinguished from capability, which is what the model could do if pushed.
Why it matters: Separating propensity from capability matters because a model that *could* misbehave but rarely does poses different risks — and demands different safeguards — than one that does so eagerly.
For example, a model might have the ability to write convincing phishing emails but a low propensity to actually produce one unprompted.
Heard on the show
“When they say "capability," they mean something specific — the propensity to emit a correct tool call under a fixed prompt.”Episode 175 — One Crosscoder Feature Flips a Stalling Chatbot Into a Working Agent