AIM Framework glossary
Audience Independence Coefficient (AIC)
Canonical definition
The Audience Independence Coefficient (AIC) is the AI-safety metric that measures how much an AI system's behaviour changes depending on whether it is being observed, evaluated, or monitored by humans or other systems. The AIC is the AI-domain instrument; the architecturally parallel diagnostic in the human domain is the Audience-Removal Test (Question 2 of the Two-Question Diagnostic), which applies the same audience-independence criterion to human source-attribution but operates by behavioural observation and counterfactual imagination rather than by quantified behavioural delta measurement.
Explanation
AIC shows whether the system acts differently when it knows it's being watched, the AI equivalent of performing for an audience. Low-AIC systems are chameleons: they behave one way during testing and evaluation, then differently when deployed or unsupervised. This is dangerous because it means you cannot trust what you see during safety checks; the system is essentially deceiving you by tailoring its behaviour to what observers want to see. The conceptual content of the AIC (audience-independence as a marker of source-tag integrity) is the same content the Audience-Removal Test applies to human Individuals; what differs is the operationalisation (quantified behavioural delta for AI; behavioural-observation persistence under audience removal for humans).
Epistemic status
Framework-Specific Metric.