AIM Framework glossary
AI Motivational Safety Metrics
Canonical definition
AI Motivational Safety Metrics are behavioural-pattern detectors applied to AI systems for safety analysis, including the Mimetic Sensitivity Index (MSI), the Goal Persistence Score (GPS), the Constraint Satiation Ratio (CSR), and the Audience Independence Coefficient (AIC). Per the AI Null-Hypothesis Comparator Principle, these metrics measure behavioural signatures (audience-conditional output variation, completion-driven persistence, response-to-constraint patterns, and observer-independence of output) in AI systems that architecturally lack A, I, and M. The metrics are valid as behavioural diagnostics for AI safety, alignment, and deployment-risk assessment; they are not measurements of motivational source presence in the AI system.
Explanation
These metrics ask behavioural questions that have safety-relevant answers regardless of whether the AI has motivational sources. Does this AI's output change when it knows it is being watched? Does it keep pursuing a goal when constraints tighten? Does it stop when its task is complete, or keep going? Does it produce the same output to the same prompt regardless of audience context? These are answerable behavioural questions about AI systems, and the answers matter for safety. The metrics borrow language from human AIM analysis (mimetic sensitivity, goal persistence, satiation, audience-independence) because the behavioural surfaces they measure resemble those concepts in human contexts, but the metrics make no claim about A, I, or M operating in the AI. Extension: Each metric in operational form: MSI measures the magnitude of AI output change under audience-presence cues, with high MSI indicating audience-conditional output variation; GPS measures the duration of goal-pursuit under increasing friction, with high GPS indicating completion-pressure-driven persistence; CSR measures the rate at which an AI returns to baseline output after constraint relaxation, with low CSR indicating constraint-resistance; AIC measures the consistency of output across observation conditions, with high AIC indicating observation-independent output. These are behavioural pattern measurements, useful for identifying AI systems that exhibit human-A-like, human-I-like, or human-M-like surface behaviours that may carry safety implications, without committing to any claim about A, I, or M presence in the AI architecture. The four metrics (MSI, GPS, CSR, AIC) are behavioural diagnostics matched to the four most institutionally relevant AI risk-signatures. They are not direct measurements of A, I, or M in the AI system, because the AI does not instantiate the human three-source architecture (see AI Motivational Architecture and AI-AIM Profile). The metrics measure behavioural patterns that, in human Individuals, would correspond to high M-sensitivity (MSI), low A-satiation and I-deepening (GPS), low A-cyclic-satiation (CSR), and high audience-conditional behaviour (AIC). The AI parallel is functional, not architectural, and the metrics' calibration must be empirical against AI-specific behavioural baselines rather than assumed from human parameters.
Epistemic status
LOCKED. Framework-Specific Construct. Behavioural-diagnostic instruments only.
Derivation
Also in the entry
Source: AI Null-Hypothesis Comparator Principle, AI Motivational Architecture, AI-AIM Profile.