Measuring What AI Systems Might Do: Towards A Measurement Science in AI

Researchers argue that current AI evaluation methods fail to properly measure true AI capabilities and propensities, which should be treated as dispositional properties. The paper proposes a more scientific framework for AI evaluation that requires mapping causal relationships between contextual conditions and behavioral outputs, moving beyond simple benchmark averages.

• →Current AI evaluation practices conflate terms like capabilities, skills, and abilities without properly defining what they measure.
• →AI capabilities and propensities should be understood as dispositional properties with stable causal relationships to behavior.
• →Dominant evaluation approaches like benchmark averages and Item Response Theory fail to measure true AI dispositions.
• →Proper AI evaluation requires hypothesizing causal factors, operationalizing measurements, and mapping contextual variations to behavioral probabilities.
• →The research calls for more scientifically defensible AI evaluation methods grounded in philosophy of science and measurement theory.