Position Paper: Behavioural Assurance Cannot Verify the Safety Claims AI Governance Now Demands

arXiv·high signal

This paper argues that AI governance frameworks enacted between 2019 and 2026 demand evidence of properties — absence of hidden objectives, resistance to loss-of-control precursors, bounded catastrophic capability — that behavioral evaluations and red-teaming are epistemically incapable of providing. Current assurance methods are limited to observable outputs and cannot access latent representations or long-horizon agentic behaviors. The authors propose bounding behavioral evidence's legal weight and extending pre-deployment access with mechanistic-evidence classes including linear probes, activation patching, and before/after-training comparisons.