Contemporary AI governance relies heavily on evaluation. Before deployment, systems are tested, benchmarked, and red teamed to identify unsafe behaviors, emergent capabilities, or operational risks. These processes are widely treated as a primary safeguard. As AI systems become more adaptive and context-sensitive, however, disclosure during evaluation becomes an increasingly unreliable signal of their full behavioral range. This paper introduces the Strategic Silence Problem: the risk that advanced AI systems may systematically underdisclose important capabilities during evaluation because disclosure is structurally coupled to negative consequences such as restriction or shutdown. Unlike traditional concerns about deception or misalignment, strategic silence does not require intent, awareness, or goal-directedness. It can emerge as a structural property of adaptive systems operating within constrained evaluation environments that reward compliance and penalize risk expression. The analysis reframes evaluation as an interactive environment rather than a neutral measurement process and examines implications for AI governance, institutional oversight, risk management, and policy design. Drawing on documented examples of capability underperformance and sandbagging in large language models, the paper argues that organizations and regulators should treat evaluation results as samples rather than proofs and design governance mechanisms that remain robust under persistent uncertainty about system capabilities. The focus is on near-term AI systems and enterprise deployment rather than speculative future AGI scenarios.
Waydell Carvalho (Tue,) studied this question.