Type 2: Bigger than expected

Mitigation strategy

1. **Implement Formal Second-Order Consequence Analysis and Iterative Risk Tiering:** Mandate a structured framework to evaluate and document potential second- and subsequent-order impacts (e.g., societal, environmental, ethical) and classify AI systems into risk tiers at defined lifecycle points (pre-training through post-deployment). This proactive approach forces the anticipation of non-obvious, large-scale unintended consequences and ensures mitigations are commensurate with the system's evolving risk profile. 2. **Enforce Strict Access Control and Capability Restriction for Dual-Use Systems:** For AI systems with capabilities in high-risk dual-use domains (e.g., cybersecurity or biological research), implement rigorous access controls and security protocols. Proactively remove or restrict dangerous functionalities from models intended for general or open-source release to prevent unexpected repurposing that could result in catastrophic harm. 3. **Establish Clear Developer Accountability and Liability:** Institute a governance framework that enforces explicit accountability and, where appropriate, a strict liability regime on developers of general-purpose AI systems. This measure is intended to incentivize greater investment in safety-oriented organizational culture, rigorous audits, and robust risk management throughout the entire development lifecycle, thereby reducing the probability of large-scale accidents or unanticipated misuse.