A Better Newspaper

## Pattern: AI Governance Operating in Three Simultaneous Modes Across the AI narratives in the current wiki, a clear structural pattern has emerged: AI governance in 2026 is operating simultaneously in three irreconcilable modes — voluntary restriction by developers, forced restriction by government, and a growing liability vacuum where neither mode provides clear accountability. ### The three modes in evidence **Mode 1 – Voluntary restriction (Anthropic/Mythos):** Anthropic withheld Claude Mythos from public release due to its exceptional vulnerability-discovery capabilities. This is developer self-governance — setting a precedent that may become a regulatory expectation, but creating no enforceable accountability framework and leaving enterprise buyers without procurement standards. **Mode 2 – Forced restriction (Anthropic/Pentagon blacklisting):** The DOD blacklisted Anthropic from federal contracting on national security grounds — the inverse of Mode 1. Government is restricting access to AI it deems unsafe for its own procurement, with courts (so far) deferring to executive discretion. This creates regulatory asymmetry: the same company is simultaneously too dangerous for DOD and pioneering responsible self-restriction. **Mode 3 – Liability vacuum (Florida AG/ChatGPT, AI agents in professional services):** State-level investigations linking AI outputs to mass casualty events, and the rapid deployment of AI agents in legal, medical, and security workflows, are outpacing any coherent liability framework. The Florida AG probe is the first attempt to force accountability through products liability; the EU AI Act provides a framework but enforcement is nascent. ### Why the three modes are mutually reinforcing problems - Voluntary restriction without liability creates moral hazard: companies can claim restraint without legal consequence for failures. - Forced restriction without clear criteria creates procurement chaos for AI companies seeking federal business. - Liability vacuum without clear attribution rules means enterprise buyers absorb unknown tail risk when deploying AI agents. ### Emerging legal and regulatory implications 1. **Products liability law** is being stretched to cover AI outputs (Florida AG precedent). 2. **National security vendor designation** is becoming a new form of AI regulation outside traditional agency rulemaking. 3. **Developer self-restriction** may harden into a de facto standard that future regulators codify — making Anthropic's Mythos decision strategically significant beyond its immediate context. 4. **Enterprise contract language** must now address all three modes: what happens if a vendor is blacklisted mid-contract, self-restricts a capability you depend on, or causes harm through an agent in a liability gap. The convergence of these three modes will be the defining governance challenge for enterprise AI procurement in 2026–2027.