Two of the world's leading artificial intelligence companies, OpenAI and Anthropic, are taking different positions on how dangerous AI systems should be stopped, exposing a growing disagreement over who should have the authority to intervene when advanced models pose serious risks.
The proposals emerged in separate submissions to California Governor Gavin Newsom following an executive order issued on September 18, 2026. The order directed state officials to examine additional safeguards for frontier AI systems, including independent safety evaluations, incident reporting and mechanisms for shutting down unsafe AI operations.
Anthropic Wants a Legal Route to Suspend Dangerous AI Models
Anthropic has proposed giving California's attorney general the authority to seek a court order suspending an AI model if independent evaluations indicate that it presents a near-term risk of catastrophic harm. Such an order could potentially cover a model's deployment, internal use or further development.
The company also supports stronger independent verification of AI developers' safety claims. Its recommendations include regular risk reports and giving qualified external evaluators access to advanced models and relevant information needed to assess potential dangers.
Anthropic's approach places greater emphasis on independent scrutiny and the possibility of legal intervention when a developer's safeguards are considered insufficient.
OpenAI Favors Enforceable Shutdown Controls
OpenAI supports a different approach. Rather than relying on a single shutdown mechanism or focusing primarily on court-ordered suspensions, the company recommends requiring frontier AI developers to maintain effective shutdown controls and use them when unsafe conditions arise and less disruptive measures are inadequate.
These controls could include stopping an AI training run, isolating an autonomous AI agent, revoking its access to external tools or networks, or suspending a particular service. OpenAI argues that the appropriate intervention should depend on how an AI system operates and which organization has control over it.
The company also supports independent technical assessments and compliance audits. Its proposal distinguishes between evaluating whether safety measures actually work and checking whether companies comply with their stated safety commitments and legal responsibilities.
Both Companies Support Stronger Independent Oversight
Despite their differences, OpenAI and Anthropic share several priorities. Both support independent assessments of advanced AI systems, stronger safety reporting and mechanisms to respond when models behave in ways that create serious risks.
The disagreement is largely about how those protections should work in practice. Anthropic emphasizes independent verification and a potential legal route to suspend dangerous systems. OpenAI focuses on operational controls that developers and other responsible organizations can use to contain unsafe activity.
Why the Debate Matters for the AI Industry
The discussion has implications beyond California. As AI systems become more capable and are increasingly connected to software tools, business processes and external services, regulators and technology companies face difficult questions about accountability and emergency intervention.
A model developer may control the underlying system, while a cloud provider, enterprise customer or third-party operator controls parts of its deployment. That division can make it difficult to determine who should act when a system behaves unexpectedly or creates a serious safety concern.
OpenAI and Anthropic's proposals offer competing approaches to this problem, but neither should be mistaken for an already enacted statewide shutdown requirement. California officials are considering recommendations, and the proposals would need further policy or legislative action before becoming binding rules.
What Comes Next
California's review could help shape future requirements for independent AI evaluations, incident disclosures and emergency safeguards. The state's recommendations may also influence broader discussions about how governments should oversee companies developing increasingly powerful AI models.
For now, the disagreement between OpenAI and Anthropic demonstrates that agreement on the need for AI safety does not automatically translate into agreement on who should control the response when a system becomes dangerous. The next stage of the debate will likely focus on enforcement, independent oversight and clearly defined responsibilities across the AI supply chain.
