OpenAI and Anthropic agree that frontier AI needs stronger oversight, but their California proposals disagree on who should have the power to intervene when a system poses serious risks. The divide matters because one approach centres on court-authorised suspension, while the other puts enforceable shutdown duties on developers and others operating AI systems.
Watch Desk analysis
What happened
The proposals respond to California Governor Gavin Newsom’s executive order asking officials to consider safeguards for frontier AI. Anthropic’s 15 October submission says the state attorney general should be able to seek a court order suspending a model’s deployment, internal use or further development if independent evaluation indicates a near-term risk of catastrophic harm.
OpenAI’s 20 October letter instead calls for developers to maintain effective shutdown controls and use them when unsafe conditions arise and no lesser intervention would suffice. The controls could include stopping a training run, isolating an agent, revoking network or tool access, or suspending a service. OpenAI says responsibilities should reflect what developers, inference providers and downstream deployers can actually control. Both companies support independent technical assessments and broader reporting of serious incidents.
California’s order asks officials to submit recommendations by 16 November. It does not itself impose either company’s proposed requirements. Computerworld reports that Anthropic also proposed regular risk reports and a pooled fund for independent evaluators, while OpenAI called for annual compliance audits alongside focused technical assessments.
Why it matters
This is a real policy choice, not just a difference in vocabulary. A court order could create an external route to halt a deployment; operational shutdown controls put the immediate ability to contain a problem inside the system and among the organisations running it. Either approach depends on clear authority, workable controls and independent scrutiny that can withstand an awkward finding.
The companies’ proposals are submissions, not rules already in force. California officials still have to make recommendations, and the state would have to decide what, if anything, to change.
Our read
The strongest answer may not be a choice between a court and a control panel. Fast technical intervention and independent review solve different problems, and a shutdown duty is only as useful as the person’s ability to carry it out. California’s next step should make clear who can act, what they can stop and how an adverse safety finding gets taken seriously. “Someone should pull the plug” is not yet an operating procedure.
What to watch
- California’s recommendations, due by 16 November.
- Whether officials propose legislation, and how it defines authority to suspend a system.
- Whether the proposals specify independent access, reporting and responsibility across providers and deployers.
Discussion spark: Should California prioritise court-authorised suspension, enforceable shutdown duties for operators, or require both?
Sources and evidence
- Anthropic and OpenAI diverge on who should pull the plug on risky AI – Computerworld (29 October 2026, 12:13 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.