AI Labs Pledge Fail-Safe Shutdowns in a Paris Accord

Some of the largest names in artificial intelligence have put their signatures to a document that asks them to build an off switch. On 9 November, at a gathering convened on Emmanuel Macron's initiative, eight national governments and ten AI companies signed the Paris Call for AI Safety and Security, a voluntary declaration built around a handful of commitments the industry has spent the past year arguing about in public.

The line drawing the most attention is the one on shutdowns. Developers who sign on commit to putting in place fail-safe mechanisms that can shut down model operations within their control, and the text is explicit that this covers training runs and internal use, not only the versions customers see. It is a direct response to a worry that has moved from science fiction into corporate risk filings, including Anthropic's own warnings about losing control of a sufficiently capable system.

What the signatories agreed to

The declaration is organised around three commitments. The first is on evaluation and risk management: a shared framework for testing frontier models, with particular attention to misuse, including cyber and chemical or biological risk, and to questions of autonomy, misalignment and loss of control. The second is on responsibility, holding each actor in the chain accountable according to its role. The third is on information sharing, through incident reporting, a common way of classifying incidents, and rapid alert systems.

Alongside the shutdown pledge, developers agreed to disclose model capabilities, limitations and known risks, to publish non-sensitive evaluation findings in a standard format, and to support review by qualified independent evaluators. The text sets no technical standard and no deadline for any of this, which is the catch that critics reached for first.

Who signed, and who did not

The governments backing the call are Bahrain, Canada, Finland, France, Norway, the Netherlands, Cyprus and Kenya. On the company side the signatories include Anthropic, Black Forest Labs, Cohere, Domyn, Hugging Face, Microsoft, Mistral AI, OpenAI, Sakana AI and Synthesia, a roster that pairs American and European labs with smaller specialists.

Two names are missing. Google and Meta, both of which ship frontier models at the same scale as several signatories, do not appear on the list. Neither company has said publicly why it stayed out, and the absence matters: a voluntary pledge that two of the field's heaviest builders decline to join tells you something about how far an industry consensus really reaches.

A pledge without teeth, for now

None of this is law. The Paris Call commits companies to principles, not to audited compliance, and the document leans on good faith at exactly the points where good faith has been thin. A promise to be able to shut a model down is only as good as the test that proves it, and the text does not say who runs that test or what counts as passing.

Still, the symbolism is hard to wave away. A year ago the debate over whether frontier systems should carry a guaranteed stop was treated as a fringe concern, closely tied to the contested question of whether an AI would resist being switched off at all. Now the companies building those systems have signed their names to the idea that the switch should exist. The harder work, turning a shared paragraph into a shared standard, sits with the same governments and labs that have spent this year disagreeing about almost everything else, among them Australia's move to a regulator with real penalties.