Canada signs a declaration on AI safety with OpenAI, Anthropic and other companies
IA en un minuto newsroom · Editor: Jon Elgezabal
In 30 seconds
France, Canada, Norway, Finland and Bahrain are among the six governments backing the new joint declaration, alongside ten AI companies. OpenAI, Anthropic and Cohere agree to open their models to some outside testing, raise the alarm early about problems and disconnect them when necessary. The United States and China stayed out, as did Google and Meta, and the document does not explain how compliance will be monitored.
Six countries and ten AI companies have signed a joint declaration that promises rigorous standards for evaluating models. Canada is one of them.
According to The Globe and Mail, the declaration, the Paris Call for AI Safety and Security, was unveiled on Monday night. Besides Canada, Norway, France, Finland and Bahrain have signed it, and the companies include OpenAI, Anthropic, Microsoft, Mistral AI, Hugging Face, Black Forest Labs and Canada's Cohere. The researcher Yoshua Bengio has also put his name to it. The companies pledge to allow some independent testing of their models, to give early warning when something fails and to switch them off if needed.
The text promises shared evaluation standards and tools, and says the signatories will concentrate on dangerous misuse in fields such as cybersecurity and chemical or biological weapons, as well as on scenarios in which people lose control of AI. It does not make clear whether or how it would be enforced. The governments of the United States and China did not take part, and neither did Google, Meta or SpaceX.
The Globe and Mail notes that it belongs to two groups of publishers that have sued Anthropic and OpenAI before an Ontario court for using their content without permission.
Why it matters · analysis and opinion
A declaration like this binds no one: it carries no penalties and no body to check compliance, which makes it closer to a public promise than to a rule. Its value lies in what it puts in writing, such as accepting that third parties examine the models and that they may have to be stopped, because later on those commitments can serve as a yardstick. The absences weigh as much as the signatures: without the two governments with the most weight in AI or several of the largest developers, it is hard to speak of a common standard. For companies that use these models nothing changes today; it is worth watching whether those independent tests end up being published, because they would allow providers to be compared on something more than their own word.
Source: The Globe and Mail · Written with the help of AI: how we make the news
