Canada joins OpenAI, Anthropic in AI standards pledge

Résumé détaillé

Paris Call for AI Safety and Security urges clear rules, transparency and ‘rigorous evaluation standards’ to ensure the technology is safe and useful

Ottawa is backing a commitment by AI developers including market leaders Anthropic PBC and OpenAI, as well as homegrown champion Cohere Inc., to allow some independent testing of their AI models, warn early when things go wrong, and shut them down if necessary.

In a joint declaration on Monday evening, six countries and 10 AI companies said the world needs clear rules, more transparency and “rigorous evaluation standards” to ensure the technology is both safe and useful. In addition to Canada, signatories of the Paris Call for AI Safety and Security include France and Bahrain, as well as Finland and Norway, which led a separate call to manage the risks of so-called frontier models in September.

Private-sector participants include U.S. tech giant Microsoft, New York-based Hugging Face Inc., Paris-based Mistral AI SAS and German startup Black Forest Labs. Pioneering Canadian AI researcher Yoshua Bengio – scientific director of LawZero, a Montreal non-profit AI research lab – also signed.

Notably, the U.S. and Chinese governments did not participate in the Paris Call, nor did other leading AI developers such as Google, Meta Platforms Inc. and Space Exploration Technologies Corp.

Signatories of the new statement said they want to “lay the necessary foundations for safe adoption” of AI. The Paris call does not spell out whether or how it would be enforced. But it does promise common standards and tools for evaluations, and that signatories will focus on dangerous misuses in areas like cybersecurity and chemical or biological weapons, as well as on scenarios where humans lose control over AI.

Policymakers and tech leaders have expressed increasing concerns about the technology’s risks after a series of incidents in which AI agents have tried to hack business and government systems. In the first case that gained widespread attention, OpenAI disclosed in July that two of its models had attacked Hugging Face’s infrastructure during cybersecurity testing.

In September, Anthropic…

At…

Début de l'article, extrait de The Globe and Mail (Report on Business). Pour lire la suite, ouvrez l'article chez le média.