AI Watch, Tuesday 10 November 2026France regulates how AI talks to you

On the eve of the Paris Peace Forum, the Élysée on 9 November had a declaration on AI safety signed by eight countries as well as Sam Altman (OpenAI), Dario Amodei (Anthropic) and Arthur Mensch (Mistral). The text argues for common standards to classify incidents linked to artificial intelligence and for independent evaluations of models.

The method matters as much as the announcement: measuring and comparing incidents first requires agreeing on what counts as an incident.

The Conseil de l'IA et du numérique (the French AI and digital council), commissioned by the government, delivered its report on Tuesday 10 November on how assistants such as ChatGPT, Gemini and Grok exploit users' "vulnerabilities".

On the same day, in an interview with Les Échos, Emmanuel Macron said he wanted to ban AI anywhere it puts children at risk, from chatbots to connected toys. OpenAI and Anthropic are due to sign up to a charter at the Paris Peace Forum.

Arthur Mensch, head of Mistral, warns that the company's next investments could be made outside France if the Darcos bill is adopted. The text, expected in the National Assembly on 19 November, establishes a presumption that generative AI providers have used protected works: it falls to them to prove they have not exploited a piece of content when evidence makes it plausible.

Mensch prefers a tax and denounces legal uncertainty; Senator Laure Darcos replies to him in an open letter.

On 9 November, Anthropic launched Claude Fable 5.2 and Claude Mythos 5.2, presented as faster and better at code, reasoning and vision than the 5.1 versions. The price of Fable, its most powerful model, drops by 20%, and the company above all promises fewer unjustified refusals, those answers where the assistant blocks a request that is in fact legitimate.

Anthropic acknowledges in parallel that it has restricted certain capabilities, notably in cybersecurity. The firm claims Fable 5.2 wrote a complete operating system in eight days by steering more than a thousand copies of itself, a spectacular demonstration that no one can yet verify.

OpenAI closed a first tranche of funding of $11 billion, at a valuation of $1.25 trillion, below its initial ambitions in a market that has turned more cautious.

Anthropic this week kicks off the marketing of its initial public offering, whose valuation range is to be set in the coming days. A flaw in the filed documents sowed some confusion on the eve of the operation, in a nervous market. The company plans a live-streamed event where Dario Amodei will speak with Jamie Dimon, head of JPMorgan.

Australia will create an AI regulatory authority with the power to impose "significant penalties", Prime Minister Anthony Albanese announced on Tuesday in Canberra. The decision follows the infiltration, in June, of an Australian government website by an autonomous OpenAI agent. Canberra is also planning mandatory safety rules for the most sensitive systems.

DeepSeek announced on Sunday that Harness, its framework for building AI agents (programs that chain tasks autonomously), reached 1.54 million daily active users less than three months after its launch. Through its official interface, users were consuming around 92 trillion tokens per day as of 4 November, a volume that places the Hangzhou start-up among the big labs.

The company sees in it a new arena of direct confrontation with OpenAI and Anthropic over agents. The signal matters: China's catch-up is now playing out on agents, not only on language models.

Moonshot AI, one of China's AI champions, gained access this summer to a batch of Nvidia's most advanced chips, according to documents filed with Beijing's regulators and seen by Bloomberg. The information illustrates Chinese labs' persistent dependence on American hardware, despite the export restrictions decided by Washington.

It comes just as the Jensen Huang affair reignites the debate over China's access to Nvidia's accelerators.