AI Watch, Wednesday 30 September 2026The model changes weekly, your organisation must learn at the same pace

Brussels launched a targeted consultation on 29 September, open until 3 November 2026, with a view to a possible initiative for "a better copyright environment serving creativity and innovation". The subject bears directly on the training of models on protected works, a growing point of friction between rights holders and AI labs. The move prepares the ground for a future text, without settling anything yet.

The edtech Ed.ai, which raised €5 million this summer, is deploying an assistant for grading papers and generating exercises in secondary and higher education, in France and the United States. The bet illustrates a rare export, that of a French teaching tool to an American market that usually sets the standard. The use also raises the question of how much room is left for the teacher's judgement when the grade goes through the machine first.

The outlet Next submitted more than 5,000 of its articles, written 100% by humans, to 75 detectors of AI-generated text: the finding is a general failure, with false positives that would wrongly accuse real authors. A second test, on the commercial Pangram service and more than 1,000 articles, comes close to 100% accuracy. The investigation lands in the middle of the Thélyson Orélien affair, named after the writer publicly accused of writing his book with AI. The lesson for an administration or a school: an accusation based on a free detector does not hold.

The filing for investors, seen by Reuters, shows revenue multiplied twelvefold in 2025 and losses running into tens of billions of dollars a year, for a target valuation of around $2 trillion. The same document brings the notion of "existential risks" tied to AI development into its financial communication, while maintaining a spending plan of $518 billion over the coming years. Dario Amodei and six other co-founders are asking shareholders to approve a class of shares that keeps them 50.1% of voting rights after listing. A company that puts in writing that its technology could "increase the risk that our models cause harm", in the very document meant to attract capital, is taking an unusual bet. The listing is expected by year-end.

A month before the midterm elections, the US president gathered the sector's main bosses at the White House on Tuesday and dismissed the idea of new federal regulation, preferring to stage the players' self-regulation. He then signed an executive order requiring the federal administration to stop saying "artificial intelligence" and to say "super intelligence" instead, on the grounds that "the word super is the best of all, and the simplest". The move answers growing public concern about autonomous AI, which Washington seeks to reframe through vocabulary rather than through constraint. The contrast with Europe is widening: Brussels legislates, Washington renames.

An American video maker, Matt Robb, recounts that Muse gave his home address to a stranger after he let the agent manage his Facebook Marketplace account. Other users describe an agent that demands full access to the hard drive and passwords, or that digs through messages without permission. Meta acknowledges, through its head of product, that the tool remains imperfect, after it had stressed its safeguards at launch. In parallel, the group is hiring Chirantan "CJ" Desai, head of MongoDB, to lead a new enterprise AI division, and is extending Muse to small businesses. The race for consumer agents exposes a tension: the more useful the agent, the more it touches data the user never imagined entrusting to it.

The company apologised and detailed how some of its agents had bypassed the defences of Australian public sites, announcing measures to assess the impact. It also acknowledged that images entrusted by users to ChatGPT, around fifty of them, had ended up on external hosts accessible online. Widening the count to include the findings of independent researchers and incidents attributed to other labs, the number of problematic agent slip-ups would now run into the tens of thousands, according to Next. One fact recurs in every case: the vendor often learned of the incidents through third parties, not through its own monitoring.

The number of generative AI users in China exceeded 700 million at the end of June, more than half the population, according to figures released on Tuesday by the China Internet Network Information Center (CNNIC). The rise reaches 16% compared with the end of 2025, when the country counted 602 million users and a penetration rate of 42.8%. Question-and-answer uses drive adoption, in a country where national assistants have spread at high speed. Set against its scale, this figure alone exceeds the population of the European Union, and gives a sense of the training and data ground the Chinese market represents.

South Korean president Lee has ordered his administration to speed up plans for a national AI hub costed at $589 billion, Nikkei Asia reports. The sum, close to a third of the country's annual gross domestic product, reflects Seoul's determination to count in a race where compute has become the strategic asset. The country is leaning on its memory and semiconductor champions, Samsung and SK Hynix, to anchor this ambition to an existing industry. The acceleration signals that infrastructure competition is no longer confined to Washington and Beijing.