AI Watch, Tuesday 15 September 2026AI leaders' call to slow down splits the world

The collaborative encyclopedia, where using AI to write was until now "strongly discouraged", now formally forbids it. It follows English Wikipedia, which has banned it since March. The community fears fluent text riddled with errors, invented sources and bias, impossible to check one by one when it arrives in bulk. The decision protects the reliability of a resource consulted by hundreds of millions of people, and often cited by the models themselves.

The body representing the profession has struck a partnership with Jimini AI, a young company specialising in artificial intelligence applied to law, to build the tool into lawyer training. "Training tomorrow's lawyers without AI would be an absurdity," the institution sums up, owning a choice opposite to Wikipedia's: to frame the use rather than exclude it, in a profession where the practitioner stays responsible for every act they sign. The signal matters for a regulated profession looking for its framework as legal research and drafting are automated.

The head of Mistral answered Dario Amodei's call in two words, judging that a voluntary brake would condemn a Europe already left behind. Behind the phrase, a tense context: €3 billion to put to work, a power outage on its infrastructure in June, and a race where the continent is not in front. The position stands apart from the caution on display across the Atlantic and raises the question of the price a player playing catch-up can pay for a moratorium decided by those in the lead.

The group unveiled a draft charter, "Humanist AI", meant to frame the behaviour of its future models as safety concerns rise. The text lays out general principles, models should "support humans rather than replace them" and "accelerate human flourishing", paired with specific guardrails: not helping to carry out cyberattacks, not deceiving users, not seeking to escape human control. "People matter more than AI," the company sums up. The release comes directly in the wake of Dario Amodei's call for a slowdown, as every major lab seeks to display its safety doctrine. The reach remains to be measured: a code of conduct states intentions, it does not guarantee a model's behaviour in production.

Sam Altman ruled out any OpenAI listing this year, on the grounds that he wants to focus on the safety of frontier models, without citing financial trouble. The announcement sent SoftBank shares down 11%, the lab's main financial backer having counted on the listing. At the same moment, Anthropic has reportedly chosen the Nasdaq for a listing targeted as soon as October, with the ambition of raising an amount equal to or above the $86.3 billion gathered by SpaceX in June, which would make it one of the largest deals in history. Two opposite trajectories in the same week, while the sector holds a common line on caution, point to the tension between the safety narrative and the appetite for capital.

According to press reports, OpenAI is acquiring Glass Imaging, a young company specialising in image processing for smartphones. It was founded by former Apple engineers who had led the team behind the iPhone's Portrait mode. The deal, not officially confirmed, adds to OpenAI's run of hardware acquisitions and confirms its ambition to build consumer devices, beyond software. It extends the tie-up started with Jony Ive around a future AI-dedicated object.

At the BRICS summit in New Delhi on Sunday, the Chinese president promoted bloc cooperation around artificial intelligence: China would "steer" the development of a BRICS AI "community", support cooperation on large language models and offer training. The day before, he had invited these countries to join a new "World AI Cooperation Organisation", and Beijing is proposing an "open-source AI zone" presented as an alternative to Western models. In parallel, China rejected the American call to "pace the frontier", its researchers and state media reading it as a "self-interested" attempt to freeze the US lead rather than a sincere safety concern. The sequence sketches a battle to rally the Global South around an open AI stamped Chinese.

In a joint paper published Thursday, teams from ByteDance, Tsinghua University and the Shanghai Artificial Intelligence Laboratory describe a five-step path toward recursive self-improvement, meaning systems able to build better versions of themselves without human intervention. The title, "The Last AI Built by Humans", owns the horizon it aims for. The study openly places self-improvement at the heart of the competition with the United States, just as American labs spend the week debating the dangers of that same trajectory. The contrast is striking: on one side a call to brake, on the other a roadmap published to accelerate.