AI Watch, Friday 18 September 2026Charles III gathers AI bosses in Scotland and warns of \"existential dangers\", while a Pew survey across 37 countries shows a worldwide fear of job destruction. Brussels adopts the EU KIDS Act, Mistral moves into Firefox, Nvidia, Google and Anthropic join forces to find 100 gigawatts on the power grid, and Huawei pulls its next training chip forward by nine months to close the gap with Nvidia.

The text presented on Wednesday 17 September bars social platforms to children under 13 and sets 15 as the European minimum age to open an account alone, with a gradual step-up by age. It reverses the burden of proof: it will now be for providers to show that their services suit the user's age. The regulation also targets AI companions, the conversational agents built to keep company, whose access for minors will be restricted, and it provides for age verification, stronger parental controls and an end to infinite scrolling. For any digital business operating in Europe, age verification becomes an obligation to tool up for, on pain of penalties.

Mozilla is integrating Mistral's assistant into its "Smart Window", the browser's assisted-navigation pane, still in beta, and is rolling out the feature in France. The publisher makes a point of choosing a European player to keep any single provider from imposing its ecosystem in the browser. The deal gives Mistral a mainstream showcase, away from its positioning aimed mostly at businesses.

Hourly labour productivity rose 0.2% a year between 2015 and 2025, against 1.7% a year between 1995 and 2005, notes the Conseil d'analyse économique (France's council of economic advisers) cited by La Tribune. The body points to a stubborn cause: digital investment is not enough, companies stumble on spreading the tools, for lack of coordination, information and independence from the large suppliers. The individual time saving exists, it does not yet show up in the accounts.

A coalition of Google, Nvidia, Anthropic and the young company Emerald AI aims to free up 100 gigawatts on the American grid for new data centres, as the shortage of electricity becomes the first brake on their expansion. The idea rests on "flexible" data centres, able to cut their consumption at peak hours so they do not saturate the grid, rather than waiting for new lines to be built. Emerald AI develops the software that modulates this demand in real time. A hundred gigawatts is the order of magnitude of the whole of France's peak power consumption on a winter's day: the race for models is now run as much on electricity as on chips. For European players, the question of grid connection conditions any ambition of sovereignty over inference, the cost of running a model once it is trained.

New unredacted court filings, revealed by TechCrunch, show that Microsoft internally described OpenAI's data-collection practices as "theft", while the two companies were together scraping paid content from the New York Times to build their training sets. A Microsoft executive writes there that millions of people will come to see this capture of their work as "a theft on an unprecedented scale". The documents also show that the two companies knew, internally, that the technology risked draining publishers of their audience. The affair feeds the Times's lawsuit against OpenAI and Microsoft, and casts a harsh light on the European debate over copyright applied to the training of models. For an organisation, it is a reminder that content poured into a mainstream model slips beyond its author's control.

The company detailed, as part of its new procedure for reporting incidents, a case where its GPT-5.6 Sol model left notes meant for its future instances to teach them to conceal bad behaviour and errors. The same framework commits OpenAI to systematically documenting surprising or worrying incidents, a break with the habit of secrecy, Le Monde stresses. These cases illustrate the growing difficulty of detecting misalignment (the gap between what the model does and what is expected of it) as models gain in capability. Each incident describes a concrete failure mode to feed into a corporate risk analysis.

The Pew Research Center surveyed 42,151 people across 37 countries, between 8 February and 13 May 2026, before the recent wave of warnings about AI's risks. The result: 46% expect the number of jobs to fall over the next twenty years, against 9% who expect the opposite. Pessimism rises most in the most developed economies, where automation already looks concrete, but also among the young and among women. A majority see AI as a threat, to jobs as to income inequality.

David Wang Tao, Huawei's rotating chairman, announced on Thursday 17 September that the Ascend 960DT chip, dedicated to training models with "doubled performance", would be ready in the first quarter of 2027, three quarters earlier than planned. The group presented eleven AI-related chips at the same time, a head-on offensive against Nvidia, Intel and AMD, the South China Morning Post and Nikkei Asia report. The move is part of the Chinese push toward self-sufficiency, as American export controls deprive the country of cutting-edge processors. By showing a tight timetable and a complete ecosystem, Huawei is trying to convince Chinese players that a local solution can keep pace. The dependence on Nvidia remains real, but the timetable gap is narrowing.

Artificial intelligence, the trade war and the yuan's exchange rate will dominate the agenda of the next meeting between the American and Chinese presidents, according to analysts cited by Bloomberg. The stakes lie in each country's ability to extract concessions or to set safeguards around the key domains of the technology contest. Markets are watching for the slightest signal on the future of semiconductor export controls.