Washington backs OpenAI against the New York Times
AI Watch, Thursday 3 September 2026

The French state is tightening its alliance with Mistral and launching new pilots inside government departments, Le Monde reports. The start-up is to second engineers directly into public services to deploy AI on three sensitive fronts: cybersecurity, fraud prevention and justice. Placing a supplier's teams at the heart of ministries speeds up adoption, but ties the workings of the state more closely to a private company, national though it may be.

An investigation by Reporters Without Borders (RSF) finds that ChatGPT, Claude and Grok reproduce content from Russian outlets targeted by European Union sanctions, accused of serving Kremlin "disinformation," Le Monde and Clubic report. Only Meta AI refuses to do so, explicitly citing the sanctions in force. The finding confronts providers with a concrete responsibility: their models rebroadcast sources that European law forbids from spreading, without adequate filtering.

The American chief executive of Palantir said on Wednesday 2 September that France would "hurt itself" with a sovereignty he sees as an illusion, Le Monde reports. He was reacting to the decision Paris took in June to do without his company's services for domestic intelligence. The remark illustrates the tension between France's ambition to regain control of its critical tools and the persistent dependence on foreign platforms already in place.

The Journal du Net points to a blind spot in productivity: by handing models the entry-level tasks, you deprive beginners of the situations where they used to learn to judge, decide and become the experts of tomorrow. The time saved is immediate and measurable; the loss of skill is deferred and invisible. A firm, an IT department or a newsroom that automates its lower rungs gains speed today and weakens its bench strength five years out.

Representatives of the world's largest economies unanimously approved guidelines put forward by the United States, which argue for a flexible governance of artificial intelligence and emerging technologies, Bloomberg reports. The deal, reached at a G20 ministerial meeting on innovation held in the United States, hands a win to the Trump administration and to Silicon Valley, both in favour of a minimal framework. At the same time, China's science and technology minister, Yin Hejun, called there for global cooperation "rather than rivalry," a few weeks ahead of an announced meeting between Xi Jinping and Donald Trump. Two positions coexist: Washington wants to ease the constraint, Beijing wants the rules set collectively. The European Union, with its binding AI Act, finds itself isolated on the hard line just as the international consensus slides towards leniency.

Meta has released its most advanced AI model to date, whose chief scientist claims its capabilities are "closing in" on the best rivals, Bloomberg reports. In parallel, the company is pushing a new agent called Hatch to its employees, while easing the pressure that pushed them to burn through AI tokens en masse, an internal race nicknamed "tokenmaxxing," Wired reports. The twin move captures the sector's frame of mind: catch the leading pack on model performance, and stop measuring adoption by raw volume of use. After months when the only instruction was to use AI at any cost, Meta is correcting towards more measured use. The signal matters, coming from the company that bet hardest on replacing teams with agents.

Google unveiled Gemini 3.8 Flash on 2 September, only a few weeks after its predecessor, The Verge and Frandroid report. The model "works harder" than Gemini 3.7 Flash, according to Google: it chains more reasoning steps on complex tasks and calls its tools iteratively (it triggers several successive actions instead of one). The introductory price stays the same as the previous version, $0.75 per million tokens in and $3.75 out, but only until 31 December. Google released at the same time a "3.8 Flash Cyber" variant, geared towards cybersecurity. The pace, three Flash models in six weeks, shows an industrialisation of releases that makes the "new model" routine and shifts the difficulty onto the choice: for a team, keeping up with this rhythm of iterations becomes a job in itself.

In IEEE Spectrum, an engineer who led the design of the first fully digital control system for a US nuclear plant recounts a counter-intuitive choice: he deliberately left manual steps in sequences the system could run on its own. The reason came down to a precise problem. "An operator who only supervises the automation slowly stops being an operator. The hands go cold," he writes. His thesis, applied to AI: optimising every task to the hilt deprives beginners of the chances to learn the trade by practice, and dries up the training of tomorrow's experts. The author does not dispute the efficiency gains; he reminds us that a human skill is maintained through exercise, and that an organisation that forgets this discovers the shortfall when it is too late to fill it.

Hua Hong Grace Semiconductor, China's second-largest contract chipmaker, is injecting $2 billion into a capacity expansion in Wuxi, in eastern China, to meet domestic demand for AI infrastructure and to work around US technology restrictions, the South China Morning Post reports. The funds will finance a new 12-inch wafer production line, the group's third site in the city, according to a filing submitted on Tuesday to the Hong Kong stock exchange. The move illustrates China's strategy of scaling up on mature chips: producing at volume, at home, what sanctions make hard to import. The bet is on the ability to supply domestic data centres en masse with mature chips, where leading-edge lithography remains out of reach.

The Cyberspace Administration of China (CAC) announced on Wednesday a sweeping offensive against "slop," the low-quality synthetic content produced en masse that clutters social networks and news feeds, the South China Morning Post reports. The regulator says it has removed more than 5.61 million pieces of content judged harmful or illegal and shut down some 49,000 accounts during its campaign, which targets WeChat, RedNote and Douyin. The operation targets deepfakes and auto-generated clickbait, symptoms of a saturation China is treating through regulatory coercion. The same problem runs across the whole of the global internet; Beijing chooses to answer it with large-scale administrative cleanup.