AI Watch, Tuesday 22 September 2026More than twenty countries call for stronger human oversight of AI, without the United States, China or France, while Washington and Beijing open a warning channel between them ahead of the Trump-Xi summit. Google launches its Googlebook at $899 to sell Gemini, Meta's Muse agent races to the top of the charts before being blocked by Amazon and flagged for a security flaw, and British Columbia sues OpenAI after a mass shooting.

Since 27 July 2026, the European regulation no longer asks for a "sufficient level" of AI literacy among employees, but the obligation to act remains fully in force, the Journal du Net analyses. For an SME, the practical question is no longer ticking a compliance box, it is proving that its teams know how to use the tools it deploys and understand their limits.

The sum, planned by 2028, will fund data centres built to absorb global AI demand, ZDNet reports, describing a Finland on its way to becoming a Nordic computing hub. The choice follows low-carbon electricity and a cold climate, two assets for cooling server rooms, and confirms that Europe's AI battle is fought as much over land and energy as over models.

To preserve their online presence, brands are building their own shopping agents and trying to improve their visibility in AI engines, Le Monde describes. The stakes are concrete: when a customer asks an assistant for advice rather than a search engine, the brand missing from the answer drops out of view at the moment of decision.

The US giant unveils a range of laptops where Gemini is stitched into the system: the model is wired to the cursor, dictation, widgets and settings, TechCrunch describes. These machines merge Android and ChromeOS and target the premium market held by Apple's MacBook and Windows PCs. Sameer Samat, president of the Android ecosystem, embraces the pairing with the phone, in the manner of the continuity between the iPhone and the Mac. The bet is clear: convince a user to buy a new device for the assistant that lives on it rather than for raw power. Google had failed to make the Chromebook a mainstream success, and it is trying again with AI as the central argument.

Muse, Meta's new consumer assistant, drew more downloads and daily active users in the United States and Canada than ChatGPT over the same period after its mobile launch, according to Appfigures estimates cited by TechCrunch. The success hit a wall: Amazon cut off Muse's access to its pages, with a message warning that "continued access by an unauthorized AI agent violates Amazon's terms of service", The Verge reports. Meta had not warned Amazon that its agent was coming. At the same time, Ars Technica describes a serious zero-day flaw: a simple ClickFix-style attack is enough to hijack this assistant, which the outlet calls "extraordinarily privileged", able to act on the user's behalf. The three-part story sums up the moment: an agent that wins users fast, acts broadly and exposes as much as it helps.

The Canadian province filed a complaint in California against OpenAI, claiming the company could have used ChatGPT conversation logs to alert police and prevent the Tumbler Ridge shooting in February, Bloomberg reports. The attack had killed nine people, including the person behind the tragedy. The grievance fits in one sentence: failing to report threats made on the platform to law enforcement. The case opens new legal ground, that of an AI provider's liability for the danger signals passing through its service. It will weigh on how providers monitor, retain and pass on what users write to their models.

The two countries agreed to open an official dialogue on artificial intelligence, paired with a mutual notification mechanism in case of an incident threatening national security, Wired reports. The agreement emerged from eight hours of negotiations in New York between Treasury Secretary Scott Bessent, Trade Representative Jamieson Greer and Chinese Vice Premier He Lifeng, days before Xi Jinping's visit to Washington. Analysts cited by the South China Morning Post see it as a pragmatic step toward crisis prevention, tempered by deep divisions over chip access, model distillation and market dominance. The channel now exists; its effectiveness will depend on what each side agrees to signal to the other.

Alibaba presented what it describes as the highest-performing AI accelerator designed in China, meant to rival Nvidia and to support a massive data-centre rollout of up to 20 gigawatts of capacity by 2032, Bloomberg reports. For scale, 20 gigawatts is close to the output of some twenty nuclear reactors running at full power. The same day, at the opening of Alibaba's AI conference, Tencent launched its most advanced image-generation model to catch up with ByteDance and Alibaba. The two announcements, timed together, signal a Chinese industry pushing its own hardware and software stack while US semiconductor restrictions remain in force. The question of chip exports hangs over this week's Trump-Xi summit.

Chinese start-up Moonshot AI has begun supplying its flagship Kimi K3 model to Amazon Web Services, one of the world's largest cloud hosts, the South China Morning Post reports. The model is now available on Amazon Bedrock, AWS's tool for building generative AI applications, presented by Amazon as "a powerful new option for code and knowledge work". The stakes go beyond mere listing: it is a test case for whether open-weight models (which you can download and run yourself) can generate revenue through third-party platforms. A Chinese model distributed by a US cloud giant blurs the usual dividing line between the two ecosystems.