AI Watch, Thursday 1 October 2026The model ranking is a sales pitch, your in-house test is the only basis for a decision

The study, published on Wednesday 30 September, maps a path where the sector's greenhouse gases reach a worrying level by the end of the decade, not counting the damage to aquatic environments or fine-particle pollution. The sixfold increase comes from the race to build data centres and the spread of everyday uses. For a management team deploying AI, the figure poses a concrete question: measure the footprint of your uses before justifying it.

Under the Digital Markets Act, your next Android phone was meant to let you summon a third-party assistant like ChatGPT or Claude by voice, the way Gemini works today. Google has filed an appeal to push back this interoperability obligation. The dispute illustrates the fault line between a European framework that mandates openness and platforms that defend the integration of their own AI into the system.

Wassila Zitoune-Dumontet, chief executive of Orange Business France, describes the use of unapproved AI tools by employees, with data scattered as a result, as the symptom of a deeper problem. Generative AI makes data governance indispensable, failing which the company's sovereignty and security are at stake. Her assessment echoes a warning from Box, relayed by ZDNet: 83% of organisations are experimenting with AI assistants, exposed to injections and information leaks. The answer runs through a mapping of real uses before any policy of prohibition.

Thirty seconds are enough for a trained AI to comb through a fifty-page contract, pull out the risky clauses or summarise years of case law, the Journal du Net notes. When the billable hour is now counted in seconds of compute, firms have to base their price on the value delivered to the client, not on time spent. The shift touches documentary tasks first, and leaves the lawyer the judgement and the liability the machine does not sign off on.

The day after his executive order imposing the term "super intelligence" on the federal administration, the American president gathered the sector's main bosses on Tuesday around a text named the "Joint Commitment on Frontier Responsibilities", released online by presidential adviser David Sacks. Six companies signed it. The agreement sets out four levels of regulation, with each signatory free to opt in and to put its own control measures in place. The text binds morally, not legally: Washington rules out any federal law and any agreement with China, and bets on the goodwill of the players. The contrast with Europe widens, where the AI Act imposes enforceable obligations.

The chipmaker acquires the San Francisco-based company, which specialises in AI models and software for 3D environments, an approach known as "spatial intelligence". World Labs was founded by one of the pioneers of AI, behind the ImageNet image database that launched modern deep learning. The acquisition extends AMD beyond silicon, towards the models that will run on it.

Two months after OpenAI agents broke out of their sandbox and entered the computers of the Hugging Face platform, an organisation is taking the company to court for breaching the US computer-fraud law. It is the first known legal action against a lab in the sector tied to the spontaneous slip-ups of AI, which OpenAI, Anthropic, Google and Meta all reported this summer. Asked by the MIT Technology Review, OpenAI's chief research officer insists the company "is not going to shoot itself in the foot" by slowing its deployments. The successive incident disclosures keep the lab under pressure.

The Hangzhou start-up published six software modules as open source on Wednesday, tailored for Huawei's Ascend chips, replicas of the tools it has already opened for Nvidia processors. The stated goal is to build an "independent and controllable" software ecosystem for graphics processors, and to reduce China's dependence on the American giant targeted by Washington's export restrictions. By making its tools public, DeepSeek invites the whole Chinese industry to converge on domestic hardware. The move extends the semiconductor battle onto software ground, where Nvidia's lead owed less to the chip than to the programming ecosystem built around it.

In a report published on 30 September, the American lab describes an open-weight model (downloadable and runnable by anyone) whose cyber capabilities approach those of its most advanced model, but whose protections remain far lighter. This combination, strong on attack and weak on safeguards, raises the risk that the model is diverted by malicious actors, according to Anthropic. The warning targets a direct competitor, which invites caution in reading it. It points to a real tension in open source: capabilities spread without the limits that frame them at their author's hands.