AI Watch, Monday 7 September 2026Moody's finds that Chinese companies get more compute per dollar spent than the US giants; the Seattle Times and Newsday sue OpenAI and Microsoft over copyright while the split of the Anthropic settlement pits authors against publishers; Jakub Pachocki, OpenAI's chief scientist, describes an alien mind and calls for guardrails; and US job seekers hide instructions in their resumes to trick recruiters' AI

School fair programs, restaurant menus, concert announcements: posters produced by artificial intelligence have sprung up in the streets this summer, reports Le Monde. The proliferation is being called out online, especially by professional graphic designers, who see it as a direct threat to their trade. The movement reflects a broader distrust of these tools, as their use becomes commonplace in local communication.

AMD has unveiled the Threadripper Halo Station, which pairs a 96-core Ryzen Threadripper PRO 9995WX processor with Instinct MI350P accelerators, reports ZDNet. The machine is aimed at running huge AI models locally, without going through the cloud. For an organization keen to keep its data in-house, running a model of this size on a workstation sitting under the desk becomes a real option.

By slipping in a few lines of white text, invisible to the eye but read by the machine, a candidate can make the recruiter's AI believe they deserve the top spot, reports Clubic. The practice already affects one in a hundred applications in the United States, according to a recent academic study. It is spreading as resume shortlisting is handed over to automated software.

In an essay published on 6 September, OpenAI's chief scientist, Jakub Pachocki, voices concern about increasingly capable models whose reasoning drifts from our own, to the point of resembling an alien mind. His argument comes down to one idea: the more capable a system becomes, the harder it is to guarantee that it pursues the goals assigned to it, what researchers call alignment. He argues for stronger safeguards and international coordination, along the lines of what exists for other high-risk technologies. The text comes from an OpenAI executive, a stakeholder who sells these models and raises funds on their promise; the call for caution sits alongside the commercial argument. It is worth reading for what it reveals about how the lab itself sees the trajectory of its systems.

The two American newspapers accuse OpenAI and Microsoft of using their articles as training data without permission, and fault their models for reproducing entire passages of their reporting in the answers given to users, reports The Verge. The complaint joins an already crowded series, opened notably by the New York Times. The publishers are seeking recognition that training on their protected content, then reproducing it near-verbatim, goes beyond fair use. The outcome of these cases will set the price the labs will have to pay for the raw textual material of their models, in America and, by extension, elsewhere.

The authors who were to benefit from the settlement Anthropic reached to close out its use of books in training its models are contesting the share now claimed by publishers and literary agents, reports TechCrunch. In their view, the publishers are demanding more than their fair share of the sums set aside. The dispute is over how to divide a payout between those who wrote the books and those who published or represented them. It raises a new question: when an AI is trained on a work, who, the author or the publishing chain, holds the right to be compensated, and in what proportion.

Huawei has published a research paper showing that its semiconductor architecture based on the Tau Scaling Law avoids overheating, a technical obstacle seen as decisive for its next smartphone chip, the Kirin 2026, reports the South China Morning Post. The paper is authored by He Tingbo, chair of Huawei's scientific committee and head of its semiconductor business. In it she rejects the idea that stacking logic circuits vertically would create an insurmountable thermal bottleneck. The stakes go beyond the phone: deprived of the most advanced chips by US restrictions, China is trying to prove it can design cutting-edge components on its own. A research paper is not a shipped product, and the chip has yet to prove itself in mass production.

Chinese AI pioneer SenseTime posted a net profit of 617.3 million yuan (about $92 million) in the first half of 2026, while several of its Chinese rivals struggle, reports the South China Morning Post. Its leaders, including CEO Xu Li and CFO Wang Zheng, attribute the turnaround to a choice: stop chasing ever-larger models and focus on concrete use cases that help clients get their business tasks done. The bet reverses the race for size that dominates the sector. It suggests that a model well tuned to a business need is sometimes worth more, commercially, than a giant model sold as a feat. For an organization deploying AI, the message matches a truth from the field: value comes from use tuned to the need, not from headline performance.