AI Watch, Friday 9 October 2026AI chips away at young people's first jobs

By a decision dated 23 September 2026, the CNIL (France's data protection authority) closed the order it had issued on 22 January against France Travail (the national employment agency), without applying the penalty payment it had set, since the body had brought its data processing into compliance. The decision closes a case opened over the handling of jobseekers' data. For public bodies equipping their services with AI and automated processing, it is a reminder that documented compliance suspends the financial penalty.

The German search engine Ecosia, which funds reforestation, is abandoning Mistral five months after choosing it and switching to open-weight models (which you can download and run yourself), several of Chinese origin, citing quality, sovereignty and French nuclear power. Arthur Mensch, head of Mistral, responded publicly: "Sorry if nuclear power is a real dealbreaker." The episode stings for a "European green tech" firm that prefers Chinese models to a French champion, and it shows how fast a sovereignty argument turns around when the product disappoints.

A study covering more than 1,500 occupations concludes that young workers are already feeling the effect of generative AI on employment, with a decline in permanent hiring in the most exposed professions, the Journal du Net reports. The signal is quiet but clear: the adjustment does not come first through layoffs, it comes through the entry doors that stop being opened. This is the ground where training and apprenticeship policy is decided.

Anthropic is revising the usage policy of its Claude assistant for the first time in more than a year, TechCrunch and The Verge report. The new rules, in force on 12 November, explicitly ban election interference, disinformation campaigns, weapons development and certain surveillance uses. The most discussed change concerns another area: the company bans repeated, purposeless "abusive or cruel" behaviour towards Claude, while making clear that ordinary frustration and criticism remain allowed. Anthropic relies on an existing feature that lets the model end a conversation on its own. Behind the anecdote of a machine's "wellbeing", the company is mainly adjusting a usage contract that its professional customers will have to reread.

At its Gemini at Work event, Google presented an agent able to plan, carry out tasks and move between business applications, TechCrunch and The Verge report. The agent can delegate part of the work to sub-agents, call on several models, and above all receive an identity of its own within the company, with its own email address. Available in the Gemini Enterprise app, it is managed from a single interface where you talk to it and assign it missions. Giving an agent an email address and a place in the directory means treating it as a digital colleague, with the questions of access, accountability and logging that this raises. For IT departments, the challenge becomes framing what this identity is allowed to do before handing it out.

After OpenAI published 722 mathematics manuscripts produced by its AI, the Association for Human Mathematics, a collective founded in September 2026, is calling on its peers to stop all collaboration with the company, Numerama and TechCrunch report. Terence Tao, a Fields medallist, reposted the text on his blog, and one of the signatories sums up the move in a sentence: "This is not a demonstration of scholarship, but of force." TechCrunch, for its part, notes that the proofs released depart from the rules set by the mathematicians OpenAI had nonetheless consulted, and do not yet meet the discipline's standards. The episode raises a question that goes beyond mathematics: when a machine produces results faster than experts can verify them, value shifts from production to the capacity to check.

Firmus Grid, a data centre host backed by Nvidia, has abandoned what would have been one of the largest IPOs in Australian history, around $5 billion, for lack of international investors, Bloomberg reports. The withdrawal comes as markets grow anxious about valuations judged excessive across the whole AI chain. The same day, the shares of Chinese optical chip makers extended their fall, and the CSI 300 index dropped to its lowest level in more than a year. These signals do not say that demand for AI is weakening, they say that investors are no longer willing to fund any promise at any price. For an organisation that relies on an AI supplier, the supplier's financial strength becomes a selection criterion again.

The Chinese startup Manus, which specialises in AI agents, has closed a funding round of more than $500 million led by Boyu Capital and IDG Capital, TechCrunch and Nikkei Asia report. Existing shareholders, including Tencent, HSG (formerly Sequoia China) and ZhenFund, put in again. It is Manus's first raise since splitting from Meta, and it places the company among the Chinese players chasing the market for agent "harnesses", the systems that get a model to carry out complete tasks rather than simple answers. The deal confirms that money keeps flowing towards Chinese agents, at a time when American valuations are tightening.

Samsung expects a record operating profit of about $80 billion (107.4 trillion won) in the third quarter, a rise of 782.5% year on year, Siècle Digital and ZDNet report. Demand for memory chips bound for AI servers is driving these results, with the final figures expected on 29 October. This jump, which also benefits SK Hynix and feeds South Korean public spending, measures the scale of the hardware boom: the maker selling the shovels in the gold rush is booking profits that few AI labs can yet show.