AI Watch, Monday 9 November 2026Fine-tuning a model erodes its safety guardrails

In an interview with Le Monde, the former governor of the Banque de France (France's central bank), who left his post in June, argues for repairing public finances through spending efficiency and three shared projects by 2029: artificial intelligence, energy and defence.

AI moves from the register of innovation to that of strategic European investment, on a par with the power grid and weapons. The argument targets both the funding of the social model and the continent's technological sovereignty.

Numerama details the model of CoreWeave, Nebius and IREN, these renters of computing power for AI (they buy chips, install them in data centres and rent out compute time). Their fragility comes from heavy debt taken on to buy processors that depreciate quickly, backed by rental contracts whose term poorly covers repayment.

When the market doubts demand for compute, these players fall first and drag down the rest of the sector.

The Journal du Net reports the proliferation of fake Aboriginal works produced by AI and sold online, from coasters to canvases. The counterfeiting deprives Australian First Nations artists of income and dilutes a culture whose motifs carry ritual meaning. The case extends the European debate on training image models from protected works without consent or payment.

The stock markets of Taiwan, South Korea and Japan fell heavily, with Taiwan extending its longest losing streak since 2022, in a global correction of AI-linked stocks. Taiwanese authorities announced measures to restore investor confidence, and the index rebounded immediately.

In Sydney, an official at Australia's central bank, Sarah Hunter, warned that the scale of AI investment will fuel demand and weigh on prices in the short term, as the bank raises rates to contain inflation. Stock-market doubt and the energy bill for building data centres are starting to echo each other.

The AI music generator faces two new lawsuits in the United States. The holders of the rights to Elvis Presley's work and the publisher of some major Eminem tracks accuse Suno of producing songs and lyrics that imitate the voice and words of these artists.

The Presley suit relies in particular on the Tennessee law that protects the right to one's image and voice (the "right of publicity"). Suno, which says it refuses to compose from an artist's name, was already facing lawsuits from record labels. The proceedings are multiplying around a single question: can a model trained on a catalogue reproduce its style without permission.

Ahead of an IPO, the Journal du Net reports that the shares held by Anthropic's employees could represent around $200 billion, paid out on condition that they stay with the company. Between taxes and holding periods before sale, this wealth will remain virtual at first.

The arrangement works as a retention lock in a talent war where an experienced researcher commands a premium.

According to Mark Gurman (Bloomberg), Apple's general manager, John Ternus, is pressing AI chief Craig Federighi to speed up Siri's roadmap and deliver an agent able to act on your behalf, not only answer (booking, sorting, chaining actions across your apps). The stated goal is to compete with OpenAI's agents and Meta's Muse.

Federighi has already delivered a working version of AI Siri on a tight schedule, but Apple acknowledges it is behind on this agentic ground. The file measures the difficulty, for a company committed to on-device privacy, of keeping pace with a market shifting toward autonomous assistants.

Yao Shunyu, chief AI scientist at Tencent and a former well-known OpenAI researcher, said on Saturday, in a rare public appearance, that models will soon be powerful enough to master biology, with serious benefits and risks. State-of-the-art systems are increasingly applied to medicine and biology, where they can speed up the discovery of treatments.

Yao warns against the use malicious actors could make of them to design pathogens. The signal comes from a Chinese executive trained in the United States, at a distance from the usual Western discourse on AI safety.

From 9 to 12 November, the Web Summit brings a large part of the global digital economy to Lisbon. La Tribune notes the place taken by China, which in a few years has become a key player thanks to a competitive cost-performance ratio, from robots to autonomous vehicles. The push is happening on European soil, just as the Union debates its own capacity to fund AI.