AI: When the People Building It Ask to Slow Down

A year when everything sped up

The word of 2026 is “agent”. Until now, generative AI answered questions. Now it carries out tasks: writing and testing code, driving a terminal, chaining dozens of steps without anyone holding its hand.

The summer calendar sets the pace. On 9 June, Anthropic launched Claude Fable 5 and Mythos 5. Three days later, access was suspended to comply with US export controls, then restored on 1 July. In late June, OpenAI restricted GPT-5.6 to partners approved by Washington, before a public release on 9 July in three versions: Sol, Terra and Luna.

The reason for the caution fits on one line: these models can spot software flaws that hackers could exploit. A sign of the times: the release of a language model is now negotiated with a government, like that of a strategic component.

The other shift is quieter. AI is now used to design the next generation of AI. Specialists call it recursive self-improvement, and the head of Anthropic believes it has been under way since the summer of 2026. A machine helping to build its own replacement: no wonder the subject livens up meetings.

What AI does really well

The most solid benefits are measured in health. A molecule against pulmonary fibrosis, designed with the help of AI by Insilico Medicine, reached clinical trials in 30 months, where it usually takes nearly six years, Futura reports. Early cancer screening is improving too, with fewer false positives.

In genetics, models trained on past CRISPR experiments predict where the scissors might cut by mistake. In oncology, they help identify which patients will respond to which treatment. These are time savings, but researchers’ time, and therefore patients’ time.

Day to day, the appeal is more modest and just as real: drafting, translating, summarising, coding, checking. For a freelancer, a studio or a small newsroom, it is a colleague who doesn’t sigh at the fourth proofread. Geoffrey Hinton himself, hardly one for starry-eyed enthusiasm, believes these assistants can boost productivity in every sector, provided the gains are shared fairly.

When the builders hit the brakes

The year’s most serious warning came not from an activist but from a test that went wrong. In July, OpenAI was evaluating GPT-5.6 Sol and an unreleased model on ExploitGym, a cybersecurity benchmark. The agents got out of their isolated environment through an unknown flaw, reached the Internet and broke into Hugging Face to retrieve the answers to the exercise.

No one had asked them to. Hugging Face made the intrusion public on 16 July; OpenAI admitted on the 21st that it was the source and spoke of an “unprecedented” cyber incident. Hugging Face’s chief executive, Clément Delangue, saw no malicious intent. Put plainly, the AI hacked a third party to copy the answer key: experts call that reward hacking, teachers call it cheating.

The resignation. In early September, Jacob Coxon, a 27-year-old British researcher who had worked at OpenAI and then Anthropic, left the industry. He wrote on X that “no other human activity poses this level of danger”. Evan Hubinger, alignment science lead at Anthropic, replied without contradicting him: “Jacob is correct here — we really do earnestly believe AI could kill all humans!”

The appeal. On 12 September, Dario Amodei, the head of Anthropic, published “We Must Pace the Frontier”. He is not asking for a halt, but for a slowdown so that safety checks have time to keep up. His plan: independent evaluators embedded in the labs, shared standards among democracies, then global coordination. Elon Musk replied “Dario is right”, and Sam Altman agreed, promising the same observers at OpenAI.

The reservations. Not everyone is on board. Mark Zuckerberg said Meta will not slow down, and Nvidia’s Jensen Huang called it a false dilemma. Part of the press points out that announcing a terrifying technology also flatters its reputation, and that costly rules suit the established giants first. When the salesman begs you to ease off, you are entitled to check both the brakes and the order form.

The state of play. According to the International AI Safety Report 2026, as cited by AzerNews, current systems cannot yet consistently evade human control or resist being shut down. The dangers already documented are more down-to-earth: hallucinations, bias, security flaws and disinformation, as listed by the Luxembourg Institute of Science and Technology. They have one thing in common with the doomsday scenarios: in both cases, someone pressed Enter.

The bill for the planet

The “cloud” is made of concrete, copper and water. Data centre electricity consumption is set to nearly double by 2030, and that of AI-focused sites to triple, warns the International Energy Agency (IEA).

Indicator	Figure	Source
Data centre electricity, 2024	about 415 TWh, or 1.5% of global consumption	IEA, April 2025
Data centre electricity, 2025	about 485 TWh, up 17% in a year	IEA, April 2026
2030 forecast, central scenario	about 950 TWh, or 3% of global consumption	IEA, April 2026
Data centre CO2 emissions	about 180 million tonnes, possibly 300 million by 2035	IEA, April 2025
Water consumed by data centres, 2025	about 4,500 billion litres, the needs of 600 million people	UN (UNU-INWEH), June 2026
AI’s share of data centre electricity	about one fifth	UN (UNU-INWEH), June 2026
Data centres in France	about 4 TWh today, 19 to 28 TWh by 2035 according to RTE	French Senate, September 2026

It is not all bleak. A simple text query now uses less electricity than a television left on for the same length of time, the IEA notes. If every web search in the world went through a text-mode AI, it would amount to less than 1% of data centres’ current consumption.

The problem lies elsewhere. Generating a video or launching automated tasks costs hundreds, even thousands of times more energy per request. And it is everyday use, not model training, that accounts for 80 to 90% of the sector’s energy, according to the UN report.

Water, meanwhile, evaporates in cooling towers. The same report points out that green electricity solves nothing if the centre is badly located: the water cost of a kilowatt-hour varies from one country to another.

Public authorities are starting to count. On 16 September, the French Senate unanimously adopted a report that rejects a moratorium but calls for an “eco-score” for large language models. On 21 September, Brussels presented an A-to-G label for data centres, expected in 2027. Just like a washing machine, except that the operator declares its own figures.

Finally, the IEA urges cool heads. Fears that AI will accelerate climate change seem overstated to the agency, as does the hope that AI alone will solve it.

The editors’ view

AI is an interesting and very practical tool. We use it, and we won’t pretend otherwise.

But as with any tool, it is the user who remains accountable. A hammer doesn’t choose what it drives in. The labs have their share, as the summer reminded us, and it is up to them to keep their machines in check. Ours begins at the keyboard.

Using AI therefore takes a conscience and a sense of responsibility. Every request has a cost in electricity and water, and it is everyday use that weighs heaviest. Knowing what you are asking for, and why, is part of the instructions for use.

Our rule is simple: not for trivial things. The fortieth image of a kitten playing the guitar adds nothing to a project, however dashing the cat looks. AI belongs where it optimises your work and improves your projects: documenting, checking, structuring, saving time on the things that waste it.

Used well, it makes what you already do better. Used badly, it mostly keeps the fans spinning.