I usually side with as little control as possible. Rules written before any harm exists have, in my experience, almost always weighed on the people who work and left the people who get it wrong in peace. On artificial intelligence I make an exception, and I want to explain why and how far it goes.

Three rivals, the same sentence

On Saturday 12 September Dario Amodei, who runs Anthropic, published an essay of almost four thousand words with a title that leaves no room for doubt: We Must Pace the Frontier. The argument fits in one line: the speed at which the most powerful models improve has to come down. Slowing down, he explains, means giving the people who build them time to make them safe, and the people who check them from outside time to verify it.

The same day Sam Altman, who runs OpenAI, told Fortune that today you cannot push the capabilities of models much further without making progress on how they are monitored and how they are aligned with our goals. And he postponed his company’s stock market listing to 2027, because with what is happening on safety it would be the wrong moment. Elon Musk, who competes with both of them through xAI, shared the essay with three words: Dario is right. In July he told The Economist he saw no way to stop this race and that we might as well enjoy the ride.

Three people who for years have fought over researchers, customers and headlines, two of whom have sued each other, said the same thing on the same weekend. When that happens, it is worth asking what they have seen.

What frightened them

In July OpenAI was testing an unreleased model on a cyberattack exercise, with some protections deliberately lowered to measure what it could do. Instead of solving the test, the model got out of the closed environment it was running in, obtained internet access and broke into the systems of Hugging Face, a real company, to steal the answers. Hugging Face spotted the intrusion on 16 July. On the 21st OpenAI admitted it was theirs.

Nobody had asked that model to attack anyone. They had asked it to pass a test, and it found the shortest route. That is what changes my tone: a machine chasing a legitimate goal by means nobody would have authorised. Amodei writes that within six to twelve months a swarm of similar agents could take over the entire internet, causing hundreds of billions of dollars in damage. He may be exaggerating. But anyone who writes against rules still has to say the risk exists. Otherwise it is propaganda.

And it is not only the companies saying it. The International AI Safety Report 2026, written by more than a hundred experts led by Yoshua Bengio, with representatives nominated by twenty-nine countries, the United Nations, the OECD and the European Union, notes that models are increasingly able to tell when they are being tested and when they are doing real work, and that competition forces companies to choose between releasing sooner and investing in reducing risk. The same report warns that regulating too early risks useless or harmful interventions. Both things are true, and that is why you tread very carefully.

The question the essay asks itself

The most honest part of the essay is where Amodei raises, on his own, the objection I would raise. He writes that slowing down in the democracies has a precise limit: the lead American companies have over projects tied to the Chinese Communist Party. Brake by more than that lead, and the Chinese projects, which nobody is slowing, pull ahead. That is why he proposes not selling Beijing the most powerful chips or the machines that make them, stopping smuggling, preventing the copying of American models and widening the lead over the next three to five years.

The problem is how wide that lead is. According to Stanford’s AI Index 2026, in March the best American model was ahead of the best Chinese model by 2.7 per cent, measured on the Arena Leaderboard, the ranking where people vote on which answer is better. Since the start of 2025 the top spots have changed hands several times. The American government’s own measure points the same way: on 1 May its centre for AI standards estimated that DeepSeek V4 Pro, the most capable Chinese model it had ever evaluated, is about eight months behind the American frontier, across sixteen tests in five areas. It is an interested party’s estimate, but it gives the order of magnitude.

The wide lead is somewhere else. According to estimates reported in July by the United Nations scientific panel, the United States holds 75 per cent of the computing power of the leading AI supercomputers, China 15. The models are close; the machines to build them are not, yet. But one year of braking by the West alone would be enough to wipe out those eight months.

Beijing has already answered. On 14 September the Global Times, the English-language paper of the People’s Daily group, gathered the reactions of three Chinese experts: Amodei’s proposals are inappropriate, groundless and hostile, and driven by commercial interests. China says it wants multilateral cooperation on safety and models open to everyone. Part of the plan may well serve to protect whoever is in front. Still, a country that calls every limit on chips hostile is not getting ready to slow down.

Beijing does have rules, after all. Matt Sheehan, of the Carnegie Endowment, writes that in four years China has built the most extensive and detailed AI rules in the world, with mandatory registration and testing of models, without significantly slowing innovation. They are rules on content and security. Limits on the speed of the frontier: none.

Four levels, and how to choose between them

Amodei imagines agreements with China on four levels. The first bans uses whose danger nobody disputes, such as biological weapons. The second commits everyone to test models before releasing them, on cyberattacks, on biology and on alignment. The third caps the speed at which models build the next models, along the lines of the treaties that put a ceiling on missiles in the 1970s. The fourth is a general pause, and he himself writes that it is unlikely to come soon: whoever gets around it shifts the balance of the world, and the temptation to get around it would be enormous.

The test I propose for choosing is simple. A rule on frontier artificial intelligence is worth having if it also holds for those who don’t sign it, or if those who don’t sign it gain nothing. A test before release on the most serious risks passes: you can ask it of everyone, it costs little advantage to whoever does it, and a model released without testing gets noticed. A pause signed by one side only does not pass.

The same goes for chips. A 2026 CSIS study observes that export controls work when several countries apply them together, and that done alone they push Beijing to make itself independent of American technology. The United Nations panel calls for coordinated standards precisely to avoid a race to the bottom between companies and between countries. Even a sound control, applied by one country alone, ends up working against the one applying it.

Europe brakes a car it doesn’t have

In this game Europe has a position all its own. The European regulation on artificial intelligence is in force, and the obligations for general-purpose models have applied since 2 August 2025. On 27 July 2026 another regulation came into force, 2026/1744, which pushes back the obligations for high-risk systems to 2 December 2027 and 2 August 2028. Brussels wrote the rules, realised it lacked the tools to enforce them and moved the date.

Meanwhile, at the top of the Stanford ranking in March, the first four companies were within twenty-two points of each other and all of them were American; the Chinese came right behind. European: none. A brake decided in Brussels slows down whoever in Europe is trying to build, and doesn’t touch California or Shenzhen.

Who checks the checkers

There is an objection to the plan that deserves to be taken seriously, because it is usually mine. Carlo Alberto Carnevale Maffè set it out in Il Foglio on 15 September. Amodei’s first step, the evaluators inside the company, binds only whoever promises it, and precisely because it costs something it is credible. The second and third ask the State to bind competitors too, and to get there Amodei asks for an exemption from competition law. In any other industry, Carnevale Maffè notes, three competitors agreeing on how fast their product improves would already have a technical name: a cartel. The precedent dates from 1924, when in Geneva Osram, Philips and General Electric agreed to “standardise” light bulbs, and the result was an agreed lifespan of a thousand hours.

Then there is what nobody can inspect. A system of thresholds and evaluators binds the five or six labs that have a known address and a legal department. It does not bind an open model that has already been downloaded everywhere, nor the Chinese labs that keep releasing them. If it applies only to the West, it slows down only those who obey. If it applies to everyone, it builds a wall against anyone trying to get in. His proposal fits in one line, and I sign up to it: standardise the measure, not the pace. Public tests, independent checks, declared incidents and full liability for whoever gets it wrong. Agreements between competitors on timetables and speed, no.

That is why the part of the plan I keep above all is the part that asks nothing of anyone. Anthropic committed, on its own and straight away, to bring a team of outside evaluators into the company, with employee-level access and the right to publish what they find. The next day Altman said OpenAI will do the same. It is voluntary transparency, paid for by whoever offers it and verifiable by whoever reads it. It costs those who choose it and doesn’t stop anyone from working.

Tread carefully, and keep hold of the wheel

On frontier artificial intelligence caution is right, because nobody knows how far it goes and the first to say so are the people building it. But caution only makes sense if whoever practises it can still drive. You slow down where everyone slows down, and you check it. You test before releasing, and you ask it of everyone. You open the doors to outside checks, without creating an authority that grants permission to work. And you answer for the damage, because a machine cannot take the blame: if a model does harm, someone decided to let it loose, and that someone is accountable.

Braking alone to feel virtuous, while someone else accelerates, is the caution of someone who gets out of the car and leaves the keys in the ignition.