The labs building the world's most powerful AI have agreed to let outsiders inspect their work. What that same AI does inside your company is almost never inspected by anyone. And what protects you today are voluntary promises.
This is what has happened since Saturday. And why, for a European company, the risk lies in what your provider is under no obligation to govern.
72 hours
On Saturday 12 September, Dario Amodei, chief executive of Anthropic, published We Must Pace the Frontier. He asks for a slower rate of improvement in the most advanced models, so that safety does not fall behind. He is not asking for a stop: "pacing does not mean halting". And he commits Anthropic, with immediate effect, to opening its doors to external evaluators with employee-level access.
Within hours he had the backing of Sam Altman, who promised the same for OpenAI; Elon Musk, with a terse "Dario is right"; and Demis Hassabis of Google DeepMind, for whom the essay "points towards the right path forward". Altman went further: he told Fortune that taking OpenAI public this year would be "ill-advised", given the state of safety.
On Sunday, in Doonbeg, Ireland, Donald Trump replied that "whoever wins AI, wins". He allowed that guardrails could be put in place. He announced none. On Monday, Beijing dismissed the labs' appeal as "fearmongering". And SoftBank, one of OpenAI's major investors, fell 13% on the Tokyo stock exchange.
What the AI slowdown argument is really about
Forget braking versus accelerating. The question is who checks.
The pace is the least novel part of the essay. What is new is the willingness to let someone from outside look in while the thing is being built. More than a declaration, it is a concrete commitment.
The trouble is that, for now, all of it is voluntary. There is no federal law in the United States, and none is expected before the November midterms. What a lab promises on a Saturday it can withdraw on any other day.
Who isn't buying it
Both camps reach for China. Amodei uses it to argue against selling China advanced chips; the White House, as a reason to slow nothing down. The numbers explain the tension. According to Stanford's AI Index 2026, private AI investment in the United States was 23 times China's in 2025, and yet in March the best American model led the Chinese ones by only 2.7%.
Nor does everyone read it as prudence. Stuart Russell, of Berkeley, thinks the plan is "completely backwards": set the safety requirements first, and only then let capabilities advance. David Sacks, a White House adviser, reminded the labs of their product-liability exposure. The vice-president, JD Vance, called it "a bit of a Trojan horse".
SoftBank suggests that caution carries an immediate price, although its fall coincided with OpenAI ruling out a listing this year.
The asymmetry for European companies
If you operate in the EU, the game is different. Since 2 August, the AI Act has required that people be told when they are interacting with an AI, and that what it generates be labelled. The obligations for high-risk systems arrive in December 2027.
Your AI provider, most likely American, is not obliged to govern what you are going to have to explain. To your board, your customers or your regulator.
And that provider can become a political problem overnight. In February, the Pentagon designated Anthropic a "supply chain risk" after the company refused to remove the safeguards that prevent its model being used for autonomous weapons and mass surveillance. In August, a federal judge ruled that unlawful. For six months, relying on that provider carried a political dimension that no contract had written down.
Three questions for your board this week
Which third-party models does the company already depend on, and what would happen if one of them changed its terms tomorrow, slowed down, sped up or turned into political news?
What would we have to explain, and to whom, if one of those systems got it wrong?
And who, in-house, checks what those systems do, against which indicator and how often? Not a policy approved in committee: a name and a measure.
My reading
If the people who know these systems best accept outside scrutiny while they build them, no company that uses them should settle for less. It is the only way to explain what you are doing when someone asks.
Measurement doesn't manufacture judgement. It makes it defensible.