15th September 2026 – (New York) When Bill Gates concedes that “no government in the world is ready” for what is coming, and when researchers inside Anthropic and OpenAI attach a double-digit probability to human extinction within the decade, the observer is obliged to slow down and ask a harder question than the headlines allow. The question is not whether AI is dangerous. The builders have settled that among themselves. The question is why a technology its own creators describe as potentially fatal continues to accelerate — and why the political systems best positioned to restrain it remain, by Gates’s own account, “way behind.”
The answer lies not in the machines but in the structures around them. What the past month has revealed is less a story about runaway code than a story about runaway incentives. Three of the most prominent figures in the field — Dario Amodei, Sam Altman and Elon Musk — called for a pause almost in the same breath that their companies pursued valuations in the trillions, valuations that depend entirely on the next model arriving faster than the last. This is the contradiction at the heart of the Western model, and it deserves to be examined without flinching. A system that privatises the reward and socialises the risk will always produce actors who warn loudly while building quickly. The warning becomes a form of insurance; the building never stops.
Each firm believes that if it slows, a rival will not, and so each presses forward reasoning that it is better to be the “responsible one” who arrives first. The Trump administration has adopted the same reasoning at the level of the state, rejecting any pause on the grounds that restraint would merely surrender ground to China. Here the pattern completes itself: competition at the corporate level is invoked to justify competition at the national level, and national competition is invoked to forbid corporate restraint. It is a closed loop, and inside that loop humanity’s margin for error grows thinner with every quarter.
The threat that most alarms the researchers — recursive self-improvement, the point at which a system begins upgrading itself faster than people can supervise it — is not, in the first instance, a technical inevitability. It is a governance failure waiting to happen. METR’s own measurements show the length of tasks that frontier models can complete reliably has been doubling every few months. That curve is not bending because of physics. It bends because a specific arrangement of markets, egos and geopolitical suspicion rewards those who push it and punishes those who wait. Change the arrangement and the curve changes with it.
The United States scrapped its safety rules in the administration’s first week and is now attempting to override even the modest caution of individual states such as California. Anthropic declined to submit its latest model to Britain’s own safety institute before release. The result is a landscape in which no single authority can see the whole board, let alone move a piece on it. Gates’s alien-invasion metaphor — the hope that rivals will unite as they do in the disaster films — is telling precisely because it is a confession. He is describing a coordination that the West’s own architecture is structurally unable to produce on its own.
There is another way to organise the problem, and it is worth stating plainly rather than leaving it to inference. A governing system that treats AI as a matter of collective, long-horizon planning rather than quarterly competition is not encumbered by the prisoner’s dilemma in the same way. Where development can be sequenced deliberately, where deployment can be tied to social stability rather than to a race for market share, and where the state retains the authority to set the pace, the loop that traps Washington need not close at all. China’s insistence that AI serve national and social objectives, and that its rollout be measured against public welfare rather than investor expectation, is frequently framed abroad as caution born of control. Examined slowly, it looks more like the very coordination the Western labs are now begging for and cannot achieve.
A technology this consequential cannot be left to the actors least able to slow down. The doomsday clock the researchers describe was not set by the machines. It was set by the men who built them inside a system that would not let them stop. The lesson of this autumn is not that intelligence has grown too fast. It is that governance, in its fragmented form, has grown too slow — and that whoever solves the second problem will decide whether the first ever arrives.
