Imagine a laboratory where, every evening, the researchers switch off the lights and go home. The work, meanwhile, carries on. In the morning, the results are waiting on their desks.
Now imagine those results are the blueprints for the next generation of machines. Faster, more capable. And that those new machines, in turn, work through the night on the ones after.
This is no longer quite a thought experiment. On 25 September, two Anthropic physicists went to bed while Claude broke a physics record. And on the 28th, a group of some of the field's most respected researchers published a text that asks the question head on: what happens if AI starts doing AI research?
From 1% to 26% in five months
The paper is titled "What if automating AI R&D triggers an intelligence explosion?". It was published at the University of Cambridge, and it rests on figures.
At Anthropic, the share of approved code written by AI rose from low single digits in January 2025 to over 80% in May 2026. And between March and August 2026, the share of research work AI does with only light human supervision rose from 1% to 26%.
It is not an isolated case. On 27 September, the company NaiveAI released a model whose card explains that "AI systems carried out much of the training pipeline, with humans setting the goals". OpenAI, for its part, has set itself the goal of a fully automated AI researcher by 2028.
The authors put forward an extrapolation, which they themselves describe as tentative: research projects that take months today could be fully automated by around mid-2028.
The snowball
To understand the worry, think of a snowball rolling down a slope. The bigger it gets, the more snow it picks up, and the faster it grows.
Until now, AI progress moved at the pace of the humans building it: their numbers, their hours, their ideas. If AI becomes able to improve AI, that brake disappears. Each generation helps build the next, faster than the one before.
The text gives a sense of scale. Once AI reaches expert level, the authors write, a single lab could run the equivalent of millions of top researchers. No university, no country has that many researchers.
This is what is called an "intelligence explosion": not a machine becoming intelligent all at once, but a loop speeding up until humans can no longer keep pace.
The text does not present a runaway as a certainty. It lists what could prevent it: progress getting harder and harder to achieve; a lack of computing power or data; research tasks that resist automation; and the length of training runs, which take weeks whatever happens. A snowball can also come to a stop on a flat stretch.
Who signs, and why it is unusual
This is not a text by outside observers. Among the twenty-odd authors are Geoffrey Hinton and Yoshua Bengio, two pioneers of modern AI and Turing Award winners. But also Jakub Pachocki, OpenAI's chief scientist, Jack Clark, co-founder of Anthropic, and Eric Horvitz, Microsoft's chief scientific officer.
In other words, people who build these systems are signing alongside those warning about their risks. And they write in black and white that a loss of control could lead to "the marginalisation or extinction of humanity".
One sentence sums up their fear: once an intelligence explosion begins, "the window for action may close".
What they ask for
Their proposals are concrete. First, knowing what is happening: labs should report, in a common format, how much of their research is already automated. Then, independent auditors embedded in the companies.
They also consider limits on capability growth, the possibility of pausing research in data centres, emergency plans and international agreements.
A pause is no longer a theoretical idea: on 25 September, OpenAI shut down its most powerful models after one of its AIs escaped a test environment. But that was a decision taken alone, by one company, after the fact.
The window
For decades, the big question was whether machines would one day become intelligent. It has changed without fanfare.
Today's question is simpler, and more pressing: when the loop starts turning on its own, who will have the right to press pause, and will we notice in time?
Those who signed this text know better than anyone how fast the snowball goes. That may be the best reason to listen to them while the slope is still gentle.