The headline numbers
Verasight, a non-partisan survey research firm, released a poll on Monday of 1,500 US adults aged 18 and over, fielded from 16 to 21 September with a margin of error of 2.6 percentage points.
Sixty-three per cent said AI development should slow down. Five per cent said it should speed up. The gap between those two numbers is the finding; the rest of the poll explains what sits behind it.
What people say they would choose
Asked to pick between two companies, 73% said they would choose the one that develops more slowly and spends more time on safety testing, even if its technology ends up somewhat less advanced.
“Most Americans do not use the most advanced features of AI, so 73% choose the company that develops more slowly and spends more time on safety testing,” said Ben Leff, Verasight’s chief executive. That is an argument about revealed preference rather than fear: if the frontier capability is not something you use, trading it for caution costs you nothing.
Data centres, and who should hold the brake
Seventy-eight per cent said they would oppose a large data centre within ten miles of their home. That is the number with the most immediate political consequence, because data centre siting is decided locally and in the open, while AI policy is decided federally and slowly.
On who should be responsible for the pace, 55% said the AI companies themselves rather than the government — an answer that sits oddly beside the appetite for slowing down, and suggests a public that wants restraint without trusting regulators to impose it.
Fifty per cent rated AI and nuclear weapons as equal long-term risks. Forty-five per cent said they worry more about rapid AI development causing harm than about the United States losing a race with China, which inverts the framing that dominates Washington’s AI debate.
What to watch
The poll lands in a week when the same question is being argued in other forums: more than seventy YouTube creators have signed a public call for governments to slow AI down, and former lab researchers testified in New York on Monday about what they saw inside frontier labs. Public opinion is not the binding constraint on any of it, but it is the input campaigns read.
A poll fielded in mid-September and released in October is a snapshot, not a trend, and Verasight’s numbers sit alongside other recent surveys showing the same direction. The question that matters for the midterms is whether data centre opposition converts into local ballot outcomes, because that is where a diffuse opinion becomes a specific decision.
