The hard thing to say about artificial intelligence is that nobody is yet sure how to stop it doing harm. Andrew Bailey, the Governor of the Bank of England, said the risks around AI are real and increasingly significant. He also said the answer is not to begin with regulation.
Writing his first article for Substack, Bailey argued that regulating AI is not the right place to start. In his view, rigorous testing should come first, to find vulnerabilities, learn how models fail, and establish the points at which society can credibly intervene. Only then does a regulatory framework have something solid to stand on. Over time, he allowed, a more formal regime may emerge from that work.
He was careful not to sound like a prohibitionist. AI development should not be halted or banned, he wrote, because the benefits are immense. His worry was narrower and more technical: frontier AI systems, the highly advanced self-learning kind, risk becoming a closed loop in which the model progressively governs itself, unless an effective way to intervene is put in place first. Testing will produce failures, models will behave unexpectedly, and that is not evidence that testing has failed, he said. It is evidence of why testing is necessary.
Two things give the intervention weight beyond one banker’s blog. He pointed to the UK’s AI Security Institute as important work that should help form a set of standards, while saying the pace of progress must accelerate, and he made the remarks in a week when OpenAI chose not to release its newest model over safety concerns, and President Trump hosted the heads of OpenAI, Anthropic, Nvidia, SpaceX, Meta and Google on the same theme.
His closing note was the most human line of the piece: we should proceed with a degree of humility. Coming from a central banker who spent the pandemic defending the plumbing of the financial system, the advice deserves a wider hearing than a Substack post usually gets.
The article was reported by the BBC, and the full text is published by the Bank of England.