TL;DR

Andrew Bailey has argued that society must keep the power to step in and set limits on frontier AI. He wants testing first and regulation later. Separately, on Wednesday 30 September, the Bank’s Financial Policy Committee (FPC) published a record warning that AI-related borrowing is spreading exposure across capital markets.

Testing before rulebooks

Writing in the Bank’s Insight series, the governor asks whether society should retain the ability to set and revise boundaries for these systems. “To my mind, the answer is unequivocally yes,” he writes. His concern is recursive learning: models that refine themselves on their own outputs risk becoming a closed loop that, without a way in, “progressively governs itself”.

Bailey is explicit that a regulatory clampdown is not where he would begin. He wants rigorous testing of models both before and after release, so authorities can find “credible points of intervention”. He credits the AI Security Institute with a strong start but says the pace has to quicken.

For a central banker, the link is cyber risk. Payment networks, market infrastructure and banks can no longer be assessed apart from AI, he argues, and lessons from testing could become shared standards for the financial system, “and perhaps more broadly across the economy”.

The debt behind the boom

The FPC record adds the balance-sheet side. It cites a Morgan Stanley estimate that worldwide AI-related debt issuance reached roughly £339bn ($450bn) by early September, over twice the 2025 total. In sterling markets, AI hyperscalers accounted for 47% of corporate bond issuance so far in 2026, though that sterling borrowing is still dwarfed by what they raise in American and eurozone markets. The committee also points to test-environment incidents in the third quarter in which autonomous models took unexpected actions, and urges firms to prepare for the resulting cyber and operational risks.

A shift in emphasis

In July, Bailey argued that no country could seal itself off from AI risks, and he called for coordinated international testing. This piece turns to the Bank’s own remit: central banks, he writes, “cannot stand aside”. It also sits awkwardly beside AI minister Kanishka Narayan’s argument this week that building British capability may do more than regulation.

Looking forward

Bailey stops short of proposing a mechanism, so the practical question is who writes the standards and when. For UK banks, payment firms and other financial institutions using AI, the direction is clear enough: Bailey wants the models they rely on rigorously tested, and the FPC is already pressing firms to get ready for the cyber and operational risks AI brings.