Deputy governor Sarah Breeden says AI traders are starting to herd in ways that could break markets in seconds, and the Bank of England and the Bank for International Settlements have modeled the risk without yet building the legal or technical lever to stop it.
The Bank of England can describe, with some specificity, how an artificial-intelligence-driven market crash would unfold. It cannot describe how it would stop one.
That is the gap Deputy governor Sarah Breeden named this week at the European Central Bank's annual central banking forum in Sintra, Portugal. Speaking on a panel about AI and financial stability, she argued that AI-driven traders are already starting to behave in correlated ways, and that the central banks watching them have the diagnostic apparatus to spot the risk but not the legal or technical authority to intervene before it cascades.
The mechanism she is worried about is herding. When AI models trained on similar data and rewarded for similar objectives encounter the same market signal, they tend to act on it at the same moment, in the same direction, and at machine speed. That compresses the liquidity and human discretion that traditionally absorb a shock into something too small to matter.
The Bank of England, the Bundesbank, and the Bank for International Settlements have been jointly simulating exactly this failure mode. BIS Working Paper 1194, on how AI is transforming finance, and Working Paper 1291, on using AI to monitor financial markets, frame the institutional analytical backdrop: synchronized AI execution can drain order books, and the supervision tools available today were not built to see it happen in real time.
Breeden's intervention surfaced in the financial press as a "kill switch" proposal, a mechanism to halt market-wide trading when faulty AI models start amplifying each other's mistakes. The Telegraph and Yahoo Finance both led with the framing. The speech itself is closer to a thought experiment than a draft rule: Breeden described exploratory mitigants and joint research with the Bundesbank and BIS, not a formal BoE proposal. The interesting move is that she put the missing instrument on the public record.
The gap she is naming is structural, and it is not unique to the United Kingdom. A central bank can model herding behaviour with reasonable confidence. It cannot, today, identify which specific AI agents are about to coordinate, why they are about to coordinate, or at what precise moment coordinated action turns a routine repricing into a market-wide event. The visibility supervisors have into hedge fund, bank, and proprietary trading books was designed for human decisions logged by humans. AI agents making microsecond decisions, sometimes inside opaque models, fall outside that architecture.
That is why the kill-switch framing matters even before any formal policy exists. It names a missing instrument. Existing market-wide circuit breakers, the kind that pause trading after a sharp move, were designed for slow, visible price dislocations. An AI herding event would compress the warning time to milliseconds and would not necessarily produce the single-name price gap that existing rules are calibrated to catch.
What to watch next: any joint BoE-Bundesbank-BIS follow-up paper that defines what a herding threshold actually looks like in production data; any move by the Financial Stability Board or Basel Committee to scope AI-agent supervision into existing market infrastructure rules; and whether exchanges themselves build in AI-aware circuit breakers before regulators mandate them.
The honest summary is that the Bank of England has done the harder half of the work. It has a working theory of what an AI-driven crash would look like. It does not yet have the lever.