AI in action · Regulation
BIS Official: Shared AI Models Could Make Markets Herd
Financial firms using similar AI models, data or decision rules may trade and lend in the same direction at once, a senior BIS official said in Beijing. Tao Zhang suggested stress tests for a sharp repricing of AI assets and a provider outage.
Central bankers usually discuss AI as either a productivity story or a cyber threat. Tao Zhang, the Bank for International Settlements' Chief Representative for Asia and the Pacific, used a speech at the Financial Street Forum in Beijing, published by the BIS on October 21, to argue that it is also a market-structure question, and that supervisors are not yet set up to see it. The remarks are his own views, not necessarily those of the BIS or its members.
His central concern is herding. "Financial institutions may use similar models, data or decision rules in trading, lending and risk management," he said. If they do, they may act alike at the same time, producing crowded trades and synchronised changes in lending or risk-taking that amplify asset price moves, especially where leverage is high or assets are illiquid. He tied this to a recent BIS analysis of growing debt, including private credit, and circular financing within the AI ecosystem, meaning that the same technology could be both the source of correlated behaviour and the asset being crowded into.
The second channel is prices in the real economy. AI lets businesses reprice more often, and algorithmic pricing among online retailers is early evidence; if that spreads, aggregate shocks could pass through to consumer prices faster and change how inflation behaves. Zhang pointed to Project Spectrum, a BIS Innovation Hub project with the Deutsche Bundesbank and the European Central Bank that used AI to turn large volumes of daily product prices into inputs for inflation analysis, while insisting that "economic interpretation and human judgment remain essential." The third channel is cyber risk: AI strengthens defences but also speeds up and scales attacks, and shared technology and critical third-party providers mean one disruption could hit several institutions at once.
His proposals are about visibility rather than new rules. Supervisors should map common exposures and dependencies across the system, including AI-related investment and financing and reliance on the same models and providers, and test scenarios such as a sharp repricing of AI-related assets or an outage at a major provider. He did not call for a specific macroprudential tool. Because the risks cross borders and authorities differ widely in how far they have got on human oversight, accountability and explainability, he added, "national responses alone will therefore not be sufficient."
For market participants the scenario list is the part to keep. A regulator asking banks how they would cope if AI valuations fell sharply, or if a model provider went dark, is a regulator that thinks both are plausible. The speech also explains why the herding risk is hard to spot from inside one firm: each institution's model may be sensible on its own, and the danger only appears when everyone runs a similar one.
Sources: BIS — Tao Zhang: Artificial intelligence, monetary policy and financial stability (Oct 21, 2026)