Speaking at International Financial Week in Hong Kong on the twenty-sixth of January, Tao Zhang, Chief Representative of the Bank for International Settlements for Asia and the Pacific, delivered a warning that was at once specific in its technical content and remarkable in its institutional setting. Artificial intelligence and digital finance, he said, posed financial stability implications that may change the way in which central banks identify, assess, and manage systemic risks [1]. This was not an academic exercise, conducted by researchers who had published a paper and hoped for citations. This was the Bank for International Settlements, the central bank of central banks, placing artificial intelligence and the risks it carried firmly onto the prudential agenda of the world's apex monetary institutions.

The parallel that occurred to observers with longer memories was not difficult to identify. In the early 1990s, the BIS had issued warning after warning about the rapid expansion of over-the-counter derivatives markets. The warnings were remarkably precise. The derivatives markets were growing at a pace that far outstripped the institutions' capacity to understand the exposures those markets generated. Banks were creating interconnections, bilateral and multilateral, that regulators could neither map nor manage. The correlations embedded in those contracts meant that a failure in one institution could cascade through others with a speed and force that no central bank had yet seen in the modern era. The warnings were noted, filed, circulated among the financial regulators of the world's major economies, and substantially ignored. The warnings continued through 1995, 1996, 1997. And then, in the autumn of 1998, Long-Term Capital Management collapsed, the Federal Reserve organised a consortium to manage its unwinding, and the financial system discovered that it had learned something important about systemic risk. The present warnings about artificial intelligence in finance carry a structural resemblance to those earlier cautions; it is worth examining what that resemblance might portend.

Zhang's speech laid out the BIS's understanding of the risk channels through which artificial intelligence might amplify systemic vulnerability. The first channel operates through the acceleration of trading and portfolio adjustments. Artificial intelligence systems, particularly those deployed for high-frequency trading and algorithmic portfolio rebalancing, can respond to market movements far more rapidly than human traders. During periods of market stress, this acceleration may intensify short-term price movements and exacerbate volatility. The second channel runs through operational dependencies. Artificial intelligence systems require infrastructure, computational power, and the specialised expertise to maintain them. To the extent that multiple institutions rely upon the same technology providers, those providers become themselves a systemic risk. A technology failure, a cybersecurity incident, or the bankruptcy of a key vendor could become, in that scenario, a financial stability event affecting dozens or hundreds of institutions simultaneously. The third channel emerges from behavioural correlation. To the extent that large numbers of financial institutions employ artificial intelligence systems trained on similar datasets and optimised according to similar objectives, those systems will tend to make similar decisions and respond to market conditions in similar ways. In a stress scenario, this herding effect could enable shocks to propagate through the financial system with exceptional speed and force.

The regulatory apparatus has not been inactive. The BIS published its report on the use of artificial intelligence for policy purposes, submitted to the Group of Twenty in October 2025 [2]. The Financial Stability Institute issued its brief on regulating artificial intelligence in the financial sector in December 2024 [3]. The European Union has extended its AI Act into the financial services domain. The Federal Reserve has issued draft supervisory guidance requiring banks with assets over one hundred billion dollars to maintain centralised inventories of all artificial intelligence models in use at their institutions. The quantity of regulatory paper on artificial intelligence has grown with remarkable rapidity; a survey of publications by the major regulators shows output rising from single digits in 2020 to double digits by 2024. Yet Zhang's remarks suggest that the BIS, at least, harbours a conviction that the pace of regulatory response may not be keeping pace with the speed of adoption and deployment in the private sector.

There is, however, a counter-argument worth entertaining. The speed of artificial intelligence adoption in finance may paradoxically be creating its own regulatory feedback loop. Every incident of consequence, however minor in absolute terms, generates new guidance, new working papers, new supervisory expectations and memoranda. The sheer volume of regulatory output on the subject of artificial intelligence risk may itself be evidence that the system is functioning as it was designed to function, that the warnings are being heard, absorbed, and integrated into the supervisory framework. Perhaps the BIS's warnings are not being ignored; perhaps they are being absorbed into the institutional apparatus so thoroughly that they no longer appear as warnings at all, but merely as background assumptions against which prudent supervision unfolds.

This interpretation may offer some comfort, but it rests upon an assumption about institutional dynamics that the historical record does not entirely support. The lesson of derivatives risk in the 1990s was not that regulators failed to identify the problem. They identified it with admirable precision and clarity. The lesson was that they failed to act with sufficient speed and force to constrain the problem before it became a crisis. The reason was not regulatory incompetence; it was political economy. The derivatives markets were generating extraordinary profits for major financial institutions. The prospect of pre-emptive regulation was thus politically unfavourable; institutions with influence lobbied against constraints; the political system preferred to watch and wait and gather more information before acting. Artificial intelligence in finance is following the same trajectory. The analytical apparatus for understanding the risk is sophisticated and growing more sophisticated. The theoretical frameworks are coherent and increasingly empirically grounded. The institutional will to constrain the deployment of artificial intelligence before the market teaches the system a costly lesson about the risks it contains remains, as it was in the case of derivatives, the binding constraint.

The Bank for International Settlements has been in the business of warning about financial risks for nearly a century. It warned about the viability of the gold standard. It warned about the sustainability of the Bretton Woods arrangement. It warned about the risks embedded in derivatives markets. The record is not one of prophecy, of ability to discern what the future holds; it is one of careful, methodical identification of structural vulnerabilities in the financial system that the political apparatus preferred not to address until it had no choice. Whether artificial intelligence in finance follows the same arc, progressing from warning through crisis to belated reform, depends ultimately on whether the present generation of regulators and policymakers can accomplish what their predecessors did not: the capacity to act before the system teaches them a lesson they have already learned.