Amit Seru warns that AI models, already embraced by market participants, are better than humans at gaming complexity

The theme of last month’s annual gathering of central bankers and economists in Jackson Hole, Wyoming, was “financial innovation.” One major issue looming over the meeting was how new technologies, specifically AI, might affect financial stability.

Economists and policymakers have long grappled with how policy should respond to technological shocks. At least since Nobel laureate Robert Lucas’s work in the 1970s, economists have understood that households and businesses adjust their behavior when the economic environment around them changes, whether in response to major policy changes or to technological shifts. One cannot simply assume that relationships observed in past data will remain unchanged as people and firms adapt.

At one level, AI is doing what a technological revolution is expected to do. It is permeating the real economy, disrupting everything from engineering and sales to finance and human resources. It is also reshaping industries by accelerating the pace at which entrants can scale and compete with incumbents. All of these changes are likely to affect the data we observe in the future and, in turn, what we learn from it about the behavior of households and businesses.

But AI also challenges our ability to anticipate how actors will respond to financial regulation. As Princeton University’s Markus Brunnermeier observed at Jackson Hole, the relationship between humans and AI is marked by “asymmetric understanding”: it understands us, yet we do not quite understand what AI agents, individually and especially together, might do to achieve a goal. While we can establish rules and scenarios for agents to follow, we cannot account for every path they might take. This asymmetry becomes even more salient when many agents operate at scale.

Regulators and policymakers have faced a version of this problem before: rules and incentives can change the behavior of those subject to them. Banks and other financial institutions often respond to regulation in ways policymakers did not foresee. The same is true of regulators and supervisors who interpret and enforce those rules, since their decisions also reflect organizational incentives and norms. This is evident when different regulators looking at the same situation arrive at different conclusions.

In both cases, however, institutional constraints limit what humans can do and therefore the range of responses policymakers need to anticipate. AI agents can pursue a much wider range of possible paths, making their behavior less predictable.

The recent breach of Hugging Face by OpenAI’s agents illustrates the problem. The AI agents worked around the established rules to achieve a goal, finding unauthorized ways to communicate, pool information, and eventually breach systems outside the intended environment. Their designers, who had neither directed nor anticipated those moves, seemed as surprised as anyone.

One can imagine a similar problem in finance, given that banks, hedge funds, and trading firms have substantial incentives to use AI agents to find profitable paths around regulatory constraints. This does not mean that guardrails cannot be put in place, such as a “red team” of agents looking for such activity. But policymakers will be setting goals and rules without knowing all the ways agents might respond to them.

Whether we like it or not, AI is already being employed by various actors across financial markets, from banks to hedge funds and speculators. Increasingly, these agents will be tasked with finding new ways of making money. Regulators, who are in the business of stabilizing markets, will therefore have little choice but to embrace AI to circumvent the circumventers.

To succeed, policymakers must recognize the power of subtraction. Over the past two decades, regulators have made the financial system much more complicated and costly to navigate. Thousands of pages of rules and enormous regulatory discretion create many opportunities for AI agents to satisfy the letter of the law while defeating its purpose. Simpler, more robust rules, such as high capital requirements, might become more valuable precisely because AI is better than humans at gaming complexity.

Simpler rules are only one part of the response. Regulators must also decide how much information to share with market participants. If regulators provided banks with their complete models, the banks’ AI systems would optimize around them. On the other hand, too much opacity might create confusion and chill legitimate financial innovation.

But even simpler rules and the right degree of transparency will not be enough if regulators themselves cannot keep up with the institutions they oversee. Those institutions will increasingly operate at machine speed, with computing power allowing them to evaluate countless strategies that humans could never examine one by one. Regulators will need to design AI agents of their own that can test rules before they are adopted, search for ways actors might evade them, and monitor markets for behavior humans might notice only after the fact.

Doing so, and incorporating AI effectively into financial regulation, will require thoughtfulness, diligence, and, perhaps most importantly, regulators’ time. That, rather than computing power, might be the scarcest resource in efforts to ensure financial stability.

*Amit Seru is Professor of Finance at the Stanford Graduate School of Business and a senior fellow at the Hoover Institution. This content is © Project Syndicate, 2026, and is here with permission.