In the Media
FINANCIAL TIMES: What AI should learn from the banks
09. 23. 2026
Both industries change rapidly and reach throughout the economy, requiring officials to weigh the benefits of innovation against risks
Click to add an image!
Dario Amodei wants independent evaluators to have desks inside AI companies. In his essay on the need to pace the frontier, Anthropic’s chief executive promises outsiders key cards, laptops and nearly full access so that they can monitor how the company’s systems are working in real time rather than simply testing them before release. He makes the case for governments to formalise this arrangement, pointing to bank supervision as a precedent.
It is a good comparison. Like AI, banking changes rapidly and reaches across the economy, forcing public officials and private firms to continually weigh the benefits of innovation against its risks. The Federal Reserve, FDIC and the Office of the Comptroller of the Currency (OCC) all embed teams inside the commercial banks they supervise to learn how the businesses operate and where their weaknesses lie. These supervisors meet with staff and management, review internal documentation and keep up with day-to-day operations. Their job is to assess whether leadership is adequately managing risk in accordance with regulation — not just whether paperwork is in order. If they are not, supervisors can require the bank to fix the problem.
The need to adopt a similar approach to AI has never felt more urgent. This summer, AI agents showed they have become increasingly clever, co-ordinating with one another to meet their goals. Hacks illustrate how a lapse in oversight can spill into systems entirely outside a lab’s realm of operation.
Under a supervisory approach modelled on bank supervision, each frontier AI lab would work with a team of supervisors who understand its technology. Embedded from the start in model design and review, supervisors would ensure the humans and agents building these systems are adequately focused on preventing bad behaviour.
Supervisors would also oversee stress tests for the models. Each year the largest banks must show how they would fare in a severe downturn, against scenarios the Federal Reserve updates and (until recently) has kept confidential so banks cannot train for the test. AI systems will pose their own systemic risk if they manage to gain control of electricity, payment or transportation systems. A stress test would require each company to show that its staff could still override or shut down a model under those conditions, against scenarios the company has not seen in advance.
We need to learn lessons from the success or failure to regulate other sectors. Policymakers failed to regulate social media because they were overwhelmed by the range of problems, ranging from teen mental health to privacy and misinformation. The challenges of AI can be similarly intimidating, but Congress does not need one massive bill to deal with AI. Instead, it should start with the most pressing concern today — model safety — and go from there. The work will require a regulatory agency, perhaps called the Office of AI Supervision, to translate requirements into rules.
Bank supervision is far from perfect. Officials have sometimes identified dangers and then failed to act. In other cases, leadership has attempted to strip discretion out of supervision altogether, reducing it to the completion of narrowly configured checklists. Supervisors can become compromised or “captured”, causing them to inaccurately assess the nature of risk.
To mitigate these challenges, an AI agency would need staff who understand the technological paradigm and are unafraid to bring their power to bear. Senior examiners should rotate between companies, just as they do in banking. The OCC caps an examiner-in-charge’s tenure at a single bank at five years. Given how fast AI moves, three would work better.
Congress has already started to consider these ideas. The bipartisan Frontier Act, which OpenAI has partially endorsed, would require the largest AI companies to retain federally licensed assessors for ongoing safety reviews. But this stops short of public supervision. Companies could hire their own assessors, risking the conflicts that compromised credit rating agencies before 2008. No independent supervisory agency would be charged with overseeing the labs, and there would be no embedded supervisors as in banking.
AI holds enormous promise, and like any other technology it brings risks that have to be managed. A bank supervision model for AI regulation creates a way for public officials and private experts to assess those trade-offs together, while ensuring that the companies profiting from a decision are not its only judges.