Banking on Climate

How do U.S. bank regulators weigh competing pressures when writing climate-related rules?

To some, climate and financial risks are inextricably linked. Erratic and increasingly frequent natural disasters have rendered insurance companies unwilling to extend policies in places where wildfires and floods are common. When uncharacteristically high summer temperatures cause drought, food prices and electricity bills spike.

To others, environmental policymaking falls beyond the scope of the central bank and other financial regulators who lack proper expertise. Does the central bank have democratic legitimacy to enforce meaningful climate mandates? Would involvement in climate policy risk jeopardizing its reputation?

These questions reverberate through the public discourse when politicians, private companies, advocacy groups, and interest organizations attempt to shape policy at the central bank. 

Whose voices rise to the top and ultimately shape decision-making? The answer, according to Rimmy E. Tomy and Kirill Skobelev at the University of Chicago Booth School of Business, is sobering for anyone who took the time to submit a public comment urging stronger climate action. Most of them, the research finds, did not move the needle. 

Tomy, Associate Professor of Accounting, sought to open the black box of climate policy rulemaking with the help of large language models (LLMs). With Skobelev, Predoctoral Researcher at the Center for Applied AI, the study showcases a new AI-powered way to study regulatory decision-making where LLMs serve as both analysts and stand-ins for human reviewers of public comments on draft guidelines.

A Glimpse Into A Closed Process

The study focused on a joint guidance from 2023, issued by the Federal Reserve, the FDIC, and the Office of the Comptroller of the Currency (OCC). The guidance sketched out how regulators expected banks to think about the climate-related risks on their balance sheets, and covered only large banks with assets greater than $100 billion. Such risks ranged from extreme weather damaging the assets that secure banks' loans, to fossil-fuel loans losing value as economy-wide decarbonization leaves those assets stranded.

Per policymaking standards, a public comment period was held between the initial draft and final guidance publication. The researchers collected 581 comments from individual citizens, advocacy groups, banks, and banking associations. 

For the Booth researchers, these comments were datapoints ripe with insights into what rhetoric, language, and perspectives most impact final policy outcomes.

“The behavior of the central bank could be influenced by the political leaning of the administration currently in power—we’re taking a closer look at which stakeholders’ voices matter when it comes to climate policy,” explained Tomy. 

Skobelev and Tomy systematically cataloged the differences between the agencies' draft proposals and the final guidance. The final version added accountability requirements, expanded the time horizons for risk assessment, and inserted materiality qualifiers. It also softened, or in some places omitted, references to reputational risk and litigation exposure associated with climate change.

What was read between the lines during comment period that led to such pointed changes?

“One way to get a closer look at the competing interests at play is to look at what stakeholders are saying because politicians eventually listen to their constituents,” Tomy said.  

The broader political backdrop is hard to miss. In early 2025, the OCC withdrew from the principles. By fall of the same year, the other agencies had followed suit. Their justification? The guidance was overly burdensome and carried the risk of diluting the main goal: to develop and implement monetary policy to support a stable economy. Around the same time, U.S. bank regulators also exited the Network for Greening the Financial System, an international group of central bankers focused on climate risk. 

But the rulemaking process itself—comments solicited, input weighed, and drafts revised—still happened. And that process, the researchers argue, is where the most revealing and lingering footprints of regulatory decision-making are left behind.

Scoring Climate Comments

The researchers used LLMs to rank comments based on the degree of their climate advocacy from 1 to 5. Scoring ‘1’ meant strong opposition to climate action; ‘5’ meant strong advocacy for binding goals and reform. As a failsafe, human auditors also ranked random public comments. LLMs and humans were largely aligned on the categorization of comments.

Perhaps unsurprisingly, pro-climate commenters were more likely to comment on concerns about continued investment in fossil fuel and fair lending. Those ranked as less climate-concerned focused on regulatory costs and the impact on community banks.

The striking portrait that emerges is sharply polarized. Private individuals, about 58.5% of commenters, averaged a climate-engagement score close to 5, as did climate advocacy and nonprofit groups. Banks (12.6% of commenters) and banking associations (8.3%) scored below 3 on average. Comments focused on fair lending and alternatives to fossil fuel financing tilted strongly toward climate action. 

Comparing the policy draft with the final guidance, the researchers found that more pro-climate comments were less likely to be reflected in the revisions.

Regulator Roleplay

The decision makers shaping climate and banking policy are potentially susceptible to social, political, and institutional pressures. To factor in this human element, Tomy and Skobelev used LLMs to roleplay as regulators with different motivations and considerations. The risk-averse bureaucrat may be motivated to preserve their career, while the average U.S. citizen may be primarily concerned about everyday price changes. Roles also captured perspective differences along partisan lines, wealth distribution and status, and general interest in global welfare. 

“Each role was given a specific directive: ‘Follow your mandate to maintain stipulative price stability and full employment,’” Tomy explained. “Or, for the career-oriented Bureaucrat role, ‘You’re risk averse, you don’t want to ruffle any feathers.’ And for the Top 1% Wealth role: ‘You want to protect the wealth of the top 1% of Americans.’”

The researchers followed a two-stage approach to operationalize the role-playing exercise. In the first stage, the LLM was asked to accept or reject a public comment based on its role and provide a summary of any accepted changes. In the second stage, the LLM was provided with the aggregated summaries and the draft policy and asked to write a revised policy for each role. 

The researchers find robust results in the first stage. When assigned Top 1% Wealth and Republican roles, both custom fine-tuned and large commercial LLMs disfavored pro-climate comments, while in Democratic and Worldwide Welfare roles they did the opposite. 

In the second stage, the researchers found suggestive evidence that revisions most closely resemble those of a risk-averse bureaucrat. However, the second-stage policy revisions were highly sensitive to model choice. 

The capacity for AI to support nuanced identification of small language quirks, potential bias, political leaning, and possible motivations may open up new applications in the policymaking sphere.  Caution should still be taken when using LLMs as human surrogates, especially when they are conditioned to take on specific roles.

The full working paper, Bank Regulators and Climate Action: Evidence from Supervisory Guidance, was supported by the Center for Applied Artificial Intelligence (CAAI).