A machine takes the minutes

A machine takes the minutes

IN DECEMBER 2024 four researchers at the Federal Reserve Board published a note with a modest purpose: to see whether off-the-shelf artificial-intelligence models could read the minutes of the Federal Open Market Committee (FOMC), the Fed's rate-setting body, as well as human beings can. The results were quietly impressive. The 11 models, set loose on some 25,000 sentences of minutes stretching back to 2010, classified the topics under discussion—unemployment, inflation, banking stress—with average F1-scores (a blend of precision and recall) of 0.80–0.93, and the best, GPT-4o, reached 0.95, beating a specialised financial model trained for the task. The detail worth pausing on comes at the end. The researchers who marked the machines' test answers were people who had spent years reading the minutes—and drafting them.

For a central bank, the minutes are not a corporate chore. They are the product. Because the Fed nudges policy mostly through words—forward guidance (its promises about the future path of rates), the "dot plot" of members' rate expectations, the chairman's press conference—the published account of the committee's debate is the transmission mechanism by which policy reaches markets. Whoever decides which arguments survive in the summary, how dissent is framed and what the committee is described as thinking exercises a quiet form of power. That editorial labour used to be confined to a handful of senior staff who sat in the room and then chose their words with care. It is now being shared with a machine.

Nobody should overstate what has happened. The Board's own note concedes that central banks do not yet use generative AI to draft their public communications. And in a speech in February 2026 on "operationalising AI", Christopher Waller, a governor, set out a deliberately managerial agenda: a common internal AI platform available to all Reserve Bank employees, "business-led" and "AI-enabled" adoption, and a checklist of guardrails—rigorous model validation, vigilance about hallucinations and human accountability. The uses he described are the sort of thing any large organisation would automate: condensing the background briefs that prepare the committee's meetings, triaging mountains of e-mail, helping software developers write tests. Jerome Powell, the chairman, has been if anything dismissive of the technology's near-term clout, telling senators in June 2025 that its effects on the American economy "are probably not great at this time."

The trouble with the clerk's view is that the clerk shapes the agenda. A machine that condenses pre-meeting briefs into "key themes" decides what the committee attends to first; an economist whose first pass through interview notes has been accelerated by a model is shaping the questions she asks next. The infiltration, moreover, flows in three directions at once. The Fed must learn to measure an AI economy it cannot yet see: the chairman has likened setting policy when official data go dark to driving a car in the fog, unsure whether jobs are slumping or price pressures mounting. It must manage the machines inside its own building. And it must cope with the machines outside that read it—the market models trained on Fed-ese, and the academic simulators built to imitate its decisions.

The last of these is further along than it looks. Researchers at George Washington University, led by Tara Sinclair, have built "FOMC in silico", a simulation in which large language models, the technology behind chatbots such as ChatGPT, adopt the personas of real committee members, grounded in their speeches, voting records and prose styles. Before the committee's actual verdict in July 2025, the ersatz committee held rates at 4.25–4.5%, exactly as the real one did, and reproduced the dissent of Christopher Waller and Michelle Bowman—described by the researchers as the first dual dissent for lower rates since 1993. Run against history, it matched the committee's rate within 25 basis points in 93% of meetings since 2000. The disquieting corollary, the researchers conclude, is that the models had absorbed something unflattering: the Fed is only partially insulated from politics.

None of this means a machine will soon set rates. It means something subtler and, for an institution whose authority rests on words, more corrosive. Four hazards stand out. Summarisers smooth; and dissent, at the Fed, is a signal—the texture of disagreement is data for markets, and a record tidied by a model loses some of it. The models are also trained on the Fed's own editorial judgement: the "ground truth" in the Board's experiment was the researchers' sense of what a sentence meant, an inheritance the machine then launders into something that looks like an objective reading. Add circularity: the words the Fed publishes become training data for the models that markets use to price its next move, and the Fed reads the markets that read it. The echo chamber closes. And accountability, which Mr Waller's own speech wraps in fine print, starts to slip: if a machine condenses the arguments, no single person can vouch for what the committee actually heard.

Mr Waller reached for the right historical analogy and, in doing so, conceded the point. He compared adoption to the arrival of ATMs, which changed the teller's job without abolishing it. True—and the reason is instructive. The teller survived precisely by becoming the person who handled the exceptions, the irregularities and the judgement calls the machine could not reach. So it will be inside the Fed. The human's residual value is the interpretation; and interpretation, at a central bank, is the whole business.

None of this argues for a Luddite retreat. The machines are good at the drudgery assigned to them—reading 25,000 sentences of minutes is exactly the work an institution should outsource—and a central bank that ignored the technology would fail in its duty to understand the economy it steers. Three lines should hold. The distinction between condensing and interpreting must remain human, because the sentence that frames the debate is the product, not the summary. Every machine-produced draft should bear a named human owner. And the Fed should treat its published words as what they have already become: inputs to the models that price its policy, deserving of the same care as the policy itself. Machines at the Fed can do everything except vouch for what their words mean. The institution's credibility is precisely that it can. It should keep that job for itself.