Speaking at the Financial Innovation Summit in Tokyo on the third of March, Governor Kazuo Ueda confirmed what had been the subject of considerable informed speculation since the autumn: that the Bank of Japan had completed the formal integration of a large language model into the briefing materials prepared for the Policy Board ahead of each rate-setting meeting. He described the system as "an extension of analytical rigour, not a replacement for human judgment," and in so doing reached, perhaps unconsciously, for the exact language that central bankers have employed whenever they have introduced a powerful new analytical tool and felt obliged to reassure their audiences that the tool would not replace the judgment it was designed to inform.

The language is familiar because the situation is familiar. When the Federal Reserve's economists began using the MPS econometric model in 1970, a joint construction of researchers at MIT, the University of Pennsylvania, and the Social Science Research Council, the same reassurances were offered. The model was a formidable analytical apparatus; it would, however, be an input to the Federal Open Market Committee's deliberations, not a substitute for them. For roughly a decade, that balance held. Then came the Lucas critique of 1976, which argued with considerable force that large-scale structural models of precisely the MPS variety were "all but useless in analysing the future effects of policy changes," because they could not account for the manner in which private agents would revise their expectations in anticipation of policy [4]. The Fed did not abandon its models; it refined them, incorporating rational expectations mechanisms, and produced what eventually became FRB/US in the mid-1990s, a model still in use today with approximately 380 equations. The lesson of that fifty-year arc is not that models are dangerous, but that institutions adopt them confidently, discover their limitations gradually, and have rarely managed the transition from confidence to caution without some embarrassment.

The Bank of Japan's new system, developed in collaboration with a Japanese technology consortium whose precise membership the Bank declines to specify in full, is trained principally on Japanese-language economic data: Ministry of Finance bond issuance records, regional branch survey results, Statistics Bureau consumer price releases, and the quarterly Tankan business survey, which generates approximately 200,000 individual responses per cycle that the Bank's Research and Statistics Department has historically required several working days to process into briefing-ready synthesis. The model ingests this material alongside global macro feeds, JGB yield curve data, and the communications of other major central banks and produces, in advance of each Policy Board meeting, a structured summary of the conjunctural position as it stands in the hours immediately before Board members convene. It does not produce a rate recommendation; it produces a synthesis.

The significance of this step resides not only in its novelty but in its institutional weight. The Bank of Japan is not a fintech. It manages the monetary conditions of the world's fourth-largest economy [by nominal GDP], holds some nine trillion yen in government bonds on its balance sheet as the legacy of a decade and a half of quantitative and qualitative easing, and concluded its negative interest rate policy only in March 2024, raising its policy rate to 0.75 per cent by December 2025, the highest level since 1995 [5]. It is, in other words, a bank in the middle of one of the most consequential monetary normalisation episodes in its modern history. The decision to introduce an LLM into the briefing pipeline at precisely this juncture is either a statement of confidence in the technology's maturity, or evidence that the volume and complexity of incoming data had become genuinely unmanageable by purely human means. Governor Ueda's remarks suggest the latter motivation is at least as important as the former.

The broader context for the deployment is the rapid diffusion of generative AI through Japan's financial sector more generally. The Bank of Japan's own survey, published in September 2025, found that approximately fifty per cent of Japanese financial institutions were already actively using generative AI, with over seventy per cent including those conducting trials [3]. The practical applications reported were overwhelmingly documentary: summarising, translating, synthesising. The BoJ's LLM integration is, in that sense, an intensification of a trend already well established among the institutions it supervises. The distinction is one of institutional consequence: a commercial bank deploying a language model to draft loan summaries is a productivity tool; the same technology applied to the briefing materials of a body with authority over short-term interest rates for one hundred and twenty-five million people is a matter of monetary architecture.

The objections worth taking seriously are three. First, the training set. A model trained predominantly on Japanese-language economic data from a period of persistent deflation and zero or negative interest rates has encountered, as its primary historical experience, precisely the anomalous conditions that the Bank is now in the process of exiting. The model's implicit priors are likely to reflect a world that is receding rather than the world that is arriving. Second, the hallucination problem, to use the inelegant but accurate terminology: large language models produce plausible-sounding text that is, on occasion, factually incorrect, and the error is not self-announcing. A synthesis document that confidently misrepresents the most recent CPI release by a fraction of a percentage point is not a trivial risk in the context of a policy meeting at which the difference between holding and moving may be argued on precisely such marginal data. Third, the opacity of the process itself: when the minutes of a Policy Board meeting record that the Board "considered available economic indicators," one will now be unable to determine from the published record whether those indicators were presented to the Board in their raw form or as mediated through a model whose internal workings remain proprietary to the vendor consortium.

The IMF's working paper on large language model analysis of central bank communication, published in June 2025, noted that LLMs "enable sentence-level understanding of meaning and intent across multiple communication dimensions" and applied a fine-tuned model to a dataset of 74,882 documents from 169 central banks spanning 1884 to 2025 [1]. That paper was concerned with analysing central bank communication from the outside. The Bank of Japan's integration runs in the opposite direction: it places the technology inside the institution, as a participant in the deliberative infrastructure rather than a tool for its retrospective interpretation. The IMF authors, one suspects, would find that distinction worth noting.

Governor Ueda is a careful man, and a careful economist. He will have considered these objections before speaking at the FIN/SUM. What he offered in response, at least in the portion of his remarks that entered the public record, was the reassurance of the qualified professional: the system is an aid, not an authority. That may well be true of the system as currently deployed. The question worth holding in mind is the one that the history of central bank model adoption has posed before and will pose again, namely whether the distinction between aid and authority is as stable over time as it appears at the moment of first deployment.