An AI sets or votes on a central bank's policy rate

Where things stand (15 Oct 2026)

I found no G20 central bank or ECB statement, pilot, consultation or bill that would let an AI system set a policy rate or cast a binding vote on one. What the banks have said points the other way:

• Russia: Governor Nabiullina confirmed the Bank of Russia uses AI but stressed it cannot make decisions in place of the Board of Directors (Expert.ru, 21 Apr 2026).
• India: Governor Malhotra called the erosion of human judgment and accountability the main risk of AI (The Hindu, 11 Aug 2026).
• Canada: The Bank of Canada's new "Prima" model is a forecasting tool, and Governor Macklem says AI is used "with human control" (Wealth Professional, Sept 2026).
• United States: Chair Warsh's Fed task forces study AI's economic effects "to inform the Federal Reserve's policy judgments" (federalreserve.gov task-forces page, July 2026). In his Jackson Hole speech (28 Aug 2026) Warsh called the Fed "a decisionmaking agency" and said even simple rules cannot yet be relied on mechanically.
• Korea: The Bank of Korea's in-house AI, "BOKI", does research, document analysis and translation (Korea Herald, 21 Jan 2026).
• ECB: ECB speeches on AI (e.g., Lane, 6 Jul 2026) are about AI's effects on the economy, not about AI making decisions.
• Commentators: Replacing the FOMC with an AI is still a podcast thought experiment. The practical step suggested there was to run an AI in parallel with the committee, which is a shadow vote and excluded here (CFR Spillover, 1 Oct 2026).

Reference class and base rate

No central bank, in the G20 or elsewhere, has ever given an AI a formal vote or delegated rate-setting to one. The nearest government case is Albania's AI "minister" Diella (Sept 2025). A year on, it still has no real decision role: the procurement agency has only finished drafting the specifications for the system (Balkan Insight, 14 Aug 2026). The FT (Apr 2026, headline and summary only, since the page would not load) judged that AI tools on corporate boards are "unlikely to be granted a vote." So the historical rate is effectively zero. The only reason the hazard is above zero is AI progress.

Paths to YES

1. New legislation. Most of the eligible rate-setting bodies have members fixed by statute or treaty (Fed, Bank of England, ECB, Reserve Bank of India, Bank of Korea, Bank of Japan, Reserve Bank of Australia, Turkish central bank, Bank of Russia, and others). Adding an AI voter there needs a law change, and none is in prospect. A plausible timeline is reform debate, then legislation, then a vote, which is unlikely to finish before about 2029–31 even if it started.
2. Internal or administrative change. In a few banks the committee or decision-maker is less tightly fixed by statute (Canada's Governor/Governing Council, South Africa's MPC). A bank there could make an AI vote part of a live pilot.
3. A member who transmits an AI's vote without discretion, confirmed publicly by the bank.
4. A government with fast decree powers (Saudi Arabia, Argentina, Türkiye) doing it for show or as a political move.

Every path also needs the bank itself to publish an explicit factual statement. Shadow, advisory and backtest uses, which are the most likely first steps, do not count.

Hazard model

The quarterly hazard rises in a straight line from 0.04% (Q4 2026, with nothing in the pipeline) to 0.18% (Q4 2031), reflecting growing AI capability and the time institutions need to adapt. I computed the cumulative probability with 1 - cumprod(1 - h) over the 21 quarters (h = linspace(0.0004, 0.0018, 21)). It comes to about 0.3% by end-2027, 1.1% by end-2029 and 2.3% by end-2031. As a cross-check, I put roughly a 20–25% chance on worlds with very fast AI progress where governments delegate aggressively (about 5–10% chance of a qualifying event in those worlds) and about 0.5% otherwise. That also gives roughly 2%.

Strongest case against my estimate

There are 20 eligible institutions, one non-decisive vote in a live pilot is enough, and AI could be clearly better than humans at forecasting by 2029–30. A tech-enthusiast governor could plausibly try it; Korea's governor Shin Hyun-song, who has led AI work at the BIS, is one example. That argues for a few percent or more. On the other side, central bankers keep stressing human accountability, new practices usually spread from small banks to large ones over many years, and even Albania's showpiece has stalled. Weighing both, I stay at about 2% by 2031 and near zero for the next few quarters.

Question. By the last day of each quarter, has the central bank of any G20 economy, or the European Central Bank (ECB), stated that an AI system set its policy interest rate or cast a vote on it?

Definitions.

• Eligible central bank: the national central bank or monetary authority of Argentina, Australia, Brazil, Canada, China, France, Germany, India, Indonesia, Italy, Japan, the Republic of Korea, Mexico, Russia, Saudi Arabia, South Africa, Türkiye, the United Kingdom, or the United States; or the ECB. This fixes the country coverage to the G20 membership at opening. EU or African Union membership alone does not make a country's central bank eligible. Institutional successors count.
• AI system: a machine-learning or other artificial-intelligence system identified as such by the confirming central bank, from any developer. Conventional calculators, fixed policy formulas, and software merely executing a rate chosen by humans do not count.
• Policy interest rate: an interest rate, or target rate range, officially designated by the bank as a monetary-policy rate. Choosing to leave it unchanged counts. For the ECB this includes any of the three key ECB interest rates. Market-determined rates, individual loan prices, and operational auction outcomes do not count unless they are themselves the bank's official policy-rate decision.
• Set: the system selected the rate or target range adopted as the bank's actual policy decision, exercising delegated decision-making authority rather than supplying a recommendation for discretionary human approval. Human authorization of the delegation, constraints on the system, or ministerial implementation of its decision do not disqualify it.
• Cast a vote: the system generated a vote on an actual policy-rate decision that was accepted as a voting contribution in the authorized decision-making body's formal procedure. It need not be decisive or on the winning side. A human may transmit the vote without exercising discretion over its content. A human independently choosing to follow AI advice is not an AI vote. For euro-area national central banks, a qualifying vote in the ECB Governing Council counts.

Resolves YES if an eligible central bank publishes an explicit factual statement confirming at least one completed instance of either setting a policy interest rate or casting a vote as defined above. The statement must identify the actual decision or voting occasion and the AI system's role; disclosing its technical name or weights is unnecessary. One qualifying decision or vote suffices, including a live pilot with actual policy authority.

Exclusions (resolve NO).

• AI used only for forecasting, research, scenario analysis, recommendations, drafting minutes or announcements, or implementing a human-selected rate.
• Advisory or shadow votes, simulations, backtests, mock committee meetings, and demonstrations without a formal voting role in an actual policy decision.
• Appointment of an AI to a committee, authorization of future AI rate-setting, or plans to delegate authority without confirmation of a completed decision or vote.
• A statement that AI influenced a decision without establishing either qualifying role.
• No qualifying publication by the relevant quarter-end.

Resolving source. Official statements, policy-decision releases, meeting minutes or accounts, institutional reports, or official transcripts published by the eligible central bank on its own domain. The confirming institution must assert the event as fact, not merely host an outside claim or a research proposal. This is a narrower resolving-source requirement than Public confirmation (A2): developer statements, press reports, and outside evaluations alone do not count. Apply the Resolution window and Resolution date (A3) conventions; the publication date, not the rate's effective date or the meeting date, determines resolution.