Central Bank AI: The Delegation the Market Is Ignoring

The Precedent: When Central Banks Bought Into Models

In 1968 the Federal Reserve adopted the MPS model, built together with MIT and the University of Pennsylvania. The logic was straightforward: replace human judgment with an equation.

The backdrop was rising inflation in the late part of the decade. That model promised stable forecasts and delivered systematic errors throughout the stagflation of the 1970s.

The structure of that choice is repeating itself today, under a different name and with far greater computational power. Three similar precedents are enough to call it a pattern. The lesson remains clear: a tool calibrated for one regime collapses when the regime shifts.

The model contained dozens of behavioural equations. Its mathematical elegance concealed a critical flaw: inflation expectations moved faster than its estimates.

Monetary institutions are drawn to tools that dilute individual accountability. An equation absorbs blame more efficiently than a governor. This dynamic explains half a century of technical choices.

The Current Pattern: Artificial Intelligence Enters the Reaction Function

Central bank AI describes precisely this phase. Central banks are testing machine learning models for nowcasting, real-time price monitoring, and textual analysis of data series.

The Bank for International Settlements is leading research on this front. Its economists are mapping how shocks propagate through inflation via direct effects and second-round effects.

The consensus frames adoption as inevitable and benign. I read the same script as 1968, amplified by the scale of available data. The difference lies in the speed at which errors propagate.

A statistical model fails slowly. A system trained on billions of parameters fails quickly and with apparent confidence.

Communication itself is now in scope. Some central banks are using language models to analyse their own minutes and measure the tone perceived by markets.

The functions involved expand every quarter: forecasting, communication, financial risk surveillance, liquidity management.

The Mechanism: Why Rate-Setters Yield to the Temptation

The cause is precise. Uncertainty about the persistence of inflation makes human judgment costly and politically exposed.

A central banker who is personally wrong loses their position. A committee that follows a shared model distributes reputational risk. AI offers exactly this cover, and does so with an aura of objectivity.

Observe the sequence. First comes research, then the internal pilot, then communication, finally the decision. Each stage erodes the space for discretionary judgment.

The causal mechanism leads to a predictable outcome. The more unstable the regime, the stronger the demand for tools that promise order.

That demand grows precisely when historical data loses its predictive value. This is the structural paradox.

Energy shocks illustrate the point. They alter the relationships between prices and output, and a model trained on the past reads these ruptures as noise.

The Weak Link: The Energy Shock and the Limits of Data

The BIS Bulletin No. 131, published on 5 August 2026^[1], ranks the recent energy shock among the most significant since the 1990s.

The authors make a point that is uncomfortable for advocates of automation: the correct monetary response depends on the persistence of pressures and the magnitude of the impact on growth, and it varies from economy to economy.

This destroys the idea of a universal model. Every jurisdiction requires different calibration, and the uncertainty surrounding these effects further complicates the choice.

A central bank AI system inherits the uncertainty embedded in its input data and amplifies it. Output quality depends on the quality of initial conditions.

The operational conclusion remains stark. Anyone calibrating a portfolio around monetary competence is assuming a stability that the current regime is actively eroding.

The BIS document remains cautious about the role of automation. That caution signals more than a thousand enthusiastic announcements.

My Position: Algorithmic Delegation Will Arrive Too Soon

I am taking an explicit position. G7 central banks will integrate AI into their official forecasting function before they have solved the regime-change problem.

This is a structural choice, independent of the electoral cycle or the governor currently in office. Institutional pressure toward reputational risk cover pushes in one direction only.

Recent history offers a test. During the pandemic, the inflation forecasts of major central banks failed spectacularly, with repeated upward revisions in 2021 and 2022. Those failures came from human-assisted models. Automation will inherit the same blind spot.

What would change my reading? One thing only: the publication of a validation framework that stress-tests models against regimes absent from training data. A counterfactual stress-test protocol, adopted and made public.

That protocol remains absent from official documents. For as long as it does, delegation precedes comprehension.

Three Implications for Capital

The market has yet to price the risk of monetary error amplified by automation. Here is what changes for those managing capital right now.

First, for family offices and sovereign wealth funds: the next 36-month horizon favours real assets resilient to monetary policy errors. Gold, energy infrastructure, agricultural land. Rate volatility will rise in step with model volatility.

Second, for chief risk officers: the scenario of a systemic forecasting error is missing from current VAR models. Add it. A model that fails in a correlated way across multiple central banks creates new tail risk.

Third, for CFOs and investor relations: the narrative of orderly disinflation could prove fragile within eighteen months. Prepare an alternative scenario before investors ask for one.

The Forecast

I am putting forward a verifiable prediction. At least one G7 central bank will announce the official integration of machine learning models into its inflation forecasting process by 30 September 2027.

Confidence: high, 72 out of 100. Horizon: 30 September 2027. Verification: an official document or a governor's speech confirming operational adoption.

The signal that would falsify the thesis is clear: the complete absence of any official announcement of this kind by the stated date.

What to Watch

Three indicators will confirm or refute the direction over the coming quarters.

• BIS and central bank research papers on the operational use of AI: frequency and tone.
• Hiring: heads of data science in monetary policy departments.
• Governors' speeches that cite forecasting automation as an operational tool, not merely as research.

Remember the rule of this desk. A structural pattern spans decades; a cycle spans a few years. This transition belongs to the first category.

The divergence between confidence in algorithms and the quality of regime data always resolves itself. The question remains how. Watch the data. Watch the committees.