Central Bank AI Governance: A Regime Shift

Central bank AI governance arrived quietly, dressed as procedure. The market reads it as compliance. I read it as a structural shift in how monetary power operates.

The Precedent Markets Forgot

In 1987, automated program trading turned a decline into a 22.6% single-day collapse on the Dow. The mechanism was simple: correlated machines, identical rules, synchronized selling. The Brady Commission documented it in 1988 (Federal Reserve, 2007).

The context of that October was specific. The structure was general. Today the same structure returns, wearing a fresh name.

Three precedents are sufficient to call it a pattern. 1987 program trading. The 1998 LTCM unwind, where shared models produced shared exits. The 2010 flash crash, when algorithms pulled liquidity in minutes (joint SEC-CFTC report). Each episode shared one trait: automation compressed decision time toward zero.

Central bank AI governance sits inside this lineage. Public institutions and private firms now run correlated models on the same data. The question for allocators is direct. Who absorbs the correlated error?

What Central Banks Are Building Now

The Bank for International Settlements runs an Innovation Hub across several centres. Its projects are public, and most readers overlook them.

Project Aurora studied machine methods to detect money laundering across payment networks. Project Gaia applied language models to climate risk disclosures. Project Agorá, launched in 2024, brings seven central banks together with private banks to test tokenized cross-border payments.

Read them together and a direction appears. Monetary authorities want machine systems inside the plumbing of finance, from settlement to supervision.

The European Central Bank has published work on machine learning for supervisory data. The Federal Reserve studies model risk in its own operations. National frameworks around enterprise AI follow the same arc: the EU AI Act entered into force in 2024, staging obligations through 2027.

The pieces look separate. They form one architecture. Public supervisors and private enterprises converge on shared tools, shared vendors, and shared silicon.

The Mechanism: Correlated Automation

Here is the causal chain, step by step.

Central banks adopt machine models for speed and scale. Enterprises adopt the same class of models, often from the same handful of vendors. Both draw on similar training data and similar cloud providers. The result is convergence.

Convergence breeds correlation. When many actors run similar logic on similar inputs, their responses align. Alignment feels efficient in calm conditions. Alignment turns dangerous under stress, because everyone moves the same way at the same moment.

This is the 1987 mechanism at institutional scale. Add one twist. In 1987 the machines were tools inside private firms. In this decade the supervisor itself holds a machine model.

The referee and the players read the same screen. That changes the failure mode. A shared model error propagates across regulator and regulated together. Done poorly, this governance manufactures the systemic risk it claims to police.

The Semiconductor Chokepoint

Strip away the software debate and one fact remains. Machine capability rests on physical fabrication.

The most advanced logic chips come from a short list of fabs: TSMC in Taiwan, Samsung in South Korea. Models replicate easily. Fabrication plants do the opposite. US export controls on advanced GPUs to Chinese firms, tightened across 2022 to 2024, rank among the most consequential geopolitical moves of the past five years.

Central banks in the West and enterprises across the OECD depend on the same compute supply. That supply concentrates in one geography exposed to one strategic flashpoint.

Enterprise AI resilience, therefore, is a semiconductor question dressed as a software question. A board that plans model strategy while ignoring fab geography plans on sand. Whoever controls the fabs controls the outcome. The governance debate in Frankfurt, Washington, and Brussels rests on hardware made along a narrow strait.

My Position, And What Would Change It

My position is explicit. This is a structural regime shift, misread by markets as a compliance cycle.

The consensus treats it as paperwork. New rules, new officers, new reports. That framing underprices the deeper change. Public authorities are becoming operators of machine infrastructure, joining the private firms they supervise inside one technical stack.

This is a change of regime. Compare it to 1971, when the dollar left gold and the calibration of monetary models had to restart. The tools reshape the institution, then the institution reshapes the market.

What evidence would change my view? A clear separation between supervisory models and enterprise models, running on distinct data, distinct vendors, distinct silicon. Independent audit standards for public-sector AI, adopted across G7 authorities. Signs that regulators keep human discretion at the decision layer.

Show me diversification of the compute base away from one strait. Then I soften the thesis. Absent that, the correlation grows, and the fragility grows with it.

Three Implications For Capital

Three implications for capital.

First, family offices and sovereign funds. Over the next 36 months, reprice concentration risk in the compute supply chain. Exposure to a single fabrication geography deserves an explicit haircut in scenario models.

Second, boards and chief risk officers. The scenario missing from most VAR models is a shared-model failure across regulator and regulated. Correlated automation sits outside historical variance. Historical calibration underweights it by design.

Third, CFOs and investor relations. The macro narrative that machine adoption lowers operational risk deserves a second read. In aggregate, shared tools raise systemic correlation. The story sold to investors today may invert within 18 months.

The Prediction

The prediction, stated plainly.

By the close of 2026, at least one G7 central bank or the BIS will publish a formal framework governing internal use of machine models, with explicit audit and human-oversight clauses. The direction of travel already points there.

Confidence: Medium, around 65 percent. Horizon: through 31 December 2026. Verification: a published supervisory or internal-governance document from a G7 authority or the BIS Innovation Hub.

What To Watch

What to watch. Three leading indicators will confirm or weaken this thesis.

• BIS Innovation Hub pipeline: fresh programmes on model governance signal acceleration.
• Vendor concentration: procurement disclosures showing many central banks buying from one provider raise the correlation risk.
• Semiconductor flows: any disruption around Taiwan translates fast into compute scarcity for public and private models alike.

Map these. The pattern is legible. The timing is the harder part.