By The Grey Ghost

Central banks built modern monetary policy around communicating clearly with markets. Agentic AI introduces a new possibility: machines capable of interpreting those signals at enormous speed, acting on them autonomously, and potentially changing the market signals policymakers rely on in return.

THE SIGNAL

For decades, central banks have worked to become easier for financial markets to understand.

Interest-rate decisions are explained.

Economic forecasts are published.

Officials give speeches.

Meeting minutes are released.

Press conferences provide additional clues about what policymakers are thinking and what might happen next.

That transparency serves an important purpose.

Monetary policy does not operate only through an interest-rate decision. It also operates through expectations.

Markets try to anticipate the central bank. Bond yields move. Currencies adjust. Lending conditions change. Investors reposition capital.

Artificial intelligence does not eliminate that system.

It could change who is doing the interpreting and how quickly they can act on what they learn.

At the 2026 Jackson Hole Economic Policy Symposium, Princeton economist Markus Brunnermeier presented research examining what happens when increasingly autonomous AI agents enter financial markets.

His paper identifies a potential condition he calls “asymmetric understanding.”

AI agents may become increasingly capable of learning how humans and institutions behave while the humans overseeing the financial system may have much greater difficulty predicting how those AI agents themselves will behave.

Brunnermeier argues that this asymmetry could make financial prices harder to interpret and place central banks at a strategic disadvantage when communicating with markets.

That is the Deep Signal.

AI does not need to control monetary policy to disrupt the relationship between monetary policy and markets.

It may only need to become extremely good at understanding the institutions making the policy.

WHAT THE MARKET IS MISSING

The obvious AI-and-finance story is about prediction.

Can AI predict stocks?

Can it forecast bond yields?

Can it identify recessions?

Can it outperform analysts?

Those are important questions.

But central banks introduce a different problem.

A central bank deliberately produces information about itself.

Speeches, policy statements, meeting minutes, economic projections, voting records and previous decisions create an enormous historical record of how policymakers have responded to different economic conditions.

Humans already analyze all of it.

AI potentially changes the scale of that analysis.

A sufficiently capable system could continuously compare current communication with previous statements, economic conditions, market reactions and policy decisions.

It could examine whether particular changes in language historically preceded changes in policy.

It could analyze how individual policymakers reacted to inflation, employment, financial stress or market volatility.

It could update those relationships every time new information appears.

None of that requires secret information.

The advantage can come entirely from extracting more meaning from public information than humans can extract at human speed.

That creates a paradox for central banks.

Transparency was designed to reduce uncertainty.

But in a market increasingly populated by sophisticated AI systems, transparency also creates an expanding body of machine-readable information about how the institution behaves.

The better markets can model the central bank’s behavior, the more difficult it may become for policymakers to surprise those markets when surprise is necessary.

And once AI agents can not only interpret information but independently take actions based upon it, the issue moves beyond analysis.

It becomes market structure.

FIRST-ORDER EFFECTS

1. Central-bank communication becomes more valuable data

Every policy statement provides information.

Every speech provides information.

Every forecast provides information.

Every decision made under a particular set of economic conditions creates another observable relationship.

That does not mean central banks should become secretive.

It means the value of their communications changes when machines can analyze them at scale.

The traditional objective is:

Help markets understand policy.

The emerging complication becomes:

How much can increasingly capable machines infer about future policy from everything the institution has previously said and done?

Brunnermeier’s research does not conclude that central-bank transparency has already failed.

It raises the more important structural question of whether monetary-policy frameworks designed for human interpretation remain as effective when autonomous AI agents become significant market participants.

2. The information race gets faster

Algorithmic trading is not new.

Financial institutions have used automated systems for decades.

Agentic AI potentially adds something different: systems capable of interpreting information, reasoning across multiple steps and taking sequences of actions with increasing autonomy.

Bank of England Deputy Governor Sarah Breeden said in June that finance is likely to evolve toward operating more autonomously and at greater scale and speed, including AI agents capable of devising and executing trading strategies.

That creates an emerging divide that has little to do with whether someone is a professional or retail investor.

The more important distinction could become:

human-speed interpretation

versus

machine-speed interpretation.

A central-bank statement can take a human analyst minutes or hours to fully analyze.

Machines can potentially compare it against enormous bodies of information almost immediately.

The information is public.

The ability to extract value from that information may not be equally distributed.

3. Similar machines could reach similar conclusions

The biggest risk may not be that one AI becomes extraordinarily intelligent.

It may be that many financial AI systems respond to the same information in similar ways.

The Bank for International Settlements, Bank of England and Deutsche Bundesbank are already examining this question through Project Logos.

The ongoing project is developing a simulated financial-market environment where researchers can compare traditional rules-based portfolio managers with LLM-based portfolio agents.

Among the questions being studied is whether common AI infrastructure could increase similarity in financial decision-making and what conditions could amplify or reduce correlated behavior.

There are no published Project Logos results proving widespread AI herding.

That is precisely why the experiment matters.

Central banks are testing the possibility before autonomous financial agents become dominant market participants.

Consider the mechanism.

A central bank releases unexpected information.

Multiple AI systems process it.

Their underlying models, economic information and objectives may overlap.

Several independently reach similar conclusions.

Capital begins moving in the same direction.

Those price movements become additional information.

Other systems react.

A feedback loop can form without any machine coordinating with another.

No conspiracy is required.

The systems simply need to interpret the world similarly.

SECOND-ORDER EFFECTS

The second-order problem is considerably larger.

Central banks don’t only influence markets.

They read them.

Bond yields tell policymakers something about inflation and growth expectations.

Credit spreads provide information about perceived financial risk.

Currencies react to changing expectations about interest rates and economic strength.

Market prices therefore function partly as information flowing back toward policymakers.

Now introduce increasingly autonomous AI agents.

A market movement could still reflect genuine changes in economic expectations.

But it could also increasingly reflect automated systems reacting to:

central-bank communication,

other AI systems,

short-term price movements,

common datasets,

or similar model-generated conclusions.

That complicates the information coming back to policymakers.

The chain can become:

Central bank communicates

↓

AI interprets

↓

AI acts

↓

Markets move

↓

Other AI systems interpret the movement

↓

Markets move again

↓

Central bank interprets the market

The central bank could therefore find itself analyzing market signals that were partly created by machines analyzing the central bank.

That is where asymmetric understanding becomes more than an academic concept.

Brunnermeier argues that sufficiently advanced AI participation could make financial prices harder to read, reducing their usefulness as signals to policymakers.

The machine doesn’t have to fool the central bank intentionally.

The market simply becomes harder to interpret because the actors producing the prices have changed.

CENTRAL BANKS MAY NEED AI TO UNDERSTAND AI

There is a logical response to a market becoming increasingly machine-driven.

Central banks themselves become more machine-capable.

Breeden has argued that central banks need to think not only about managing AI adoption across finance but about how AI should change the way central banks perform their own responsibilities.

That could eventually produce a financial system containing several interacting layers:

AI trading agents

AI portfolio-management agents

AI fraud and risk systems

AI market-surveillance systems

AI regulatory tools

AI central-bank analytical systems

Humans would not necessarily disappear from the decision chain.

Humans could continue setting mandates, objectives, risk tolerances and final policy.

But an increasing amount of the information being processed underneath those decisions could move machine-to-machine.

That represents a very different financial system from the one for which today’s monetary-policy communication frameworks were designed.

THE TRANSPARENCY PARADOX

This is the part worth watching most closely.

Modern central banks generally learned that predictability and communication could improve monetary policy.

AI doesn’t make that principle wrong.

It introduces a competing consideration.

Information meant to make policy understandable to people is also information machines can analyze.

And machines may eventually extract relationships from those communications that policymakers never deliberately intended to signal.

A central banker might consider two phrases practically interchangeable.

An AI system examining thousands of historical variables might discover that one phrase has historically correlated with a particular future policy outcome.

Whether that relationship is economically meaningful or merely statistical is another question.

But the machine can trade on the relationship either way.

The issue therefore isn’t simply whether central banks disclose too much.

The deeper question is whether communication frameworks built for human interpretation need to evolve once machines become a major part of the audience.

Reuters reported that Brunnermeier’s Jackson Hole presentation raised precisely these kinds of unconventional questions about central-bank communication in a world of increasingly capable financial AI.

WINNERS

Financial institutions with superior AI capability

The advantage could come from processing public information better and faster rather than possessing secret information.

Institutions with stronger models, proprietary data, better infrastructure and specialized financial AI could gain an increasingly important analytical edge.

Central banks that build their own capabilities

Central banks that understand autonomous systems early may be better positioned to interpret markets as those systems become more important.

The Bank of England is already arguing that technological change should alter how central banks themselves work.

AI monitoring and financial-risk technology

A financial system populated by autonomous agents creates demand for systems capable of identifying correlated behavior, unusual interactions and emerging systemic risks.

Organizations studying systems rather than individual models

The major question may eventually become less about whether one AI makes a good decision and more about what happens when thousands of AI systems interact simultaneously.

LOSERS

Human-speed market participants

Humans do not suddenly become incapable of investing.

But their ability to respond first to public information becomes increasingly difficult to preserve against systems capable of processing enormous datasets almost instantly.

Smaller institutions without advanced AI infrastructure

If sophisticated financial AI becomes expensive and strategically important, competitive advantages could concentrate among institutions capable of developing or purchasing the strongest systems.

Policymakers relying exclusively on historical market signals

If the composition of market participants changes, the meaning of market movements may change with it.

Historical assumptions about what a bond-market move represents may become less reliable if autonomous machines increasingly participate in producing that move.

Anyone looking only for catastrophic individual AI failures

The systemic risk does not require a rogue AI.

Thousands of perfectly functioning systems can still create instability if they independently react to the same information in similar ways.

WHAT HAPPENS NEXT

Watch Project Logos.

Not because it has already proven that AI portfolio managers will destabilize markets.

It hasn’t.

Watch it because central banks are now building environments specifically designed to observe how LLM-based financial agents behave under controlled market conditions.

Questions worth watching include:

Do different LLM portfolio managers converge toward similar decisions?

How strongly do they respond to identical information?

Does model diversity reduce correlated behavior?

What happens under market stress?

Can autonomous agents create feedback loops?

Do existing financial safeguards work when decision-making occurs at machine speed?

And eventually:

Can central banks reliably understand a market increasingly populated by agents capable of studying the central banks themselves?

The immediate answer remains unknown.

That’s why this qualifies as a Deep Signal rather than a prediction.

The pieces are beginning to appear before the final outcome is visible.