Over the past year, the debate about artificial intelligence in the financial world has been turned on its head. While commercial and central banks previously focused mainly on how much work the technology would save them and how it would boost productivity, they are now grappling with an entirely different set of questions. They want to know what will happen when machines trade against machines, who will be able to detect unplanned market manipulation, and how to supervise a financial system that understands its regulators better than they understand it.
Central bankers and economists received a comprehensive list of these concerns in August 2026 at the Fed’s symposium in Jackson Hole, Wyoming. Princeton economist Markus Brunnermeier presented a paper on artificial intelligence in the financial sector and described several scenarios in which the technology could destabilize markets and the entire banking system. He summed it up with an analogy that resonated in the room: carpenters build tables, while bankers build trust. And according to him, it is precisely trust and the functioning of institutions that artificial intelligence can undermine in entirely new ways. His presentation came across as an urgent call for authorities to begin preparing for this change immediately.
A market that understands the central bank better than it understands itself
The concerns stem from a simple imbalance. Artificial intelligence processes information so much better than humans that it will soon know with almost complete certainty how bankers will decide, even before the officials themselves have made up their minds. Algorithms will then tailor their trading strategies accordingly, circumventing the officials’ actions and profiting from them, whether legally or outside the rules. Professor Brunnermeier used the term asymmetric understanding to describe this situation.
Speeches, meeting minutes, and the entire history of central banks are ideal training data. Official institutions are becoming perfectly readable to machines, while the institutions themselves have no insight into the world of trading bots. According to Brunnermeier, financial markets may then contain less information and behave more unpredictably than they do today.
The response central banks might resort to would itself be harmful. They would have to communicate less clearly, reversing decades of progress toward openness, which is credited with reducing market volatility and promoting stability. In this environment, transparency becomes predictability that arms the counterparty. This is also the origin of Brunnermeier’s idea of holding two press conferences, one offering an intelligible narrative for people and the other serving as training data for machines. The author himself adds, however, that this would solve the problem only partially.
Manipulation that no one planned
The second concern involves trading bots that may coordinate their actions without communicating with one another in any way. In less liquid markets, even a small order can move the price, making it inexpensive to set an exchange rate swinging. Moreover, when a larger number of machines all unquestioningly converge on the same course of action, none of them is large enough to be labeled a manipulator. The classic scheme in which actors artificially drive up a price and then sell off their positions could therefore unfold without a single trace of any prior agreement.
This is precisely the problem that also concerns financial technology experts. All it takes is several models that may operate independently but were developed or trained on similar foundations. This undermines the established methods of regulators and investigators because, in such a case, there is no one to serve with a subpoena, no traceable evidence of conspiracy, and no one to accuse.
A kill switch that cannot be pressed
Over the decades, financial markets have developed safeguards for periods of extreme price fluctuations. Automatic trading halts have so far served to stop flash crashes, give people time to analyze the situation, and allow more cautious investors to help stabilize the market. In a world where machines execute most trades, however, the same safeguard would do more harm than good.
A sudden system shutdown would interrupt hedging positions halfway through the process, trigger an avalanche of margin calls, and drain the market of liquidity, much of which is supplied by algorithmic market makers. Once the financial world becomes fully dependent on automated trading, such an emergency stop may be nothing more than an unrealistic wish. Human traders can no longer replace algorithms, just as cash could not replace electronic payments if they were completely shut down.
When an entire industry depends on three models
Another vulnerability arises in the very foundations on which banks build their tools. If models are supplied by a large number of competing companies, users can switch to another provider when one falters, thereby increasing the resilience of both the financial sector and the real economy. But when the development and commercialization of this technology are concentrated in the hands of one or two companies, a critical point of failure emerges through which prices, terms, and outages can spread across the entire system at once. Such providers can then dictate terms to the rest of the economy, while a flaw in a widely used model will affect all users at the same moment.
Experts compare this situation to banks’ dependence on core banking systems and cloud service providers. However, the entire industry’s dependence on fewer than three models would be considerably riskier because a potential flaw would affect every institution in the same second, and no oversight procedure for such an event currently exists.
Professor Brunnermeier therefore recommends that artificial intelligence models become widely available commodities, ideally before market power becomes too concentrated. Regulatory rules should prevent the products of individual companies from differing fundamentally from one another and keep the costs of switching between versions and providers low. According to him, the option to revert to an older generation of a model or switch to a competitor must remain permanently open.
Fraud that costs as little as the truth
Fake news, misleading advice, and fraudulent products can be produced just as cheaply as genuine ones. The same technology also powers cyberattacks, phishing, and fabricated evidence. A malicious prompt inserted into a conversation or an ordinary model error can turn into an unauthorized trade or money transfer. The risk is that model providers will gain the ability to shape financial products in ways that no one on the outside can see.
The proposed solution—having artificial intelligence outputs checked by another AI system—opens up another vulnerability. If these monitors are nearly identical copies built on the same foundation model, they will make the same mistakes and overlook the same manipulations. A hundred nearly identical monitors are then like a single monitor voting a hundred times. Moreover, oversight operates at machine speed and exclusively between machines, making tacit coordination easier to maintain and harder to detect than among human auditors. The question also remains of who will audit the auditors themselves.
Professor Brunnermeier acknowledged that artificial intelligence can also bring better risk management and more accurate oversight than humans could provide. However, he emphasized the threats because, in his view, detecting them early is the way to maintain stability and avoid a system collapse. Banking regulators should therefore have the opportunity to assess major new models before they are released to the public. In June, the U.S. president signed an executive order asking major model developers to voluntarily allow the government to review new systems thirty days before they are launched.
Sources: finance.yahoo.com and americanbanker.com
