Artificial intelligence has moved from the margins of finance to its core. Trading desks use it to price assets and execute orders. Banks use it to score credit, flag fraud, and screen transactions. Insurers use it to underwrite. Asset managers use it to build portfolios and summarize research. The promise is efficiency: faster decisions, lower costs, fewer human errors. The worry, increasingly voiced by regulators and researchers alike, is that the same technology could make the financial system more fragile in ways that are difficult to see until something breaks.
The problem of everyone doing the same thing
The most frequently cited danger is correlation. Financial crises tend to happen not because one institution makes a bad bet, but because many institutions make the same bad bet at the same time and then try to unwind it simultaneously. AI could intensify that dynamic. A relatively small number of foundation models, data vendors, and cloud providers now sit beneath a wide range of financial applications. If firms are training on similar data, using similar models, and receiving similar signals, their behavior may converge u2014 buying the same assets in calm markets and selling them in unison when conditions turn.
Speed compounds the issue. Automated systems can act in milliseconds, far faster than humans can intervene. The “flash crash” episodes of the algorithmic trading era offered a preview of how quickly liquidity can evaporate when machines withdraw at once. AI systems that adapt on the fly, rather than following fixed rules, are harder to predict and harder to reason about after the fact.
Opacity and accountability
A second concern is explainability. Traditional risk models, whatever their flaws, could usually be interrogated: a supervisor could ask why a loan was denied or why a position was sized a certain way. Many modern AI systems do not offer clean answers. That creates practical problems for compliance u2014 fair-lending rules, for instance, assume decisions can be justified u2014 and deeper problems for risk management. If a bank’s leadership cannot articulate why its models behave as they do, it is difficult for them to know when those models are wrong.
There is also the matter of who is accountable when an automated system fails. Responsibility may be spread across a model developer, a data supplier, a cloud host, and the financial firm that deployed the tool. Diffuse responsibility tends to produce gaps rather than overlapping safety nets.
Concentration and contagion
The technology supply chain itself is a source of systemic risk. Financial regulators have long worried about the industry’s reliance on a handful of cloud providers. AI deepens that dependency. An outage, a corrupted model update, or a successful attack on a widely used system could affect many institutions at once u2014 a form of contagion that does not travel through balance sheets but through infrastructure.
Meanwhile, AI is also a tool for bad actors. It lowers the cost of convincing fraud, synthetic identities, and market manipulation, forcing defenders into an arms race with attackers who face fewer constraints.
What comes next
None of this argues for keeping AI out of finance, which is no longer a realistic option. It argues for treating AI as infrastructure rather than as a product feature: stress-testing models for correlated failure, requiring human circuit breakers, mapping dependencies on third-party providers, and giving supervisors the technical capacity to inspect systems they did not build.
Financial history suggests that new technologies are usually absorbed smoothly until, abruptly, they are not. The value of asking these questions now is that the answers are cheaper before a crisis than after one. Read More
Leave a Reply