By Amir Naser Hojati
Individually prudent AI trading systems can become collectively destabilising when they respond to the same signals, making behavioural concentration a new systemic risk.
Europe’s next AI-market risk may not come from a defective model. It may come from many well-designed models reacting to the same information at the same time. After more than 15 years of trading futures markets, I have learned that markets often absorb disagreement better than synchronisation. The practical implication is straightforward: regulators and financial firms should supplement firm-level model controls with system-level tests that measure how automated strategies behave together when volatility rises, and liquidity disappears.
How Can Safe Algorithms Create an Unsafe Market?
Financial regulation understandably begins with the individual institution. Was the algorithm properly tested? Are its risk limits appropriate? Can a human intervene? Does it behave as intended?
Those questions remain essential, and Europe has strengthened the framework around them. The EU AI Act entered a new enforcement phase on 2 August 2026, while ESMA’s February 2026 supervisory briefing on algorithmic trading focused on governance, testing, pre-trade controls and outsourcing, and explicitly addressed the expanding use of AI.
But markets do not experience algorithms one at a time. They experience their combined orders.
Consider several large institutions using independently developed systems. Their code, models and vendors may differ, yet the systems observe many of the same variables: volatility, liquidity, momentum, correlations, macroeconomic releases and order-book conditions. A shock arrives. Volatility rises, market depth deteriorates and each system decides that reducing exposure is prudent.
Individually, each decision can be rational. Collectively, the decisions can reinforce one another. Selling consumes liquidity; thinner liquidity increases price impact and volatility; higher volatility triggers further deleveraging. No algorithm necessarily malfunctions, but the market-level outcome becomes unstable.
That is the gap between model safety and system safety. A firm can pass its internal controls while still participating in a destabilising feedback loop that no single institution intended.
Why Could AI Intensify Behavioural Concentration?
Crowded trades are not new. Human traders copy one another, fund managers use similar risk frameworks and investors often move toward the same exit during stress. Yet human decision-making also introduces friction: one trader hesitates, another interprets the signal differently, and another may provide liquidity because the move looks excessive.
Automation can reduce some of that diversity. Models trained on overlapping data and optimised for similar objectives can discover similar relationships without sharing source code or communicating with one another. Correlation can emerge from common inputs and common incentives.
The 2010 Flash Crash offers a useful non-AI precedent. The joint CFTC-SEC review found that an automated sell programme interacted with algorithmic trading as liquidity thinned, while many automated liquidity providers temporarily paused. The episode did not involve modern AI, but it demonstrated how automated responses can interact with stressed liquidity and transmit instability across linked markets.
Recent ECB research makes the concern more concrete. In simulation-based experiments, researchers compared different AI architectures in a stylised redemption setting. Q-learning agents showed a high degree of coordination and could produce extreme, bank-run-like dynamics under stress, while large language model agents behaved more heterogeneously. The result does not prove that a market crisis will be caused by AI, but it does show that architecture and interaction can materially change system-level outcomes.
For executives and risk committees, the lesson is practical: model diversity should not be judged only by whether firms bought different software or trained different models. The more useful question is whether systems respond differently when the same market variables deteriorate.
What Should a Market-Wide Stress Test Measure?
Europe already uses stress testing to ask how institutions behave under adverse conditions. The same principle can be extended to increasingly automated markets without requiring firms to reveal proprietary source code or trading logic.
ESMA, national supervisors and the European Systemic Risk Board could define common scenarios and ask firms to evaluate how their automated systems would respond. The ESRB is already responsible for macroprudential oversight and systemic-risk monitoring, making it a natural partner for exercises that move beyond the firm-by-firm view.
A useful test could include scenarios such as:
- a sudden volatility spike combined with a severe reduction in displayed market depth;
- an unexpected central-bank decision that breaks a previously stable correlation;
- a rapid widening of spreads across several related instruments;
- a sharp increase in margin or internal risk limits that forces simultaneous deleveraging.
The objective would not be to predict the next crisis. It would be to measure response concentration: how many systems attempt to sell, hedge or withdraw liquidity within the same short interval, and how sensitive that behaviour is to deteriorating liquidity.
This changes the supervisory question from “Did each firm’s controls operate correctly?” to “What happens if many correct controls produce the same action at once?”
What Can Regulators and Firms Do Now?
A full regulatory regime is not necessary before useful information can be collected. Europe could begin with anonymised, aggregated exercises and a small set of behavioural indicators.
Exhibit 1. From model safety to system safety
| Firm-level oversight | System-level complement |
| Model testing | Cross-firm response correlation |
| Pre-trade controls | Concentration of simultaneous orders |
| Risk limits | Common deleveraging triggers |
| Human intervention | Market-wide reaction speed |
| Vendor/model governance | Dependency on common data and signals |
Source: Author’s framework, informed by ESMA’s supervisory briefing and the ESRB’s systemic-risk mandate.
For regulators, three early priorities stand out:
- Measure behavioural concentration during stress, not only normal-market correlations.
- Track whether volatility, liquidity or margin thresholds trigger similar responses across institutions.
- Combine ESMA’s market-supervision perspective with the ESRB’s system-wide risk mandate.
Firms can act before any new rule arrives. Boards and risk committees should ask whether an automated strategy’s exit assumptions still hold when many peers are trying to exit simultaneously. Stress tests should therefore include not only adverse prices, but adverse liquidity and delayed execution. Firms should also examine whether supposedly independent strategies share the same hidden dependency—for example, a volatility filter, liquidity threshold or risk-off trigger.
For boards and risk committees, three questions are immediately useful:
- If our risk-off trigger fires, how many comparable systems are likely to react to the same signal?
- Does our stress testing assume liquidity will remain available precisely when we need to exit?
- Are our strategies genuinely behaviourally diverse, or merely different in code, vendor and architecture?
This is not an argument against AI in markets. AI can improve execution, identify patterns and strengthen risk management. The point is that a conservative model can still become part of an aggressive market-wide move when thousands of other conservative models reach the same conclusion at the same moment.
What Should Europe Do Next?
Europe’s AI and market rules are becoming more sophisticated at the level of the individual firm. The next step is to add a system-level lens. Supervisors do not need source code from every institution, nor do they need to forecast the next shock perfectly. They need to identify where individually prudent automated behaviour could become collectively destabilising. The safest model can still belong to a dangerous crowd. Europe already knows how to stress-test institutions; it should now learn how to stress-test their interactions.
Amir Naser Hojati