Automated trading crashed the market once, in 2010, with no AI involved

How automated selling can feed on itself

Financial markets now run at a speed no person can follow. Much of the trading in US stocks and stock-index futures is done by high-frequency trading: automated programs that place and cancel orders in fractions of a second. These programs supply much of the market's liquidity, the standing orders to buy and sell that let a trade happen at a steady price. The worry is about what happens when many of them do the same thing in the same instant.

The reasoning runs like this. Automated strategies are often built from similar ideas and fed similar data, so they tend to react alike. When a shock arrives, correlated programs can all sell, or all pull their orders, in the same second. The traders who normally cushion a fall step back to protect themselves. Liquidity thins, each sale pushes the price down further, and that fall triggers the next round of selling. Because the loop closes in milliseconds, no human can lean against it in time. The economists who studied the day most closely call this a demand for immediacy. In a stressed market, automated traders compete to get out first, and that competition costs everyone slower than them.¹ The International Monetary Fund frames the same mechanism as increased market speed and volatility under stress, especially when many strategies respond to a shock in a similar way or shut down at once.²

This is a specific worry, and it is worth separating from two the course has already covered. It is not automated traders quietly learning to hold prices high, the subject of algorithmic collusion. It is not many firms making the same biased decision from the same model, the subject of algorithmic monoculture. The claim here is narrower and faster: speed plus sameness can make a market fall out from under itself in minutes.

Inside the twenty minutes the market broke

It has happened once, clearly, on May 6, 2010. The day was already tense. Between 9:30 a.m. and 2:00 p.m., with Europe's debt crisis in the news, the Dow Jones Industrial Average drifted down about 1.5 percent, to 10,712.³ Then, at 2:32 p.m., a single mutual fund began to sell. It used an automated execution program to sell 75,000 E-mini S&P 500 futures contracts, worth about $4.1 billion, as a hedge. The program was told to sell at a rate of 9 percent of recent trading volume, without regard to price or time.⁴

In a calm market that order might have taken hours. On a stressed one it ran in about twenty minutes.⁴ High-frequency traders bought the first waves, then, to stay flat, resold to each other. The official report calls it a hot-potato effect. In the fourteen seconds between 2:45:13 and 2:45:27, these traders passed more than 27,000 contracts back and forth, about 49 percent of all volume, while adding only about 200 contracts on net.⁴ Volume looked enormous while real buyers had vanished. In the four and a half minutes to 2:45:27, the E-mini fell more than 5 percent and the SPY exchange-traded fund fell more than 6 percent.⁴ The major indices, already down more than 4 percent for the day, plummeted a further 5 to 6 percent in minutes.⁴

Then it reversed almost as fast. Buyers returned, and by about 3:00 p.m. most stocks were trading at sane prices again. The indices closed down about 3 percent.⁴ The real damage was in the gap. More than 20,000 trades across 300-plus securities executed at prices 60 percent or more from where they had stood moments earlier. Some printed as low as a penny, some as high as $100,000, before the exchanges canceled them.⁴

The one real crash ran on rule-based code

Nothing in that account was artificial intelligence. The sell program followed one fixed instruction, match 9 percent of volume and ignore price, the kind of rule you could write on an index card.⁴ The high-frequency traders ran fixed strategies too. The most careful study of the day, an audit-trail analysis by Andrei Kirilenko and colleagues later published in the Journal of Finance, is blunt about their role. High-frequency traders did not cause the Flash Crash, it finds, but contributed to it by demanding immediacy ahead of other market participants.¹ A large enough sell order can set off a liquidity-based crash on its own.¹

There is a second actor, added years later. In 2015 the Commodity Futures Trading Commission charged a London trader, Navinder Sarao, with manipulation and spoofing. For more than two hours he ran a layering program that placed large sell orders he meant to cancel, applying close to $200 million of downward pressure. That pressure, the agency said, contributed to the order-book imbalance and the conditions that led to the crash.⁵ Notice the care in that wording. The 2010 report did not name him, and no account claims one trader caused the day. Whatever weight you give the spoofer, his tools were rule-based too. The one crash everyone points to ran on plain automated code, built and instructed by people, years before a language model could hold a conversation.

The AI version of the worry lives in simulations

So why worry about AI at all? Because the machines doing the trading are changing. The fear is that trading agents built on machine learning, and increasingly on large language models, could be far more alike than today's hand-written strategies. Trained on similar data and reasoning in similar ways, they might crowd into the same positions and unwind them together, reproducing 2010's correlated withdrawal with even less human oversight.

The strongest evidence for this is a simulation, not a market. In a 2026 study, researchers including the MIT economist Andrew Lo ran language-model agents as traders and found that better models can make the system worse. Frontier models behaved more alike as they grew more capable. When the agents shared a common misinformation environment, that shared behavior became a liability.⁶ It is a careful result and an artificial one, run in an agent-based simulation rather than a live exchange.⁶ Regulators voice the same mechanism. The Bank of England warns that AI-driven strategies could push firms toward correlated positions through similar model designs, and that advanced models could exploit opportunities in a destabilising way.⁷ The IMF reports that market participants rank herding and concentration among their top AI worries, and lists flash crashes among the failures that faster, guardrail-thin decisions could bring.² What none of them can point to is a crash on record caused by AI trading agents. That event has not happened.

What would settle whether AI makes crashes worse

The people who press the argument today are mostly central banks and the IMF, along with some of the researchers building the agents, and their asks are concrete. The IMF wants regulators to recalibrate circuit breakers and review margin practices for the possibility of rapid, AI-driven price moves.² The Bank of England wants oversight extended to agentic AI, systems that take actions on their own to reach a goal.⁷ Some of these tools already exist. After 2010 the SEC adopted a limit up-limit down mechanism that blocks trades outside a set price band. It also updated the market-wide circuit breakers, which never triggered on May 6.⁸

The honest picture is that confidence outruns proof on both sides. Anyone certain AI will cause the next crash is reasoning from a single precedent that ran on rule-based code, plus simulations. Anyone certain it will not is ignoring that the 2010 mechanism was real and that automation has only deepened since. The IMF makes the two-sided case itself. The same report says AI could reduce risk by improving risk management and deepening liquidity, even as it could sharpen sell-offs.²

What would settle it is the kind of evidence the record does not yet hold: a market break traceable to machine-learning agents, or large multi-agent studies run at something like real market scale. Neither exists yet.