AI and Capital Markets── Algorithms execute 60–70% of U.S. equity trades, and dependence on the same models is creating a new kind of instability

01Algorithms now move 60–70% of equity trades

In the United States, 60–73% of equity trading volume is executed by algorithms. A 2020 SEC staff report acknowledged that algorithmic trading, including high-frequency trading, has become central to market structure. The ratio is climbing in Europe and Asia as well; on the Tokyo Stock Exchange, roughly 70% of orders are placed automatically.

These numbers mean algorithmic trading is no longer a specialized technique—it is the market's default mode. The question is whether that default ensures stability or undermines it.

02On May 6, 2010, roughly $1 trillion vanished in 36 minutes

At 2:32 p.m. EDT on May 6, 2010, the Dow Jones Industrial Average fell 998.5 points—about 9%. Roughly $1 trillion in market capitalization disappeared in 36 minutes, most of it within just a few minutes. The joint SEC–CFTC report published on September 30, 2010, identified the immediate trigger: a large institutional investor had routed an order to sell approximately $4.1 billion worth of E-mini S&P 500 futures—75,000 contracts—through an automated execution algorithm.

High-frequency traders initially absorbed the selling pressure, but within minutes liquidity evaporated. The report described a "complete evaporation of liquidity in the marketplace." More than 20,000 trades were executed at prices more than 60% away from levels just moments earlier. Many of those trades involved retail-customer accounts.

03Post-crash reforms addressed speed, not structure

In response to the Flash Crash, the SEC approved two measures in 2012. The first was the Limit Up-Limit Down (LULD) plan, which temporarily halts trading in an individual security when its price moves beyond a set band—5% for S&P 500 constituents, 10% for most other stocks—relative to its average price over the preceding five minutes. It was made permanent in 2019. The second was a revision of market-wide circuit breakers, now set to halt trading when the S&P 500 falls 7%, 13%, or 20% from the previous close. During the COVID-19 sell-off in March 2020, the Level 1 breaker at 7% was triggered four times.

These mechanisms limited the depth of sudden declines. But they did not answer the deeper questions: why does liquidity vanish all at once, and why do multiple participants move in the same direction simultaneously?

04The IMF identified dependence on the same AI models as a herding catalyst

In October 2024, the IMF published Chapter 3 of its Global Financial Stability Report, titled "Advances in Artificial Intelligence: Implications for Capital Market Activities." The report named herding and market concentration as the most significant risks arising from the wider adoption of AI in capital markets.

One concrete finding stands out. AI-driven ETFs exhibit significantly higher portfolio turnover than conventional actively managed ETFs. While a typical active equity ETF turns over its holdings less than once a year, AI-driven ETFs do so roughly once a month. During the March 2020 market turmoil, several AI-driven ETFs increased their turnover further, showing a tendency toward synchronized selling under stress.

The mechanism is straightforward. When many managers use the same foundation model, the same training data, and the same feature sets, their trading signals converge. In calm markets the outputs may appear diverse, but when a stress signal fires, models trained on similar data respond in the same direction at the same time.

05Pharmaceutical stocks move on clinical trial results—AI has compressed the reaction time

The risks of algorithmic trading are directly relevant to the pharmaceutical and healthcare sectors. A 2023 study published in Scientific Reports built a machine-learning framework to predict stock-price reactions to 5,436 FDA clinical trial announcements from 681 companies between 2018 and 2022. The framework combined BERT-based sentiment analysis, a Temporal Fusion Transformer for return forecasting, a graph convolution network for event-relationship extraction, and gradient boosting for price-change prediction.

What the study illustrates is how AI has shortened the time between a trial result becoming public and the market's response. Analyses that once took human analysts hours are now completed by NLP models in seconds, and algorithms execute trades immediately. As the window between disclosure and price reaction shrinks, the space for human judgment narrows correspondingly.

06Model concentration stems from data oligopolies and foundation-model dominance

The root of herding behavior lies not in model design choices but in infrastructure.

First, data supply is concentrated. Financial-market data providers are few, and the IMF report characterizes this as a "data oligopoly." Models trained on the same data produce similar outputs regardless of who built them.

Second, foundation models are concentrated. Development of large language models and deep-learning frameworks is dominated by a small number of firms. The FSB's November 2024 report flagged third-party dependencies and service-provider concentration as key financial-stability vulnerabilities. Its October 2025 follow-up maintained these concerns.

Third, success breeds imitation. Strategies that deliver high returns are published, replicated, and scaled, causing the market as a whole to tilt in one direction. The IMF warned of a "winner takes all" scenario in which a small number of dominant platforms intensify this concentration.

07Practical responses: preserving diversity and governing speed

Responses are already in motion on both the regulatory and the operational sides.

On the regulatory front, circuit breakers and LULD continue to be refined. The Bank of England's December 2025 Financial Stability Report flagged the risk that AI-driven trading could amplify volatility and called for stronger surveillance. The IMF issued a 2025 technical note outlining regulatory considerations for accelerated AI use in securities markets.

On the operational side, ensuring model diversity is now a priority. Even when using the same foundation model, varying training data, feature selection, and risk-management thresholds can reduce output correlation. Some managers run multiple models in parallel and execute large trades only when the models agree—a built-in check against synchronized errors.

Stress testing is evolving too. Traditional tests assume "the market drops 20%." To account for model-synchronization risk, a new premise is needed: "What happens if every manager running the same model sells at the same time?"

08Efficiency and stability are separate questions

Algorithmic trading has made markets more efficient. Spreads have narrowed, execution costs have fallen, and price discovery has accelerated. But efficiency and stability are distinct questions. The 2010 Flash Crash demonstrated that an efficient market can shed roughly $1 trillion in market value in 36 minutes.

The introduction of AI has not changed this dynamic. It is reproducing the same category of risk through a new channel: dependence on the same models. The herding risk the IMF identified in 2024 operates on fundamentally the same mechanics as the liquidity evaporation of 2010. The difference is that speed has increased further and the synchronization of decisions has deepened.

Institutions are trying to catch up to the speed of technology. Circuit breakers are a retrospective safety valve; they cannot address the structural risk of model concentration. What the FSB and IMF are calling for is not speed limits alone but transparency and diversity in model use. The stability of capital markets will be preserved not by slowing technology down but by governing the homogeneity that technology creates.