AI and Market Herding: What Regulators Have Found

Updated October 2026. First published April 2024. The original was a broad survey of AI; this version takes the one question from it that regulators have studied since: whether AI makes markets herd.

The 2024 post asked what happens if everyone trades with the same kind of AI. If every model predicts a stock will fall, who's left to buy it? Central banks and market regulators have since written about that question at length. They mostly agree the risk is real in principle. None of them has shown that it has happened.

The Argument

The plainest early statement came from Gary Gensler, then SEC Chair, in a July 17, 2023 speech at the National Press Club. He said AI "could promote herding with individual actors making similar decisions" because they were "getting the same signal from a base model or data aggregator." He went further: "AI may play a central role in the after-action reports of a future financial crisis." He and Lily Bailey had made the same case in a 2020 paper, "Deep Learning and Financial Stability," which the FSB cites.

The Financial Stability Board laid it out in more detail in its November 14, 2024 report. It names four AI-related vulnerabilities that could increase systemic risk: third-party dependencies and service provider concentration, market correlations, cyber risk, and a combined category of model risk, data quality, and governance.

On correlations, the FSB gives four reasons firms end up with the same models and data. They copy each other's choices. A third-party model can improve as more firms use it, which pulls in more users. Only a few models perform well enough to choose from. And vendors don't always disclose their training data, so firms can rely on the same sources without knowing it. Correlated AI in trading and portfolio management, the report says, "can amplify volatility, exacerbate liquidity crunches during downturns, and increase the probability of flash crashes."

The report is more specific about large language models. Most of them share an architecture and much of the same web-crawl training data, so firms that build on them may get correlated outputs. Whether that matters depends on the use. If firms use those models mainly for drafting, training, and finding information, "the impact on market correlations is likely to be minimal." The risk grows when the same models start setting investment strategy or risk limits. The FSB adds that automation can make it worse. Many firms pulling the "kill switches" on their automated systems at the same moment could deepen a stress event.

Others reached similar views. The IMF's October 2024 Global Financial Stability Report expects AI adoption in capital markets to "increase significantly in the near future," and lists "increased market speed and volatility under stress" and "reliance on a few key third-party AI-service providers" among the risks. It found that existing frameworks already address many of them. The Bank of England's Financial Policy Committee warned in April 2025 that more advanced AI trading strategies could lead to "firms taking increasingly correlated positions and acting in a similar way during a stress."

What the Evidence Shows So Far

Mostly, it shows adoption. In the 2024 AI survey by the Bank of England and the FCA, two of the most common current or planned uses were optimizing internal processes, at 72% of firms, and fraud detection, at 64%. Neither is a trading decision. The FSB's October 2025 follow-up focused on how to monitor AI adoption and its vulnerabilities, and asked authorities to close their data gaps. Its first report had already said that limited data on AI use makes these vulnerabilities hard to measure.

So in October 2026, herding driven by shared AI models is a plausible risk that regulators are watching, with no documented episode behind it. The nearer risk is the one every report lists: many firms depending on the same few providers for models, cloud capacity, and data. For a single firm, that's an operational risk, and you can work on it now.

Controls That Already Apply

A broker-dealer running an AI-driven strategy doesn't need a new rule to be responsible for it.

The SEC's Market Access Rule, Rule 15c3-5, requires a firm with market access to block orders that exceed preset credit or capital thresholds and to reject erroneous or duplicative orders before they reach the market. The CEO certifies the controls every year. An order a model generates goes through the same checks as one a trader enters.
FINRA's Regulatory Notice 15-09 on algorithmic trading expects firms to supervise a strategy from development on, test it independently of the people who wrote it, keep a plain-language description of each strategy, and keep a way to "quickly disable the algorithm or supporting platform." It also says supervision continues "after any algorithmic strategy is put into production." A model that retrains itself keeps changing after production, which makes that line harder to meet.
Provider concentration is a vendor question. FINRA's Regulatory Notice 21-29 leaves the firm responsible for whatever it outsources, including the model inside a vendor's product. Treasury's 2024 interviews found that vendors often get that model from yet another vendor.

One step costs almost nothing. Ask your strategy owners and your vendors which base models and data sources they rely on, and write the answers down in one place. If the same two names appear everywhere, you've found your firm's own correlation risk.

This post is for informational purposes only and does not constitute legal advice. For guidance on specific regulatory obligations, consult your counsel or compliance advisor.