Artificial intelligence is reshaping how markets move, how trades are placed, and how fortunes are won and lost. But the deeper you look, the clearer a troubling pattern becomes: the system is only as good as the humans who built it — and markets have a long history of punishing that kind of confidence.

April 2026  ·  Analysis

On the afternoon of May 6, 2010, something strange happened on Wall Street. In the space of about twenty minutes, the Dow Jones Industrial Average fell nearly 1,000 points — roughly a trillion dollars in market value — and then, almost as inexplicably, recovered most of it. No war had broken out. No major company had collapsed. The culprit, investigators later concluded, was a cascade of automated trading algorithms that had misread market conditions and begun selling in a self-reinforcing spiral, each system’s panic triggering another’s.¹ Human traders eventually intervened and stopped the bleeding. But for those twenty minutes, the machines had been in charge — and they had no idea what they were doing.

That episode, known as the Flash Crash, is more than a historical footnote. It is a preview of a world that is arriving faster than most people realize: one in which artificial intelligence doesn’t just assist financial decision-making, but increasingly drives it. And while the technology has advanced enormously since 2010, the fundamental tension it exposed has not gone away. If anything, it has deepened.

What AI actually does — and doesn’t do

The word “intelligence” in artificial intelligence does a lot of misleading work. When a trader or investor uses an AI system to scan for opportunities — flagging momentum plays, identifying undervalued stocks, parsing earnings reports — the system is not reasoning. It is pattern-matching. It has been trained on historical data and is identifying configurations that resembled profitable outcomes in the past. That is genuinely useful. But it is not the same as understanding why those patterns worked, whether they will continue to work, or when the conditions that made them reliable have quietly changed.

The distinction matters enormously in financial markets, where context is everything and the past is a notoriously unreliable guide to the future. A model trained primarily on the long bull market that followed the 2008 financial crisis has a built-in view of the world — one where certain relationships between asset classes, interest rates, and growth expectations held roughly true for over a decade. When those relationships broke down in 2022, as inflation surged and central banks raised rates at the fastest pace in a generation, models that had never experienced such a regime had no reliable framework for what was happening.

“A model trained on a bull market has never experienced a liquidity crisis. It has no framework for what it hasn’t seen.”

This is not a flaw that better engineering can simply patch away. It is a structural feature of how these systems are built. Nassim Nicholas Taleb, the former options trader and author of The Black Swan, spent years arguing that conventional financial models dangerously underestimate the probability of rare, extreme events. His core argument was that quantitative models assumed the world was more predictable than it actually is — and that this false confidence was not a minor inconvenience but an existential risk.² AI-driven trading systems inherit this problem and, in some ways, amplify it. They are extraordinarily good at finding patterns. They are not good at recognizing when pattern-finding is the wrong tool for the moment.

The data problem runs deeper than the internet

The most commonly cited concern about AI in financial markets is the quality of the data it trains on. The internet contains noise: misinformation, spin, coordinated manipulation, and low-quality analysis. Feed an AI on that, and the outputs will reflect it. This is true — particularly for systems that incorporate social media sentiment or news feeds, where bad actors can and do attempt to move markets through manufactured narratives.

But the data quality problem goes deeper than most people acknowledge. The issue is not simply that the internet contains bad information. It is that every layer of the AI development process is shaped by human judgment — and human judgment is fallible, biased, and context-dependent in ways that are extremely difficult to detect or correct.

Modern AI systems are not just trained on raw data. They are shaped through a process called Reinforcement Learning from Human Feedback, or RLHF, in which human evaluators rate model outputs and the system learns to produce responses that score well with those evaluators.⁴ This means the values, assumptions, and blind spots of the people doing the evaluating get baked into the model. A 2023 paper by researchers at MIT, Stanford, UC Berkeley, and Cambridge identified this as a fundamental limitation — the quality of RLHF is entirely dependent on the representativeness of the human feedback, and biases in that feedback propagate into the model in ways that are often invisible to anyone, including the model’s own builders.⁵

For a trading system, this matters in subtle but consequential ways. If the training data reflects a particular view of how markets work — shaped by a specific era, a specific culture of finance, a specific set of assumptions about risk — then the model encodes that view as a kind of ground truth. It will not question it. It will not flag it as an assumption. It will simply act on it, with the same confidence it brings to everything else.

“The system doesn’t inherit the wisdom of the humans who built it. It inherits their assumptions — including the ones they didn’t know they had.”

The automation problem: when speed becomes a liability

There is a seductive logic to full automation. Algorithms don’t panic. They don’t get greedy. They don’t override their own rules because a position feels uncomfortable. For strategies operating on sub-second timeframes, there is no human alternative — the edge exists precisely because no person can act that fast. These are real advantages, and serious quantitative traders have built significant businesses around them.

But automation amplifies errors at a scale and speed that human trading cannot match. A human trader who misreads a setup loses on one trade. An automated system that misreads a setup — or encounters market conditions it was not designed for — can execute that same error thousands of times before anyone intervenes. The Flash Crash made this vivid: the market recovered only when human agents stepped in.¹ The machines, left to themselves, had no off switch.

This risk has not gone unnoticed at the regulatory level. Gary Gensler, the former chair of the U.S. Securities and Exchange Commission, spent several years warning publicly that AI in financial markets poses a systemic threat that regulators are not adequately prepared for. In a 2020 paper co-written while he was a professor at MIT, Gensler argued that widespread AI adoption in trading was likely to produce “monoculture” effects: situations where large numbers of institutions rely on the same underlying models, trained on the same data, making similar decisions at the same moment.⁶ “Broad adoption of deep learning,” he wrote, “may increase uniformity, interconnectedness, and regulatory gaps, leaving the financial system more fragile.” In a 2023 interview with Bloomberg, he went further: “This technology will be at the center of future financial crises.”

The concern Gensler and others have raised is not that any single AI system will fail catastrophically. It is that the aggregated behavior of many AI systems — trained similarly, responding to the same signals, operating without meaningful human oversight at the moment of decision — could produce the kind of synchronized, self-reinforcing crisis that is much harder to stop than the failure of any individual firm.

What considered practice actually looks like

None of this means AI has no place in trading. Used carefully, it is a genuinely powerful tool — exceptional at processing large volumes of data, identifying statistical regularities, and executing rules-based decisions without the emotional interference that derails human traders. The point is not to avoid AI. The point is to understand precisely what it is, and what it is not.

Practitioners who use it most effectively tend to keep their data inputs tightly controlled, working from curated proprietary datasets rather than internet-connected feeds where noise and manipulation are harder to filter. They treat AI outputs as a starting point for analysis rather than a conclusion — using the system to surface candidates and flag anomalies, while retaining human judgment about whether current market conditions actually match the historical context the model was trained on. And they are cautious about full automation, particularly for decisions where errors compound quickly: position sizing, risk management, and any scenario where a bad signal, executed at scale, could do lasting damage.

That caution is not sentimental. It is structural. Markets are adaptive, adversarial systems. Every edge that becomes widely known — or widely systematized — tends to erode. An AI system that was profitable in one regime may be silently losing ground in another, and it will not tell you. It will keep executing with the same confidence, against conditions it was never designed to understand, until the losses make the problem undeniable.

“The machine executes on the world it was trained on. The trader has to know when that world no longer exists.”

The deeper insight — the one the Flash Crash demonstrated, that Taleb has spent a career elaborating, and that Gensler spent years warning regulators about — is that the risk is not simply in any individual system. It is in the collective assumption that these systems are more reliable than they are. That assumption, held widely enough, across institutions large enough, with enough capital behind it, is how a manageable problem becomes a crisis.

The machine doesn’t know what it doesn’t know. That has always been true of financial models. What is new is the speed at which that ignorance can now move through a market, and the number of people who, not fully understanding what they are using, are prepared to let it.

References & further reading

1. U.S. CFTC & SEC joint report, Findings Regarding the Market Events of May 6, 2010 (September 2010). The official regulatory account of the Flash Crash and its causes.

2. Nassim Nicholas Taleb, The Black Swan: The Impact of the Highly Improbable, Random House (2007). Taleb’s core argument that conventional risk models systematically underestimate rare, high-impact events. His earlier book Fooled by Randomness (2001) applies similar arguments specifically to financial markets.

3. Ji et al. (2023); Bang et al. (2023), as cited in: Comprehensive Review of AI Hallucinations: Impacts and Mitigation Strategies for Financial and Business Applications, Preprints.org (May 2025).

4. The foundational paper on RLHF: Christiano et al., Deep Reinforcement Learning from Human Preferences, NeurIPS (2017).

5. Casper et al., Open Problems and Fundamental Limitations of Reinforcement Learning from Human Feedback, arXiv:2307.15217 (2023). Authors include researchers from MIT CSAIL, Stanford, UC Berkeley, and the University of Cambridge.

6. Gary Gensler & Lily Bailey, Deep Learning and Financial Stability, MIT (2020); Gensler public remarks on systemic AI risk, SEC.gov (September 2024); Gensler interview, Bloomberg Businessweek (August 7, 2023).

7. Algorithms Probably Caused a Flash Crash of the British Pound, MIT Technology Review (October 7, 2016).

8. Greg Jensen (Bridgewater Associates), quoted in Fortune (2023): using large language models to pick stocks is “a hopeless path.”