From hedge funds to wealth managers, Wall Street has embraced artificial intelligence in search of an investing edge. Researchers are now asking what happens when more investors turn to similar AI models to find one: buying the same stocks, reacting to the same headlines, and sometimes making the same mistakes.
Processing Content
The emerging research suggests that scenario could have market-wide consequences. AI may make investors faster and more informed while also making their trades more crowded, their systems easier to fool and their risk-taking harder to control. Several recent studies suggest widespread adoption could shorten the lifespan of profitable trading signals as investors converge on the same opportunities. Others find AI models systematically take more risk than intended or can be manipulated through the information they consume.
Taken together, the papers point to a shift in the debate around AI in finance, from whether the technology can help investors beat the market to how markets change when growing numbers of investors rely on the same machines.
READ MORE: AI is mostly ‘Moneyball’ for advisors — but it’s becoming ‘WarGames’
The concern stems from a simple premise: Markets work because investors disagree. Every day, portfolio managers read the same earnings release, jobs report or Federal Reserve statement and reach different conclusions about what to buy or sell. Artificial intelligence promises to help them process that information faster. But if thousands of firms increasingly rely on similar models trained on similar data, those differences may begin to narrow.
Researchers at New York University found evidence of that dynamic already emerging. Studying nearly one million institutional fund holdings, Shuchen Meng and Xupeng Chen found portfolios have become increasingly similar as AI adoption has spread across the investment industry. The trend was especially pronounced among firms making greater use of the technology.
The implications could extend beyond portfolio construction and into market structure. Their model suggests a profitable trading signal could lose half its excess return in about 18 months, down from five to seven years before AI became widespread. As more investors reach the same conclusion at roughly the same time, profitable ideas become crowded more quickly. For active managers, that means today’s winning strategy could become tomorrow’s crowded trade much sooner.
“Each marginal AI entrant shortens the lifespan of every exploitable pattern at an increasing rate,” Meng and Chen wrote in the paper titled AI-Driven Alpha Decay: Algorithmic Homogenization, Reflexive Signal Erosion, and the Paradox of Intelligent Markets. “When everyone uses similar AI, the collective outcome differs qualitatively from the sum of individual benefits.”
The findings are a warning for buy-side firms that have been increasingly tapping advanced AI in research and portfolio construction. In a survey by the Alternative Investment Management Association last year, 58% of fund managers said they expected to use AI more in their investment process, up from 20% two years earlier.
Other researchers argue AI could introduce new points of failure by making investment systems vulnerable to manipulated information.
READ MORE: LPL’s AI challenge: Moving fast without overwhelming advisors
At the University of Liechtenstein, Advije Rizvani, Giovanni Apruzzese and Pavel Laskov designed 10 LLM-based trading models to forecast share prices with sentiment analysis for a portfolio of stocks. All generated positive returns during a 14-month investment period through April 2025. Yet every model was fooled after researchers made subtle changes to financial news headlines that were barely noticeable to human readers, such as swapping letters for nearly identical-looking characters or embedding hidden text. In the worst case, a model’s overall return fell by roughly 18 percentage points after manipulation targeting a single stock on a single day.
“One wrong decision could propagate to other days and another decision that the systems are making,” said Rizvani. “Even just for one day, it could lead to a very catastrophic consequence.”
None of the studies proves that artificial intelligence will make markets more fragile. Most rely on simulations, controlled experiments or limited datasets, and the researchers themselves are careful about drawing broad conclusions. Crowded trades, especially in tech stocks, have been a hallmark of US markets since long before the rise of AI, without any catastrophic consequences.
Yet a common theme runs through the emerging literature: the same technology that promises to make investors more informed may also make it easier for crowded trades, bad information and overconfidence to spread through markets.
A third line of research points to a familiar problem. AI may inherit one of investors’ oldest weaknesses: taking too much risk.
Elm Partners Management’s Jerry Bell, Victor Haghani and James White tested four popular AI models on a simulated trading challenge, asking them to place bets on the S&P 500 and Treasury bonds after reading advance copies of Wall Street Journal front pages. Claude and ChatGPT matched the directional accuracy of elite macro traders, correctly calling market direction more than 50% of the time.
But the experiment exposed a critical blind spot. All four models consistently took on far too much risk, running daily return volatility of 20% to 40% against a recommended range of roughly 7% to 15% for the investor profile they were given.
“We train the AIs to be like people, and then we’re finding that the AIs are like people — being overconfident, taking too big positions,” Haghani said.
READ MORE: The AI tools meant to save time are burning out advisors
The first generation of AI research in finance asked whether machines could compete with investors. The next generation is asking what happens when investors increasingly compete through the same machines. Every investing breakthrough promises an edge over everyone else. The emerging question is what happens when everyone has access to the same one.
“Blindly trusting large language models to make sound decisions would be unwise,” said Apruzzese at the University of Liechtenstein. “If everybody just takes AI because they think it’s good and helps them make more money, and without thinking about the consequences, they may risk losing a lot.”
