When Wall Street strategists rely on the same AI system: trading becomes more crowded, risks become more concentrated, and errors spread faster.

Wall Street's accelerating adoption of artificial intelligence (AI) is reshaping the microstructure of the market, sparking new concerns in the industry about increasingly crowded trading days, systems being easily deceived, and uncontrollable risk exposure. As hedge funds and wealth management firms compete to adopt similar AI models and datasets to find investment advantages, the divergence of views among market participants is shrinking. Recent research indicates that this trend towards algorithm homogenization is leading to highly similar investment portfolios and significantly shortening the lifespan of profitable trading signals, directly threatening the ability of active fund managers to achieve excess returns. At the same time, AI-driven trading systems are revealing significant vulnerabilities and blind spots. Multiple tests show that these models not only make mistakes when faced with subtly manipulated financial information, leading to significant single-day net asset value drawdowns, but also systematically assume risk volatility far exceeding expectations. These findings mark a shift in focus in the financial industry's discussion of AI. The core of market attention has shifted from "can technology help investors beat the market" to "how will market structure evolve when a large number of investors rely on the same machines." Homogenized strategies shorten profit cycles The efficiency of financial markets is built on investor divergence, but the widespread use of AI is breaking this premise. Researchers Shuchen Meng and Xupeng Chen from New York University found that as AI becomes more widespread in the investment industry, the similarity of investment portfolios is constantly increasing, a trend that is particularly pronounced in institutions heavily using this technology. This homogenization has a direct impact on market structure. The research model shows that before the widespread use of AI, a profitable trading signal could last five to seven years, but now its excess returns will halve within about 18 months. As more and more investors arrive at the same conclusions almost simultaneously, today's profit strategies will quickly become tomorrow's crowded trades. "Each additional marginal AI participant is shortening the lifespan of all exploitable patterns at an increasing rate," Meng and Chen noted in their paper, "When everyone is using similar AI, the collective outcome will be different in nature from the sum of individual interests." Buy-side institutions' reliance on AI continues to deepen. A survey conducted by the Alternative Investment Management Association (AIMA) last year showed that 58% of fund managers expect to use more AI in portfolio construction, up from 20% two years ago. Information manipulation exposes system vulnerabilities In addition to exacerbating crowded trades, AI models' reliance on input information also introduces new single point of failure risks. Researchers Advije Rizvani, Giovanni Apruzzese, and Pavel Laskov from the University of Liechtenstein designed ten trading models based on large language models (LLM) to predict stock prices through sentiment analysis. While these models all generated positive returns over a 14-month investment period, they were all vulnerable to manipulated information. Researchers made imperceptible modifications to financial news headlines that human readers would struggle to detect, such as replacing visually similar letters or embedding hidden text, and all models were successfully deceived. In the worst case, manipulation targeting a single stock on a single day led to a decrease of approximately 18 percentage points in the models' overall return. "A wrong decision can propagate to other days and affect other decisions the system is making," Rizvani said, "Even just one day can have very disastrous consequences." Uncontrollable risk exposure and overconfidence While AI inherits human traders' analytical abilities, it also appears to inherit humanity's oldest weakness: overassuming risk. Jerry Bell, Victor Haghani, and James White from Elm Partners Management tested four popular AI models in a simulated trading challenge. After reading the front page of The Wall Street Journal, Claude and ChatGPT achieved accuracy rates of over 50% in predicting the direction of the S&P 500 index and the US Treasury market, comparable to top macro traders' performance. However, the experiment revealed a critical flaw. All four models continuously assumed too much risk, with daily return volatility ranging from 20% to 40%, significantly higher than the recommended risk range of 7% to 15% set for them. "We trained AI like humans, and then we found that AI is indeed like humans overconfident, building too large positions," Haghani commented. As financial AI research moves from "can machines compete with humans" to a new stage of "how investors compete through the same machines," market participants need to reassess the true costs of technology. As Apruzzese warned, "Blindly trusting large language models to make wise decisions is unwise. If everyone adopts AI just because they think it's good and can help them make money without considering the consequences, they may face significant risks of loss." This article was translated from "Wall Street Horizon" by GMTEight Editor: Chen Siyu.