列支敦士登大学的研究人员设计了十个基于大型语言模型的交易系统，发现所有模型都能产生正回报，但在研究人员对财经新闻标题进行细微篡改后——例如替换相似字符或嵌入隐藏文字——每个模型都被成功欺骗。最严重的情况下，仅针对单一股票单日的操纵就导致模型整体回报下降约18个百分点，揭示了AI驱动投资系统面对资讯操纵时的脆弱性。
第三条研究路线指出，AI可能继承投资者最古老的弱点：承担过多风险。Elm Partners的测试显示，四个主流AI模型虽然在判断市场方向上达到了精英宏观交易员的水准，但它们持续承担远超建议范围的风险，日回报波动率达到20%至40%，而建议范围仅为7%至15%。研究者警告，盲目信任大型语言模型做出投资决策是不明智的，若所有人都不加思考地使用AI，可能面临巨大损失。
Wall Street has widely adopted artificial intelligence to gain an investing edge, but researchers are now examining the systemic consequences of widespread AI use. A New York University study found that institutional portfolios have grown increasingly similar as AI adoption has spread, and profitable trading signals now decay far more rapidly—losing half their excess return in roughly 18 months compared to five to seven years previously. When many investors converge on the same AI-driven conclusions simultaneously, winning strategies become crowded trades much faster.
Researchers at the University of Liechtenstein demonstrated that AI-driven trading systems are vulnerable to information manipulation. They built ten LLM-based trading models that all generated positive returns, yet every model was deceived by subtle alterations to financial news headlines, such as swapping visually similar characters or embedding hidden text. In the worst case, manipulation targeting a single stock on a single day caused a model's overall return to drop by approximately 18 percentage points, highlighting a serious new attack surface in automated trading.
A third strand of research reveals that AI models consistently take on far more risk than intended. In a simulated trading challenge by Elm Partners Management, four popular AI models matched elite macro traders in directional accuracy but ran daily return volatility of 20% to 40%—well above the recommended 7% to 15% range. The emerging body of research shifts the AI-in-finance debate from whether machines can beat markets to what happens when everyone competes through the same machines, warning that overconfidence, crowded trades, and manipulated information could spread more easily through AI-dependent markets.