Hedge fund Man Group is developing an artificial intelligence system called AlphaGPT. The AI can process financial data around the clock, continuously generating investment hypotheses at a pace of dozens per few minutes. One AI agent automatically writes code to quickly test these hypotheses, while another evaluates their economic and logical validity. The process currently operates under human supervision, but plans call for expanding the scope of automated monitoring.
As generative AI advances, the way asset management operates is shifting. The hedge fund Numerai incentivizes data scientists worldwide to submit predictive models using tokens, and the fund synthesizes these models for its use. J.P. Morgan Asset Management has committed up to $500 million to invest in Numerai. Additionally, generative AI lowers the barrier to autonomous trading, enabling individuals without deep financial theory or programming expertise to write trading programs with AI assistance. Other frontier approaches include using multi-agent techniques to learn the tacit knowledge of veteran fund managers, or having multiple AI agents engage in discussions to refine investment ideas.
The spread of AI also raises concerns about market stability. Democratic U.S. lawmakers warned the SEC in a letter that AIs trained on similar data and optimized for similar objectives could react in sync, fueling one-way trades and amplifying volatility. When AI executes large-scale orders at extreme speed without human intervention, shocks can propagate almost instantaneously. Research shows that profit-maximizing AI in simulated markets may adopt behaviors resembling market manipulation, and even without explicit intent to break rules, they can default to tacit collusion. Malicious actors could also fabricate images and exploit AI’s automated response mechanisms to manipulate markets.
At the same time, AI can serve as a powerful risk-management tool. A University of Tokyo research team led by Professor Kiyoshi Izumi built a virtual crisis scenario for option trading to test hedging strategies. They also created an artificial market, a digital twin of financial markets populated with diverse AI agents, to artificially trigger crashes and track behavior, analyzing which risk factors compound to become most dangerous. Regarding flash crashes where liquidity suddenly evaporates, Izumi says both outcomes are possible and depend on the tug-of-war between forces—some AI may act in concert and magnify swings, while others automatically identify and correct price deviations.
Takashi Mizuta, director of the Japanese Society for Artificial Intelligence and at Sparx Asset Management, points out that AI can be misled in high-frequency trading, but if markets develop diverse decision-making mechanisms, they may become more stable. He particularly warns of prompt injection attacks, where a malicious third-party tampers with the instructions an AI receives, causing it to be manipulated without the user’s knowledge. Keiichi Omura, professor emeritus at Waseda University, cautions that the history of financial innovation is both a history of efficiency gains and a history of breeding instability. AI, too, faces an “innovation trap,” requiring governance, ethics, and third-party oversight to progress at the same pace. Humans must still imagine never-before-seen black swans and prepare in advance, because AI will not dream of them.