Dare Not Hand Money to ChatGPT! Expert Warns 'Herd Effect' May Trigger Systemic Risks

As AI investment tools rapidly rise, Vasant Dhar, a New York University professor who brought machine learning to Wall Street over 30 years ago, has issued a warning: he won't entrust his own funds to ChatGPT to manage until AI models have been verified. While AI can assist in data analysis and investment scenario simulation, answers are easily influenced by the way questions are asked; if a large number of investors rely on the same model for trading, it could amplify market volatility.

In an interview with MarketWatch on Monday (27th), Dhar revealed that he once asked ChatGPT whether he should go long on NVIDIA (NVDA-US), BYD, and the S&P 500 index, and then asked the same question again, only to receive different answers. This demonstrates that the output of large language models is not stable, and conclusions may change based on the way questions are phrased, background data, and preset conditions.

He emphasized that while AI can help organize information, simulate scenarios, and execute corporate valuations, the analysis must be based on reliable data. If the model has not been tested and it's unclear how it will change conclusions under different investment assumptions, investors should not buy or sell stocks based solely on the answers from a chatbot.

In addition to individual investors potentially suffering losses due to incorrect answers, Dhar is also concerned about the impact of AI on the overall financial market. He pointed out that if a large number of investors trust the same AI and enter or exit the market based on the same signals, if the model makes a wrong judgment, it could trigger collective buying or selling, further amplifying price volatility.

This situation is known as the "herd effect," where investors no longer make judgments based on their own research but follow other market participants and take the same actions. As more and more people use the same or similar AI models, trading directions may become more concentrated, causing local errors to quickly spread throughout the financial system.

Dhar stated that the market is still some distance away from fully relying on the same AI, but as the application of AI in the financial sector continues to expand, in the future, in order to outperform the market, the key for investors may not only be obtaining AI signals, but also identifying when the model is making mistakes.

As to whether AI will ultimately reduce market volatility or cause greater instability, there is currently no clear answer, and the result will depend on whether financial institutions establish sound verification, supervision, and checks and balances mechanisms.

Dhar even believes that in the future, AI may help central banks detect asset bubbles or the accumulation of financial risks, finding signals that human decision-makers often overlook. However, like humans, AI can also make mistakes, so in the absence of accountability and review systems, it is not advisable to directly replace important financial decisions.

Regarding how ordinary investors should use AI, Dhar stated that first, they must clarify whether they are truly passionate about investment research, or only hope to grow their assets in addition to their full-time job. The time, knowledge, and risk tolerance of these two types of investors are different, and the strategies they should adopt are also different.

For those who do not have the time or interest to deeply research the market, he suggested that instead of frequently engaging in short-term trading, they should choose an index they identify with and hold it for the long term. If they believe that the US economy can continue to grow, they can consider investing in the S&P 500 index; if they are optimistic about a few technology giants continuing to dominate the industry, they should still personally research the company's fundamentals and should not completely entrust investment judgments to AI.

Dhar warned that individual investors without a complete trading system who frequently engage in short-term operations usually only increase the probability of losses, and the more frequent the trading, the faster the capital may be lost.

In contrast, large language models have a certain value in long-term investment research. In the past, evaluating the long-term value of companies like NVIDIA or SpaceX required analyzing highly uncertain industry prospects and inferring multiple future scenarios, mainly relying on human investors' judgments. Now, AI can help handle some of the work, but should be positioned as a research partner, not a replacement for investors making the final decision.

Dhar also introduced in the interview the Damodaran Bot, abbreviated as DBOT, a tool he developed. This tool is based on Aswath Damodaran, a professor of finance at New York University and known as the "valuation guru," and can analyze corporate value by simulating Damodaran's valuation methods based on the investment arguments put forward by the user.

DBOT uses the enterprise free cash flow valuation method, mainly examining four factors: operating profit margin, revenue growth rate, weighted average cost of capital, and reinvestment efficiency. Compared with human analysts, the biggest advantage of this tool is speed, as it can apply the same method to the entire S&P 500 index constituents in a short time.

However, DBOT does not simply copy Damodaran's views and sometimes comes to conclusions that are vastly different.

Taking SpaceX (SPCX-US) as an example, the company's valuation at its first public offering was $1.77 trillion, Damodaran's estimated reasonable value was approximately $1.2 trillion, but the valuation given by DBOT was only about $600 billion, less than half of Damodaran's valuation, and its stance was more conservative than the human expert it was modeled after.

DBOT separately examined SpaceX's rocket launches, Starlink satellite network, and AI business centered on xAI. The analysis concluded that the AI department requires a huge capital investment, but future returns are highly uncertain, so it raised questions about some growth assumptions.

In contrast, Damodaran used more data provided by SpaceX in his analysis, so he arrived at a higher valuation. DBOT, on the other hand, lists its own and the market's assumptions for the four key valuation factors side by side, allowing investors to judge which predictions are more reasonable, rather than directly giving a single buy or sell order.

Dhar stated that this case shows that the value of AI is not only in speeding up analysis or copying expert opinions, but also in helping investors verify the assumptions adopted by human analysts. However, the numbers derived from the model do not equal the correct answer, and the final judgment still needs to be made by the investors.

Dhar's involvement in AI financial applications dates back to 1994. At that time, he left academia to join Morgan Stanley (MS-US) and brought machine learning technology, which was previously used to identify and predict consumer purchasing behavior, into the proprietary trading department.

At that time, quantitative research on Wall Street was mainly dominated by physicists and economists, with the former preferring physical models and the latter mostly adopting linear models, making it difficult for machine learning to fully integrate into either camp, thus receiving many doubts.

However, the models developed by Dhar later improved the performance of the trading team. He continued to develop AI models at Morgan Stanley and Deutsche Bank (DBK-US) and, in 1998, spun off SCT Capital to establish one of the earliest hedge funds operating entirely based on machine trading strategies. Currently, the company's algorithms still use AI for trading every day.

With the evolution of technology, Dhar's methods have evolved from early machine learning to deep learning and visual models. He is currently also collaborating with sovereign wealth funds to develop systematic trading signals for commodities.

Dhar believes that the important turning point in AI development can be traced back to 2015-2016, when machine translation began to demonstrate near-human capabilities in some language combinations. In 2017, the Google research team published the paper "Attention Is All You Need," proposing the Transformer architecture, which later became the foundation of large language model technology.

After the release of ChatGPT in 2022, AI has further evolved from a specialized tool with limited uses to a general technology that can be used by the general public. Dhar believes that machines can now complete many tasks that previously had to be handled by portfolio managers and professors personally, but this does not mean that humans have become obsolete.

On the contrary, the truly important work has shifted to asking the right questions, examining the assumptions adopted by the model, and judging whether the analysis process and final conclusions produced by AI are trustworthy.