He brought AI to Wall Street in 1994 — but won’t trust ChatGPT with his money
MarketWatch Link: LINK
When should people trust a machine over their own judgment?
Vasant Dhar has been building artificial intelligence long before anyone called it that. In 1994 he left academia to bring machine learning to Wall Street, then founded SCT Capital, one of the first hedge funds built entirely on machine-driven trading. His firm’s algorithms still trade markets every day.
Dhar is a professor at the NYU Stern School of Business and the NYU Center for Data Science, where his research centers on prediction, AI governance and a question he keeps returning to: When should people trust a machine over their own judgment?
Dhar explores that question in his book, Thinking With Machines: The Brave New World of AI, and on his podcast and newsletter, Brave New World.
In this recent interview, edited for length and clarity, Dhar shares why he would not hand his own money to a chatbot, what his AI valuation tool found when it disagreed with the human expert it was built to mimic, and how investors can decide where AI actually belongs in their own portfolio.
MarketWatch: Are retail investors using AI effectively or are they making expensive mistakes?
Dhar: This gets to the heart of when algorithms do better than humans, and when humans do better. People should not be doing frequent, short-term trading on their own. Unless you have a real system or algorithm, you will lose money doing that, and the more you trade the faster you will lose.
Long-term investing is different. Until large language models came along, AI could not help with decisions like whether to buy Nvidia or SpaceX, since those require thinking through uncertainty and imagining future scenarios. That used to be a purely human exercise.
Now LLMs can think through those decisions with you, which makes them useful partners. But trusting them blindly is a mistake unless they are grounded in real data. I asked ChatGPT whether to go long on Nvidia, BYD and the S&P 500, among other names, and got an answer. I asked again later and got a different one. The variance is real, and it’s highly sensitive to how you frame the question.
MarketWatch: You built an AI tool around one specific investor’s thinking. What is it?
Dhar: It’s called the Damodaran Bot (DBOT), after Aswath Damodaran, the “dean of valuation.” You ask it to value a company the way Damodaran would, based on a specific investment thesis you give it.
It uses a free cash flow to the firm approach, based on four key drivers: operating margins; revenue growth; weighted cost of capital; and reinvestment efficiency, measured through a company’s sales-to-capital ratio. It’s all standard methodology that Damodaran has made publicly available.
The advantage of the DBOT is speed. While Damodaran might publish one or two valuations a month, the bot can run that same analysis across the entire S&P 500 with the push of a button.
MarketWatch: Has the bot ever disagreed with what Damodaran would have concluded himself?
Dhar: Yes, and SpaceX is a good example. The IPO valued the company at $1.77 trillion. Damodaran’s own valuation came in around $1.2 trillion. The bot’s number was roughly $600 billion, less than half of Damodaran’s figure, making it even more bearish than the human it was modeled on.
The bot ran a deep analysis on SpaceX’s three businesses: the rocket-launch business, Starlink and the AI business built around xAI. It found too much uncertainty in the AI segment, which requires massive capital spending with an unclear payoff. Damodaran’s own analysis took the company’s numbers more at face value than the bot did, which pushed back against some of the assumptions.
The bot’s report shows a table comparing its assumptions for those four value drivers for SpaceX against the market’s assumptions, letting investors decide for themselves which numbers look reasonable.
MarketWatch: If someone has money to invest, would you trust AI to help decide where it goes?
Dhar: That depends entirely on the person. The first question is whether you’re genuinely passionate about investing, or just trying to grow money on the side while working a full-time job. The answer differs in each case.
If you don’t have the time or interest, put the money in an index you believe in. If you believe in the long-term growth of the United States, buy the S&P 500 and forget about it. If you believe a handful of companies will keep dominating technology, invest — but still do the research yourself.
Would I trust AI to make that decision for me? No, not unless I had a system whose outputs I’d already tested, and I’d want to see how credible its answers are and how much they change with a different thesis. I would never just ask an AI whether to go long on a stock and act on the answer alone. That would not be a good idea.
MarketWatch: What is the bear case on AI in investing? Does it ever get dangerously wrong?
Dhar: All the time. AI can be wrong, just like people can. But one question I’ve started getting from reporters is what happens if AI gets so good that everyone starts using it. Does that raise systemic risk?
The answer is yes, because it creates herding. If everyone trusts the same AI and acts on the same signal, you could get bigger shocks to the system when something goes wrong. In a strange way, AI starts to become the market itself, and outperforming the market means knowing when the AI is wrong.
We’re not at that point yet — not even close — but the scope of AI in finance is expanding. I’ve thought about whether AI could eventually play a role of a central banker, detecting bubbles or risk buildup that humans might miss. We genuinely don’t know yet whether on balance AI will lead to more stability or volatility. A lot depends on the checks and balances we put in place.
AI changes where the real work happens. People can ask the machine to run the analysis and spend their own time asking better questions instead. I recently wrote a piece asking whether I’m becoming obsolete, as a portfolio manager and a professor, since machines can now do so much of what I used to do by hand. The short answer is yes, but only as long as they can ask the right questions and interpret intermediate results correctly.
MarketWatch: Your fund, SCT Capital, still trades using machine learning. What was it like bringing these methods to Wall Street in the 1990s?
Dhar: Both exciting and difficult. When I went to Morgan Stanley in 1994 after several years in academia, Wall Street was full of physicists and economists. The physicists liked physical models, and economists liked linear models, so machine learning didn’t fit cleanly into either camp and was viewed with suspicion.
It left plenty of low-hanging fruit for me. I’d already seen machine learning work well at ACNielsen in identifying and predicting consumer buying patterns, and I brought those tools to trading. Fortunately, they substantially improved the performance of the proprietary trading group I worked with. I kept building AI models at Morgan Stanley and Deutsche Bank, then spun out SCT Capital as an independent firm in 1998.
The methods I use have evolved since, from machine learning to deep learning to what we have today. I now use vision models, the same technology that powers driverless cars, and I work with sovereign wealth funds on systematic signals for commodities trading. My algorithms trade every day using AI.
MarketWatch: When did it become clear that AI had crossed into something fundamentally new?
Dhar: Things look different from the inside than the outside. Around 2015 and 2016, machine translation started becoming genuinely good, almost perfect in some language pairs. That was an early sign we were starting to crack natural language.
The real turning point was in 2017, marked by the publication of a paper out of Google called Attention Is All You Need, which introduced the architecture behind today’s large language models. By the time ChatGPT launched in 2022, it wasn’t much of a shock to people in the field, but it was a pivotal moment for AI, showing that it could engage in a coherent conversation about anything with anybody in natural language.
What ChatGPT did was turn AI from a narrow application into something everyone could use and identify with. That was the real moment of transition, from an application to a general-purpose technology anyone could use. That’s the world we’re in now.