AI in action · Investor protection
Study: AI Assistants Steered Investors Toward Index Funds
Retail investors who began using a ChatGPT- and Gemini-based assistant raised the index-fund share of new investment by 7.6 points, an NYU Stern study found.
The worry about chatbots and money is easy to state: an assistant that agrees with everything will tell an overconfident investor that their favourite stock is a great idea. A working paper dated November 8, 2026, by Valerie Baldinger of New York University's Stern School of Business, tests that worry with real trades rather than surveys, and finds the opposite tendency.
The data come from a US wealth management platform that added a conversational assistant built on OpenAI's ChatGPT and Google's Gemini models in November 2024. Because the platform links clients' brokerage and retirement accounts, the study can match 69,598 prompts and the AI's answers to what investors bought afterwards. The author worked as an unpaid contractor for the platform and reports no financial interest in it.
The first finding concerns the advice itself. Investors expressed interest in buying ETFs or mutual funds in 3.5% of their prompts, but 18% of the AI's answers recommended buying funds. Passive funds were recommended more often than individual stocks across topics, both when users asked for information and when they asked directly for advice, and the pattern held across model versions. The paper reads this as evidence against sycophancy in financial advice.
The second concerns behaviour. Comparing investors before and after they started using the assistant, with a staggered difference-in-differences design, the share of new investment going to index funds rose by 7.6 percentage points. New purchases carried less stock-specific risk and the funds bought had lower expense ratios. The paper's measure of the cost of under-diversification, the expected return given up by holding a portfolio less efficient than a diversified one with the same risk, fell by 43 basis points a year on average, and by 101 basis points for the less diversified half of users.
The findings sit awkwardly next to an earlier National Bureau of Economic Research working paper, reported in June, which asked chatbots to build stock portfolios directly and found them averaging 41% in semiconductor shares, about double the S&P 500's weight, with picks that tracked media coverage. The difference is the question: asked to pick stocks, models pick popular stocks; asked how to invest, they lean towards the textbook answer.
Both studies carry limits. Baldinger's sample is one platform's users, who are wealthier than the average US equity holder, and the paper has not yet been peer reviewed. The practical point for investors is about prompts more than tools: the same model can push towards concentration or diversification depending on what it is asked to do.
Sources: Valerie Baldinger, NYU Stern — Generative AI and Retail Investment, working paper (Nov 8, 2026, PDF); Investopedia via Yahoo Finance — New research warns AI stock advice heavily favors tech (Jun 15, 2026)