MIT Professor Highlights Risks and 'Art' of AI Financial Prompting

Andrew Lo, director of MIT's Laboratory for Financial Engineering, warns that the utility of generative AI in personal finance depends heavily on the quality of the user's prompts. While tools like ChatGPT, Claude, and Gemini can explain broad concepts such as investment diversification or the benefits of ETFs, they often struggle with precise mathematical tasks. The risk is particularly high in tax planning and specific numerical calculations. Lo notes that large language models (LLMs) often produce 'hallucinations,' delivering incorrect answers with a tone of authority that can mislead unsuspecting users who may not double-check the output. The trend toward AI-driven finance is growing rapidly. According to an Intuit Credit Karma poll of 1,019 adults, 66% of generative AI users have sought financial advice from these tools, with adoption exceeding 80% among Millennials and Generation Z. Alarmingly, approximately 85% of those users acted on the AI's recommendations. To improve outcomes, Lo suggests a 'science' of prompt engineering. Rather than asking generic questions, users should provide detailed constraints, including tax brackets, risk tolerance, and specific goals. He recommends instructing the AI to act as a 'fee-only fiduciary,' forcing the model to frame advice within a legal standard of the client's best interest. Ultimately, Lo emphasizes that AI should be treated as a conversational tool requiring iterative prompting—sometimes over 20 exchanges—and rigorous manual verification of all outputs to ensure accuracy in personal financial planning.