Study finds AI financial advice improves savings and diversification but lacks nuance and rebalancing

A new study by researchers at MIT Sloan evaluates the quality of financial advice generated by large language models (LLMs) and finds that following AI recommendations can lead to improved financial behaviors, such as higher savings rates and more diversified investments. The research, co-authored by Taha Choukhmane, assistant professor of finance at MIT Sloan, simulated the financial outcomes of 1,000 adults over their lifetimes after receiving AI-generated advice. The study compared the advice generated by user-written prompts to more structured, academic-style prompts, as well as to baseline financial behaviors without AI assistance.

The findings indicate that AI advice consistently encouraged behaviors aligned with academic financial principles, such as saving during working years, reducing stock exposure after age 45, and investing in diversified stock funds. However, the study also identified significant limitations. AI advice struggled to adjust to financial shocks, such as unemployment, often recommending overly conservative spending cuts even when savings were available. Additionally, the models frequently allowed portfolios to drift without active rebalancing, a key component of long-term financial planning.

The quality of AI advice improved when prompts were more detailed and structured, resembling academic prompts that included full financial information and explicit assumptions about the economic environment. For example, prompts that specified age, job status, income, savings balances, and economic assumptions led to better-aligned advice. In contrast, typical user prompts—such as requests to invest a small monthly amount without context—produced less reliable guidance.

The study also highlighted disparities in advice quality based on user characteristics. Prompts written by men, financially literate users, or those experienced with LLMs resulted in higher-quality advice, potentially widening wealth gaps. Choukhmane noted that regular users often do not frame prompts in ways that elicit optimal financial advice, emphasizing the need for better prompt design and user education.

The researchers concluded that while LLMs can serve as an affordable and accessible source of financial guidance, their advice should be supplemented with human judgment, especially in complex or dynamic financial situations. The study underscores the importance of addressing biases in AI-generated advice and improving prompt engineering to maximize the benefits of AI financial tools.