MIT Study: AI Financial Advice Is Surprisingly Good

MIT Sloan researchers analyzed the quality of financial advice from GPT-5.2, GPT-5.6, and Gemini 3 Flash. Based on generic prompts written by 1,000 people and structured prompts designed by financial academics, they simulated the financial outcomes for individuals aged 22 to 89 who followed the advice over the long term.

Overall, the LLMs consistently offered the following directions:

• Increase savings during working years.
• Draw down savings after retirement.
• Invest actively in diversified equity funds.
• Reduce equity allocation after age 45.

Following this advice produced a substantial savings buffer for most people aged 30 and older, and yielded better outcomes than existing financial behavior in terms of savings rate, stock market participation, diversification, and age-appropriate risk-taking.

However, the LLMs did not respond flexibly enough to economic shocks such as unemployment. They recommended cutting spending excessively even for unemployed individuals who had savings buffers, and also showed a tendency for asset allocation to drift over time rather than being actively rebalanced.

How users phrased their questions also affected performance. Using a structured academic prompt that specified age, employment status, income, savings, life expectancy, retirement timing, employment/income risk, and tax and Social Security assumptions improved the quality of advice. Even in this case, however, active portfolio rebalancing remained lacking.

The researchers assessed that LLMs have the potential to provide cheap, accessible financial guidance while reducing the costs and conflicts of interest associated with traditional financial advice. At the same time, the findings show that real-world use requires designing prompts that sufficiently reflect individual circumstances and economic shocks, and continuously reviewing the model's advice.