MIT: LLMs Give Solid Financial Advice — But Bias Creeps In Through Your Prompt

MIT Sloan researchers put AI financial advice to the test and found it better than expected. Feeding prompts from 1,000 adults into GPT-5.2, GPT-5.6, and Gemini 3 Flash, then simulating the results across full life cycles, they found the models reliably nudged users toward sound behavior: save during working years, diversify into stock funds, dial back equity exposure after 45, and draw down in retirement. Following the guidance produced meaningful savings buffers for essentially everyone over 30, suggesting LLMs could be a cheap, accessible alternative to human advisors and their attendant fees and conflicts of interest.

The weaknesses are in the nuance. The models lean on rules of thumb and adapt poorly to changing circumstances — telling someone who just lost their job to slash spending even when they had savings to cushion the blow — and they let portfolios drift instead of actively rebalancing. Structured, detailed prompts that spell out age, income, employment risk, and economic assumptions produced markedly better advice, but even then the models under-rebalanced. The practical takeaway is that outcomes hinge heavily on how the question is framed.

The more troubling finding is that advice varies by who is asking, opening real wealth gaps. Prompts written by men or by financially literate users drew recommendations for higher equity allocations, compounding into roughly $50,000 (4%) more wealth by age 60; users with no prior AI experience were steered toward lower saving rates, leaving them about $100,000 (6%) behind. Roughly two-thirds of the gender gap traced to differences in how men and women worded their prompts, but the remaining third came from the model shifting its advice when the same prompt was simply labeled as coming from a woman — a pattern that could reflect reasonable demographic inference or learned training-data bias. With no agreed benchmark for how advice should vary across demographics, researchers say the fix for now is better prompts and explicitly asking the model to guard against bias.