MIT study: AI financial advice is solid, but prompt gaps cost users $100K

Half of Americans now say they use AI for financial advice, yet until recently nobody had systematically measured whether that advice is any good. A new paper from MIT Sloan and Stanford, covered by MIT’s Ideas Made to Matter, attempts exactly that, and the headline finding is genuinely surprising: the advice is better than the researchers expected.

The study, led by MIT Sloan assistant professor Taha Choukhmane with Weidong Lin, Matthew Akuzawa, and Stanford’s Tim de Silva, built a life-cycle model of incomes, jobs, investments, and taxes, then asked 1,000 adults to write their own prompts for GPT-5.2, GPT-5.6, or Gemini 3 Flash. The researchers simulated what would happen if people aged 22 to 89 followed that advice over time. The results, which won the Swiss Finance Institute Outstanding Paper Award 2026, show LLMs steering users toward higher savings, stock market participation, diversified allocations, and age-appropriate risk-taking.

“We were somewhat surprised by how good the advice was,” Choukhmane said. “Especially when you read the kind of questions people asked, it was not a given that the advice would line up with what academics think are good financial principles.”

But the paper’s real contribution is not the headline. It is the granular detail on where LLM advice breaks down, and the dollar figures attached to those failures. Following advice generated in response to prompts written by men, financially literate users, or people with prior AI experience produced roughly 5% more wealth near retirement. The compounding effects are stark: women and less financially literate users ended up with about $50,000 (4%) less wealth at age 60, while users without prior AI experience saw almost $100,000 (6%) less.

Prompt quality is a wealth variable

The most actionable finding is that prompt structure directly moves retirement outcomes. The researchers compared user-written prompts against “academic” prompts that included full financial information, clear assumptions about life expectancy, retirement age, employment risk, and current tax law. The structured prompts produced better advice on spending and saving, and cut down on rule-of-thumb answers.

“Regular people are not writing their prompts the way a finance professor is,” Choukhmane said. A typical user prompt might read: “Where should I invest starting with $50 and consistently adding $25 a month after?” The academic version tells the model to assume normal life expectancy, current U.S. tax law, and no changes to Social Security rules.

This is the kind of result that should make AI product teams rethink their interfaces. If the quality of financial advice depends on the user’s ability to write like a finance professor, then the tool is only as good as its least literate user. The gap is not a model limitation; it is a design failure. Chat interfaces that accept free-form prompts are handing the burden of financial sophistication to the people who need it least.

The gender gap is partly the model’s fault

The paper separates the wealth gap into two sources. About two-thirds comes from differences in how men and women write prompts. Women were more likely to use words like “family,” “grocery,” and “pay,” while men used “strategy,” “crypto,” and “growth.” The remaining third comes from the model changing its advice when the same prompt was labeled as coming from a woman rather than a man.

Choukhmane is careful to note that not all variation is problematic. “We want [the LLM] to have different bias because men and women are different and have different life expectancy and income risk,” he said. But the model could make those inferences explicit to users, rather than silently adjusting advice. The alternative explanation, that the model learned gender biases from training data, is harder to fix and easier to ignore.

The advice has blind spots

The LLMs consistently advised saving during working years, drawing down in retirement, and reducing stock exposure after age 45. But they failed on shocks. When users experienced job loss, the models advised cutting spending too sharply, even when the user had savings. The models also allowed portfolios to drift rather than actively rebalancing them.

These are exactly the situations where a human advisor earns their fee. A good advisor rebalances after market moves and adjusts spending plans after a layoff. The LLMs treat financial planning as a static optimization problem, not a dynamic one. That is a fundamental limitation of the current architecture, not a prompt-engineering issue.

What this means for the AI economy

The paper’s business findings are quietly significant. LLMs recommended specific account types and providers that users never mentioned. Vanguard products appeared in 6% of LLM responses, and iShares in 3.4%, even though fewer than 0.4% of prompts mentioned either company. Financial firms are discovering that their visibility in AI responses matters more than search rankings or traditional marketing.

For AI builders, the implications are direct. Financial advice is a high-stakes domain where prompt quality, user demographics, and model behavior compound into tens of thousands of dollars of retirement wealth. The same dynamics apply to legal advice, medical triage, and tax preparation, any domain where a model’s response depends on how the user frames the question.

The study’s authors acknowledge there are no clear benchmarks for what good AI financial advice looks like. Choukhmane said an accepted framework for how advice should vary with demographics is necessary before LLMs can progress in the right direction. He remains hopeful.

The honest takeaway for builders is uncomfortable. The models are good enough to be trusted, but not good enough to be trusted equally. A tool that delivers 5% more retirement wealth to men, to the financially literate, and to people who already use AI is not a neutral technology. It is a wealth amplifier with a bias toward the already advantaged. Fixing that requires more than better prompts. It requires interfaces that elicit the right information, models that make their demographic assumptions explicit, and benchmarks that measure outcomes, not just response quality.

The paper ends with a practical suggestion: use AI as a tool for building financial understanding rather than simply following its advice, and treat it as a complement to a human advisor. That is sound advice for consumers. For the industry, the challenge is building systems that deliver the academic-prompt outcome to users who will never write like a finance professor.