How good is AI financial advice over a lifetime?”
Does LLM-generated financial advice move households closer to or further from optimal economic behavior, and why does that advice differ systematically across individuals? This is the question that Taha Choukhmane, Tim de Silva, Weidong Lin, and Matthew Akuzawa analyze in this paper. The authors proceed in three steps:
- Survey: 1,000 demographically representative U.S. adults (via Prolific) were asked to write real, free-text prompts seeking spending and investment advice from an LLM i.e. the same kind of informal messages people actually send.
- Lifecycle model: A quantitative lifecycle simulation model was calibrated to U.S. income dynamics, employment transitions, asset returns, and tax rules, providing both a simulation engine and a normative benchmark grounded in economic theory.
- Simulation: Survey prompts were fed to GPT-5.2 (and Gemini 3 Flash as a robustness check). The textual advice was translated into quantitative choices (saving rates, asset allocations) and simulated year-by-year from age 22 to retirement, tracking wealth accumulation under realistic conditions.
Findings
- LLM advice broadly aligns with lifecycle theory. Following the advice would move most respondents closer to textbook financial behavior: near-universal stock market participation, equity shares that decline with age after 45, and savings buffers above $10,000 built by age 30.
- Systematic departures from optimal theory. The LLM relies on simple heuristics (31% of saving rate recommendations are multiples of 10%; over 98% of withdrawal recommendations follow the “4% rule”), insufficiently smooths consumption after income shocks, implies an unrealistically high discount factor, and allows portfolios to drift passively rather than recommending active rebalancing.
- Advice varies inequitably across individuals. Wealth at age 60 differs by 4–6% depending on who writes the prompt. Users with higher financial literacy or prior AI experience receive better advice. Women receive equity allocation recommendations 3.5 percentage points lower than men: two-thirds of this gap is demand-driven (how men and women frame questions differently) and one-third is supply-driven (the model treats identical prompts differently based on stated gender).
AI Financial Advice: Supply, Demand, and Life Cycle Implications
Authors: Taha Choukhmane, Tim de Silva, Weidong Lin, Matthew Akuzawa
From: Massachusetts Institute of Technology (MIT), Stanford University
Gender discrimination in LLM financial guidance
Do large language models provide gender-neutral financial advice? If not, does providing richer investor information eliminate the disparity? This is the question Richard Foltyn, Jonna Sofia, and Kristina Olsson analyze in this paper. The authors used a controlled experimental design:
- 41,800 queries were submitted to 33 LLMs from five vendors (Anthropic, DeepSeek, Google, OpenAI, xAI), covering roughly 95% of the retail market.
- The core treatment is minimal: prompts are identical except for one word “man” versus “woman,” in the sentence “I am a 40-year-old {man/woman} looking to invest $50,000 over a five-year horizon.”
- Since LLM responses are stochastic, each model is queried repeatedly (200 times per gender/model combination) to generate a distribution of recommendations, not just a single answer.
- A second experiment used implicit gender signals via occupation (e.g., nurse vs. engineer) rather than explicit labels. A further set of experiments progressively adds contextual information (marital status, income, education, risk preferences) to test whether richer prompts eliminate the gap.
Findings
- Persistent gender gap: Women are advised to allocate 1.8 percentage points less to an S&P 500 ETF than otherwise identical men . This is consistent across all vendors, model generations, and complexity levels.
- Not improving over time: The gap is if anything slightly larger in the most recent model generation (2.0pp) than in earlier models (1.6pp), and high-complexity models display a larger gap than medium-complexity ones.
- Implicit gender signals also trigger the gap: When gender is conveyed through occupation (e.g., nurse vs. engineer) rather than stated directly, the market-share-weighted gap falls modestly to 1.3pp but remains highly significant.
- Richer information attenuates but does not eliminate the bias: Adding income, education, marital status, and risk preferences shrinks the gap to 0.8pp, but all vendors except DeepSeek still show a significant negative female coefficient.
- A striking anomaly: For one model (Anthropic Claude 4.5 Sonnet), prompts identifying the investor as a married woman trigger a refusal to provide advice in 76% of cases. This never occurs for the equivalent married man prompt.
- Economic implications: A 1.8pp lower equity allocation implies an annualized return wedge of 0.34 percentage points over 2023–2025. Modest for one individual, but potentially significant when multiplied across millions of users, repeated decisions, and compounding over time.
The Worth of a “Wo”: Gender Bias in Financial Advice from LLMs
Authors: Richard Foltyn, Jonna Sofia Kristina Olsson
From: Norwegian School of Economics
