AI financial advice is surprisingly good, especially if you ask right questions

A new study reveals that people are increasingly turning to AI for financial advice, though the long-term impact on their financial well-being remains unclear. Assistant Professor Taha Choukhmane from MIT Sloan School of Management co-authored the research, which assessed the quality of financial advice given by large language models (LLMs).

The researchers discovered that following AI recommendations could result in considerable savings for individuals over 30 years old. AI consistently advised saving during working years, drawing down funds in retirement, investing heavily in diversified stock funds, and reducing stock exposure after age 45. However, AI struggled to adapt to shocks like unemployment and failed to actively rebalance portfolios.

Better prompts improved the quality of LLM advice, according to the study, yet the AI still often generated insufficient active portfolio rebalancing. The researchers created a model of how people's incomes, jobs, investments, and taxes typically change over their lives. They then asked 1,000 adults to create prompts seeking spending and investing advice from GPT-5.2, GPT-5.6, or Gemini 3 Flash. Simulations showed how people aged 22 to 89 would fare when following the AI's advice over time.

The study found that LLMs can provide affordable, accessible financial guidance, helping users beat the high costs, biases, and conflicts of interest often associated with traditional human advisors. The advice from LLMs was generally good, particularly when questions were asked in an academic manner. However, the models struggled with subtle aspects of financial planning and failed to adjust sufficiently to changing circumstances.

For example, they advised people who had lost their jobs to cut spending too drastically, even when they had savings.

The advice from LLMs also varied depending on the user's gender, financial literacy, and experience, resulting in wealth gaps. Men, more financially literate users, and those with prior AI experience generated about 5% more wealth close to retirement. This disparity arose from differences in the types of questions asked, with women using terms like "family," "grocery," and "pay," while men preferred "strategy," "crypto," and "growth."

Additionally, the LLM's advice might change when the same prompt was labeled as coming from a person of a different gender, reflecting potential biases learned from the training data.

Despite the promising results, Choukhmane emphasized that there are no clear benchmarks for AI financial advice. When a framework for how advice should vary with demographics exists, LLMs can progress in the right direction.