A new arXiv preprint audits how large language models give financial advice when a client's stated risk tolerance changes. The authors note that AI financial advisers typically set a stock allocation based on a single risk-tolerance question, but people do not answer that question consistently.
Using a national US panel, the researchers found that 41% of households chose a different answer when asked the same risk-tolerance question two years later. This instability raises questions about the reliability of advice that hinges on one response.
The paper suggests that LLM-based financial advisers should account for the variability in risk-tolerance answers, rather than treating a single answer as fixed. The full findings are not yet available, but the abstract highlights the potential for inconsistent portfolio recommendations.