Conference Agenda

How does AI-generated personal financial advice from Large Language Models (LLMs) compare to economists’ normative models? We develop and implement a method to evaluate generative AI financial advice by simulating thousands of life cycle paths for consumption, saving, and portfolio choices under realistic income, employment, and asset return scenarios. Our approach compares these LLM-generated paths against optimal choices from a standard life cycle model and can parsimoniously summarize LLM advice across prompts and models by estimating structural time and risk preferences. Applying our method to OpenAI’s GPT-5 mini, we find that the advice qualitatively aligns with standard life cycle theory but deviates systematically in four key ways: (i) recommended consumption and saving paths imply unrealistically high patience, with estimated intertemporal discount factors well above one; (ii) recommended choices often reflect simple heuristics, such as round savings rates, fixed-percentage withdrawal rules in retirement, and common asset allocation rules-of-thumb; (iii) LLM recommendations exhibit substantially more inertia in portfolio rebalancing and moderately less consumption-smoothing in unemployment than our normative benchmark; and (iv) holding all else constant, recommendations vary systematically with demographics (e.g., recommending lower equity shares for women) and between repeated identical queries.