An MIT study has found that AI-generated financial advice performs better than expected across the full arc of household finances — good enough, the authors conclude, to count as an affordable alternative to professional guidance for people who could never afford a planner [1].

The research, led by MIT Sloan's Taha Choukhmane with co-authors at MIT and Stanford, tested what actually happens when ordinary people take advice from chatbots. About 1,000 participants wrote their own prompts seeking spending and investing guidance from GPT-5.2, GPT-5.6, or Gemini 3 Flash. The team then simulated lifetimes — ages 22 through 89 — in which households followed that advice, benchmarked against a life-cycle model encoding what good decisions look like given realistic incomes, taxes, and shocks [1].

The core finding is quietly radical. Following chatbot advice steered virtually everyone toward higher savings, stock-market participation, diversified portfolios, and age-appropriate risk reduction — behavior that compounds into meaningful saving buffers for anyone over roughly 30 [1]. Against the population's actual baseline behavior, the machines scored well. The advisory industry's core claim, that planning requires a credentialed human in the loop, survived the study in narrower form than the industry would like.

The failures were equally specific. Models handled steady states well and transitions badly: they told newly unemployed users to slash spending even when buffers existed, relied on rules of thumb where judgment was needed, and allowed portfolios to drift rather than rebalance [1]. Structured, information-rich prompts — the kind a finance professor writes — closed much of that gap. Ordinary people do not write those prompts, which is where the study turns from benchmark to warning. Advice quality tracked the prompter: men, the financially literate, and AI veterans extracted recommendations worth roughly 5% more terminal wealth, leaving women and novices tens of thousands of dollars behind at retirement [1]. Two-thirds of that gender gap came from prompt phrasing alone.

The paper's sibling analysis in this edition examines the usability gap underneath those numbers; the technical takeaway here is methodological. This was not a survey of chatbot opinions but a simulation of adopted advice across simulated lives — closer to how epidemiologists model treatment effects than to how tech journalists grade answers [1]. Its weaknesses are the mirror of its strengths: simulated households follow advice more consistently than humans do, and the models tested will be superseded before the findings are settled science.

Industry response has followed predictable lines. Advisory bodies emphasize that controlled conditions miss life events, family dynamics, and behavioral coaching — all true, all beside the distributional point (see the companion analysis). The Financial Planning Association's caution that controlled studies do not replicate client relationships is fair; so is Choukhmane's observation that the people who most need financial advice are precisely the people who cannot pay for it [1]. For them, the relevant comparison is not chatbot versus planner. It is chatbot versus nothing.

That reframing is the study's real contribution to the disruption debate X keeps having. If the counterfactual is expert human advice, AI looks like a discount substitute with gaps. If the counterfactual is the median American's unadvised financial life, it looks like infrastructure. Regulators will spend the next several years deciding which comparison governs fiduciary rules written before either existed [1].

-- KENJI NAKAMURA, Tokyo