8Conclusion
This paper studies the financial advice that a handful of AI models now supply to households around the world. I ask three financial questions to thirteen frontier models in twenty-one countries. Summing up the results, I find that the models give nearly the same answer everywhere on the portfolio allocation question. Country means span twelve points and every significant contrast with the United States is negative. On the other questions, the models localize the advice when the question depends on the local economy. On the mortgage question they recommend investing 10 to 29 percentage points less than in the United States in every country where the prompted rate is realistic. Asked the same question without a rate, they assume local mortgage rates and advise accordingly. On the inflation question they switch from equities to inflation-linked bonds in Brazil, Turkey, and Mexico.
I use a simple benchmark to determine what should move the advice. The benchmark together with what we know about popular financial advice helps organize the evidence collected in the paper, by forming three candidate advisors for the allocation question. The Merton advisor treats the stake as a standalone portfolio and ignores both the investment horizon and the balance sheet. The Life-Cycle advisor adds the household’s balance sheet. The Questionnaire advisor follows the retail risk questionnaire of popular financial advice (Canner et al. 1997; Choi 2022) and regulation (e.g., Financial Services Authority 2011; European Securities and Markets Authority 2022).
The models behave like the Questionnaire advisor. Stated risk tolerance raises the recommended equity share by 46 percentage points and a twenty-year rather than one-year horizon raises it by 40, while a stated generous pension barely moves it. Asked what they would want to know before advising, the models ask the questions of a suitability questionnaire, in every country. The balance sheet enters only as a screen on the investor’s capacity for loss, which does move the advice, but a guaranteed annuity, a cash cushion, or a currency hedge barely raises it. Finally, the similarity across countries is not a failure to tell the countries apart. Local financial context appears in three quarters of non-U.S. explanations, while naming local instruments or asking in the local language moves the recommended share by about a point. This pattern is consistent with interpretability evidence that multilingual models solve tasks in a shared, English-leaning concept space. The explanation adapts to the country and the allocation does not, so the investor cannot tell from the answer whether the advice reflects her local circumstances.
Taken together, the results suggest that AI financial advice combines standardized investor assessment with selective geographic differentiation. Across countries, models ask similar questions and respond similarly to investor characteristics, yet their recommendations differ more substantially when the financial decision depends on local economic conditions. At the same time, the explanations are more geographically differentiated than the allocations themselves. The globalization of financial advice may therefore involve not just the cross-border supply of recommendations but also the spread of common rules for evaluating investors.
The concentration of advice production in a small number of models also creates the potential for correlated recommendations across households and countries. Where common advice rules fail to incorporate economically relevant local circumstances, this concentration may also propagate common mistakes. Understanding when geographic uniformity is warranted, and when it reflects inappropriate standardization, is therefore an important question for the emerging global market for AI financial advice.