One advisor for the whole world?

2Literature

There is a growing literature on AI financial advice, which also relates closely to the literature on robo-advising and traditional financial advice. Beginning with this newer literature, Choukhmane et al. (2026) elicit prompts from a representative U.S. sample, simulate the life-cycle paths that following the advice would generate, and confront those paths with the prescriptions of a calibrated life-cycle model. They find that current models give moderately risk-averse advice that is broadly consistent with life-cycle theory, yet relies on round-number heuristics and varies with the gender and financial literacy of the person asking. The reliance on round numbers is also a feature of the results of this paper. Ross and Lo (2026) use a different approach, where they draw a thousand synthetic client profiles with near-orthogonal characteristics, collect allocations from a single vendor’s model family, and fit interpretable surrogate models. They find evidence for “heuristic collapse”: self-reported risk tolerance is the most important predictor of advice, while age, income, horizon, and liquidity contribute only marginally. Similarly, Fieberg et al. (2025) elicit portfolio recommendations from 32 LLMs for 64 hypothetical investors who differ in risk tolerance, investment horizon, sustainability preferences, gender, and investment experience, for hypothetical investors in China, Germany and the United States. Most models recommend implementable ETF portfolios whose risk exposure rises with the investor’s stated risk tolerance, and these portfolios earn historical risk-adjusted returns similar to those of the robo-advisors. The recommendations overweight domestic securities, consistent with a home bias inherited from the training data. Finally, Foltyn and Olsson (2026) query 33 models from five vendors with prompts that differ only in the word “man” or “woman,” and find that the models advise women to hold 1.8 percentage points less in equities than men. The gap falls to 0.8 points but does not vanish when the prompt adds marital status, education, income, and risk tolerance. This paper adapts their stock-bond question as its primary outcome and replaces the gender treatment with the country and language of the question.

This paper is also relevant for the literature on robo-advising and algorithmic financial advice. Earlier robo-advisors improve diversification and reduce behavioral biases (D’Acunto et al. 2019), deliver substantial welfare gains (Rossi and Utkus 2024), and broadly move household behavior closer to normative prescriptions (D’Acunto and Rossi 2023). In credit markets, machine-learning scoring changes the distribution of mortgage approvals across racial groups (Fuster et al. 2022). LLMs differ from these robo-advisors, because they are general-purpose rather than domain-restricted and because they reach essentially every language and country at once. These models therefore loosen the supply-side constraints on financial advice that partly determine who accesses and benefits from advice (Reuter and Schoar 2024).

What kind of advice do investors usually get? Canner et al. (1997) show that popular advisors recommend a higher ratio of bonds to stocks for more conservative investors, contrary to the mutual-fund separation theorem. Choi (2022) documents how popular personal financial advice departs from the prescriptions of economic theory. In other studies, advisor fixed effects explain more of the variation in client portfolios than client characteristics do (Foerster et al. 2017), which suggests that human advisors give one-size-fits-all advice. AI models reproduce both these patterns, but do so in many different countries at once.

Finally, this paper contributes a new test to the literature on how multilingual language models process and retrieve knowledge. There are two strands of this literature. Interpretability studies find that the middle layers of multilingual models operate in a shared concept space. This space, where the models solve problems, lies closer to English than to any other language (Wendler et al. 2024; Tang et al. 2024; Wu et al. 2025; Lindsey et al. 2025). Knowledge-probing studies instead find that the same model often asserts different facts depending on the query language (Kassner et al. 2021; Qi et al. 2023). These two findings make different predictions for the prompts in this study. Shared processing predicts nearly the same advice in every language, as the portfolio is solved once in the middle layer in English. Language-keyed retrieval predicts that advice inherits the conventions of each language’s financial text, meaning that a prompt in German would bring up German-language texts with a specific type of advice. The results support the first interpretation, shared processing, where the local detail added by the models appears in the explanation rather than in the allocation. The questions that the models would like to ask investors also vary very little across countries, which again supports shared processing.