Source: arXiv

Paper

Do People Follow AI Advice? Evidence from a Pension Portfolio Choice Experiment — Choi, Kim, Kovach, Lee, Shin, Tzavellas (arXiv 2608.11371, Aug 2026). Korea-touching authorship includes Sejong University and KAIST (with coauthors at U.S. universities).

What happened

General-purpose AI can spit out retirement allocations in seconds. Whether that changes real behavior depends on pass-through — how much of the advice difference ends up in final portfolios. This preprint runs an incentivized experiment with 400 employed adults in South Korea who are enrolled in workplace defined contribution (DC) pension plans. Participants first allocate a hypothetical DC balance across eleven products, then may revise after seeing one of two fixed AI-generated recommendations.

A 2×2 design randomizes (1) recommendation content — relatively aggressive versus conservative portfolios from frontier GPT models given identical product information — and (2) whether a short rationale accompanies the numerical advice. Participants know the recommendation is AI-generated but are not told which model produced it. The authors treat the model contrast as variation in advice content, not as a horse race between named vendors.

The breakdown

Approximately 37% of the experimentally induced difference between the aggressive and conservative recommendations passes through to final portfolios. That causal contrast shifts expected return, volatility, risk-grade mix, and how many products people hold — but produces no detectable difference in computed Sharpe ratios. In other words: AI advice here moves risk exposure, not a clean risk-adjusted-performance scoreboard.

On the revision margin: 81% of participants revise after seeing the advice. Among revisers, 95% move toward the assigned recommendation and implement about half of the suggested adjustment. Incomplete aggregate pass-through therefore mixes an extensive margin (nearly one fifth never revise) with an intensive margin (revisers stop partway and still keep weight on their own baseline). Short rationales do not detectably change average pass-through.

Why it matters outside Korea

If you work abroad with Korean colleagues on DC plans, design fintech for Korea-market retirement menus, or brief HQ on “will ChatGPT change how Koreans invest?,” this is a behavior-measurement note, not a product endorsement. Overseas HR and benefits teams should expect partial, selective uptake: enough that advice content still shapes risk people end up bearing, not enough that a chatbot replaces the user’s prior judgment. Researchers tracking generative-AI advice-taking get a rare causal pass-through estimate in a multidimensional, jointly constrained allocation — closer to real DC menus than a single yes/no click.

Soft constraint for readers: this brief paraphrases an academic experiment. It is not investment, tax, or legal advice, and it does not tell you to follow any chatbot’s pension mix.

What travelers and expats should watch

• Do treat AI pension suggestions as directional inputs you may partially adopt — the study’s revisers typically move toward advice but only about halfway.
• Don’t assume a polished rationale will change how much of the advice sticks; here, rationales showed no detectable pass-through shift.
• Expect aggressive-versus-conservative advice to show up as different risk levels and product mixes, not as a guaranteed Sharpe improvement.
• Re-check any personal DC or retirement choice against plan documents and a licensed adviser — especially if you hold Korea workplace pensions while living abroad.
• Open the OA PDF before citing the 37% / 81% / 95% figures in a deck; this brief cites abstract-level claims only.

Context

Read this as a pass-through paper about human–AI advice integration, not as proof that AI is a good (or bad) robo-adviser. Korelay frame: the famous story is “people follow ChatGPT”; the useful read is that Korean DC workers in this design transmit recommendation differences substantially but incompletely, changing risk exposure while retaining substantial weight on their own first draft — and that adding a short rationale may not buy you more compliance.

Source

arXiv:2608.11371 — abstract and paper framing cited for briefing; open the OA PDF for design, preregistration, and full results. Not investment advice.