At a glance
- The assistant answered questions about linked brokerage and retirement accounts at Mezzi, a US wealth app and SEC-registered investment advisor that added a tool built on OpenAI's ChatGPT and Google's Gemini models in November 2024.
- Single-stock purchases fell 5.9 percentage points in the same estimate.
- Advisors reviewing a client who acts on that answer see total-market beta concentrated in a single ticker.
Retail investors who adopted an AI assistant lifted index funds' share of their monthly purchases by 7.6 percentage points, according to an NYU Stern working paper built on 69,598 prompts. The assistant answered questions about linked brokerage and retirement accounts at Mezzi, a US wealth app and SEC-registered investment advisor that added a tool built on OpenAI's ChatGPT and Google's Gemini models in November 2024.
Single-stock purchases fell 5.9 percentage points in the same estimate. Before adoption, index funds accounted for about 33% of monthly purchases. The paper, "Generative AI and Retail Investment" by NYU finance PhD student Valerie Baldinger, uses a staggered difference-in-differences design comparing about 1,900 investors who adopted the assistant at different times. The assistant does not execute trades; investors decide whether to act on an answer. The estimate may not carry to tools that can place orders.
Advisors reviewing a client who acts on that answer see total-market beta concentrated in a single ticker. VTI, Vanguard's total-market ETF, was the most recommended ticker in six of the seven ChatGPT and Gemini models studied. Within Lipper's Multi-Cap Core category, VTI took 53.9% of investor ticker recommendations, against 15.8% of the category's assets, so the models named it more than three times as often as its asset share would imply. The remaining questions are which model version made the recommendation and what it displaced in the portfolio.
Investors rarely asked for a fund. They expressed interest in buying an ETF or mutual fund in 3.5% of prompts, while the assistant recommended a fund in 18% of its answers, about five times as often. When a prompt named no ticker and did not mention stocks, 89.7% of the assistant's buy recommendations were for ETFs or mutual funds.
The shift lowered annual return loss by 43 basis points from a pre-adoption average of about 407 basis points. Return loss is the expected return an investor forfeits by carrying risk the market does not pay for because it can be diversified away. For the less-diversified half of investors it fell 101 basis points; for the more diversified half the effect was not statistically significant.
The paper is Baldinger's job market paper and has not been peer-reviewed. She served as an unpaid contractor for Mezzi and reports no financial interest in the company. Mezzi has promoted an earlier version of the research on its website. The population is app users who chose to link accounts, which limits how far the 7.6-point figure travels beyond that sample.
AI models pitched VTI at 3.4 times its asset share
VTI's share of ticker recommendations vs its share of Lipper Multi-Cap Core assets, %
Share of recommendations
53.9%
- Most-recommended ticker in
- 6 of 7 models
Share of category assets
15.8%
- Recommendations ÷ assets
- 3.4x
Show as table
| Share of recommendations | Share of category assets | |
|---|---|---|
| Figure | 53.9% | 15.8% |
| Most-recommended ticker in | 6 of 7 models | |
| Recommendations ÷ assets | 3.4x |
Reported by the Passive & Indexing desk from ETF Trends, NYU Stern working paper, "Generative AI and Retail Investment" by Valerie Baldinger and NYU Stern working paper, as reported in the ETF Daily data pack. Published . Updated Nov 19, 2026 · 12:19 p.m. ET. Corrections