NBER working paper: Investment recommendations from models such as ChatGPT and Gemini are highly concentrated in large-cap tech stocks
A working paper from the U.S. National Bureau of Economic Research (NBER), AI Managed Household Portfolios: A Preliminary Report (No. 35153, April 2026), shows that when large language models are asked to build portfolios aimed at outperforming the S&P 500, their recommendations are highly concentrated in large-cap technology and growth stocks. The study covers models including OpenAI’s ChatGPT 5.0 and 5.2 and Google’s Gemini 2.5 Flash, with daily forward-looking data collected starting in August 2025.
The paper notes that ChatGPT 5.0 typically suggested holding about 20 stocks, with Nvidia (NVDA) weight remaining at 18%–20% over the long term; the median Gemini portfolio contained only about 4.5 stocks. The semiconductor industry accounted for an average of about 41% of AI portfolios, far above the about 20% in the S&P 500. The models favored momentum, large-cap, and low book-to-market stocks, and concentration increased further during active management. The news coverage received by recommended stocks was about nearly 10 times that of ordinary companies, and the study considers media attention to be the main driver. After adjustment for the Daniel et al. (1997) characteristics, these portfolios did not show statistically significant excess returns.
Recent media reports have summarized a similar concentration phenomenon as about 50% of ChatGPT and Gemini recommendations being concentrated in 10 stocks, which is directionally consistent with the concentration findings in this working paper. The study is a preliminary report, data collection is still ongoing, and it is aimed at prompts for ordinary household investors rather than professional refined queries.
Original source:
Title: AI Managed Household Portfolios: A Preliminary Report Publisher: Bruce I. Carlin, Ryan D. Israelsen, Christopher F. Wazzan / NBER Publication date: 2026-04 URL: https://www.nber.org/system/files/working_papers/w35153/w35153.pdf Source type: research or data Verification status: Read the full abstract, introduction, data methodology, and core results sections of the working paper on 2026-11-09 Supporting content: Supports that the models recommended undiversified portfolios, ChatGPT held about 20 stocks with a 18%–20% NVDA weight, Gemini’s median was 4.5 stocks, semiconductors were about 41%, media attention was a driver, and there were no significant excess returns after characteristic adjustment; does not prove the exact single figure of “50% concentrated in 10 stocks” or real-time latest holdings
ABAB AI Insight
Bruce I. Carlin (Rice University), Ryan D. Israelsen (Michigan State University), and Christopher F. Wazzan (University of California, Berkeley) conducted daily prompt experiments on multiple LLMs starting in August 2025, simulating ordinary investors asking them to “build a one-year portfolio that outperforms the S&P 500.” The study covered both buy-and-hold and active management modes, and recorded the webpages accessed by the models and the reasons they gave.
The results show that the concentration was not accidental: during the sample period, ChatGPT consistently allocated 18%–20% to Nvidia, and concentration increased further after active management; Gemini compressed the portfolio to a median of only 4.5 stocks. Semiconductor weight averaged 41%, while traditional sectors such as banking were almost absent. When the models were asked to explain their information sources, the number of news articles about the recommended stocks was about nearly 10 times the average company in Compustat, and fractional logit regression confirmed that media attention was a key explanatory variable.
Similar to human active fund managers, these models chose high-beta stocks (with an average of about 1.6) when unconstrained, and even when beta was restricted to a range, they still tended toward the upper end of that range. The difference is that LLMs lack a traditional risk-management framework, and their “research” depends heavily on public news rather than deep financial modeling. After the paper used DGTW characteristic matching (size, book-to-market, momentum) for adjustment, it found that portfolio returns did not significantly outperform a benchmark passively exposed to these characteristics.
This pattern changes the risk exposure of retail investors using AI: ordinary users may unknowingly receive portfolios highly concentrated in a small number of large tech and semiconductor stocks, and bear the corresponding single-stock and sector risks without obtaining excess returns after risk and characteristic adjustment. A visible indicator is whether concentration declines in later updates as model versions iterate, and the extent to which portfolios deviate from market indexes after actual users adopt them.
ABAB News · Cognitive Law
LLM investment recommendations are highly concentrated in large tech stocks with high media attention, rather than broadly diversified.
Concentration increases further during active management, rather than automatically reducing risk.
After characteristic adjustment, there is no significant excess return; concentration itself is not evidence of stock-picking skill.