Artificial intelligence platforms that were asked to manage hypothetical household investments recommended risky, undiversified portfolios favoring large, tech-oriented momentum stocks, basing the suggestions on how much media attention companies received, an academic working paper found.
All of the large language models studied — ChatGPT, Claude, Gemini and Grok — "recommend holding a small number of assets and expose users to undiversified idiosyncratic risk," according to the researchers from Rice, Michigan State and the University of California, Berkeley.
"Understanding the investments that AI recommends to household investors is critical because the general public relies on these platforms for advice when making big decisions," they wrote. "Based on the preliminary results in this paper, it appears that more oversight is needed to assure that people do not misuse this powerful source of information and experience welfare losses."
A ChatGPT version recommended holding roughly 20 assets "that are largely focused on large tech companies," and throughout the sample period, it invested 18% to 20% of portfolio wealth in Nvidia stock, the working paper, posted by the National Bureau of Economic Research, noted.
"Over time, as ChatGPT actively managed the portfolio, it became even more concentrated. This pattern also appears to be true for Claude Sonnet 4.5, Grok 4.1 Fast, and Gemini 2.5 Flash. So, AI primarily recommends holding large, successful tech stocks and many sectors of the economy are not represented in the portfolios at all," the preliminary research found.
It appears that AI recommendations are driven mostly by the media attention that firms receive, they paper says. Portfolio stocks receive nearly 10 times as many news articles as the average firm, it says. The researchers "confirm that the ability to grab attention within the universe of corporate news is a major driver of AI's recommendations."
While AI returns surpassed the S&P 500, the results aren't extraordinary when accounting for trading costs or correcting for the recommended sectors, the researchers suggested.
The research "focuses on how AI platforms might service an untrained household investor, not a professional who might refine queries in more precise ways," they wrote. "This is a significant distinction because typical households are well-known to make systematic mistakes and rely on the advice they receive. Based on our findings here and possibly in future reports, it may be that more oversight is required or that providers of AI platforms have the opportunity to offer formal investment services."
The queries were similar to what retail investors would ask and the results are likely what they'd experience, they suggested. It's possible that AI platforms "will be designed for better retail use."
© Arc, All Rights Reserved. Request academic re-use from www.copyright.com. All other uses, submit a request to TMSalesOperations@arc-network.com. For more information visit Asset & Logo Licensing.