AI Stock Picks Overweight Big Tech and Semiconductors, Raising Potential Risks for Investors

Key Takeaways

• AI-generated portfolios tend to be highly concentrated in large-cap tech stocks, especially semiconductor companies.

• Researchers found that large language models leaned more on media coverage than company fundamentals when making stock picks.

• That media-driven approach produced less-diversified portfolios and could leave investors more exposed to losses in crowded sectors.

Investors turning to AI chatbots for stock picks may end up with portfolios heavily concentrated in Big Tech and semiconductor stocks—without gaining a meaningful edge over the market, according to a new National Bureau of Economic Research working paper.

In the study, the researchers asked different large language models (LLMs)—including ChatGPT 5.0, Grok 4.1 Fast, Claude Sonnet 4.5 and Gemini 2.5 Flash—to build two kinds of investment portfolios. One was passively managed, meaning the AI picked stocks and held them. The other was actively managed, meaning the AI could change its picks daily.

The resulting portfolios were aggressive and heavily concentrated in Big Tech, with Nvidia (NVDA) a near-universal pick. While many investors have been putting their money into these stocks, the NBER researchers noticed that LLMs were driven less by fundamental analysis than by the volume of media coverage.

What This Means For You

Many investors have been looking to AI for portfolio advice. But the new research suggests LLMs may have inherent biases for certain types of information that could expose investors to more risk than they assume.

According to the study, the AI models’ recommendations were primarily based on the amount of media coverage a company received, with recommended stocks having nearly 10 times as many news articles about them as the average Compustat company. (Compustat is an S&P database that collects information on public companies in the U.S.)

“The ability to grab attention within the universe of corporate news is a major driver of AI’s recommendations,” the researchers wrote.

The sites the LLMs used to gather information were primarily corporate websites, specifically those of semiconductor and other major tech companies. About two-thirds of the LLM visits were to company websites, with the remainder going to news services, government organizations and sites that provide stock analysis, the study said.

Semiconductor stocks made up an average of 41% of AI portfolios, about double their 21% share of the S&P 500 index, the NBER study found. That’s a problem for regular investors. Concentration like that could amplify losses if the sector turns, and the loudest names in the news cycle tend to be the ones that investors have already run up—it’s why they’re in the headlines—meaning you could be buying near the top.

And while the LLMs’ portfolios did yield higher returns than the market as a whole, beating the S&P 500, the researchers found that these returns were not outsized when accounting for trading costs or their bias toward tech stocks.

“AI takes risk, recommends a narrow set of assets, focuses on specific industries, and does not appear to exhibit better performance than passive characteristic-based benchmarks,” wrote the researchers.

On top of that concentration, the AI portfolios were built from unusually volatile stocks. Without any restrictions, the models picked stocks with an average market beta of 1.6, meaning the portfolios would swing about 60% more than the S&P 500 index in either direction. Even when the researchers asked the models to match the S&P 500’s level of volatility, the portfolios still hugged the top of the allowed range, and sometimes broke through it.

Gemini provided the least diversified portfolios, holding a median of just 4.5 stocks, compared with 10 for Claude and Grok, the researchers said. And rather than diversifying further, the portfolios grew narrower as the models actively managed them over time.

The paper is preliminary, the researchers caution, and the Claude, Gemini and Grok results are based on only about four weeks of trading data. They also note that the prompts in the study reflect what a typical retail investor would type, not the more refined queries a professional might use to push the models harder.

Read the original article on Investopedia