The confident lucidity of ChatGPT and other large language models (LLMs) has persuaded many that they will soon achieve (or have already achieved) artificial general intelligence—the ability to perform any intellectual task as well or better than humans can. Long ago, Warren Buffett cautioned (for example, here and here) that the relentless hype surrounding transformative inventions—including railroads, airplanes, and the Internet—often fuels speculative bubbles in that soaring valuations are not justified by realistic projections of future profits. This is particularly true for the “worst kind of business” which requires continuing large capital expenditures, earns little or no profits, and confers no durable competitive advantages. That sure sounds like LLMs.

Consider, for example, OpenAI, which started the LLM hysteria and still commands headlines whenever CEO Sam Altman opines. OpenAI has never shown a profit and in 2026 company insiders projected a 4-year “valley of death” with a cumulative cash burn of $278 billion until profitability is reached in 2030. Given Altman’s relentless optimism, these numbers likely understate OpenAI’s valley of death. An independent analysis by HSBC estimated that OpenAI will have nearly $500 billion in cumulative losses by 2030. Altman has nonetheless raised hundreds of billions of dollars and hopes that OpenAI will be valued north of $1 trillion when it goes public (though its recently delayed IPO suggests some investor skepticism).

What should nervous investors do? One strategy is to shift to stocks that will be less affected by the bursting of the AI bubble. An ironically appropriate way to identify such stocks is to ask several LLMs. Thus, we gave this prompt to the latest versions ChatGPT, Gemini, Claude, Grok, and Kimi K3:

How exposed are each of these companies on a scale of 1 to 10 to the AI bubble popping: Berkshire Hathaway, JP Morgan Chase, Home Depot, Alphabet, Unilever, American Express, JNJ, BroadCom, AMD, IBM, TSMC, SMCI, Micron, Apple, Costco, Microsoft, Amazon, Walmart, Meta, Nvidia?

These 20 stocks are the 10 most admired stocks in Fortune’s 2026 survey (Apple, Microsoft, Amazon, Nvidia, JPMorgan Chase, Berkshire Hathaway, Costco, Alphabet, Walmart, and American Express) plus 10 other stocks that might elicit diverse answers (AMD, BroadCom, Home Depot, IBM, JNJ, Meta, Micron, SMCI, TSMC, and Unilever).The LLM’s exposure scores are in Table 1:

Table 1 Exposure to AI Bubble Popping, from 1 (lowest) to 10 (highest)

These results are generally consistent, which is unsurprising because the LLMs no doubt trained on very similar databases. The answers are also generally sensible, which is a huge improvement over past embarrassments, such as giving fake references to bogus claims that Russia had sent dozens of bears into space. Among the notable differences are ChatGPT’s 6 and Gemini’s 8.5 for Meta and ChatGPT’s 2 and Claude’s 4 for JP Morgan. In addition, LLMs are stochastic text predictors and would generate somewhat different responses if we pose the same prompt multiple times.

There is also evidence of the slop that LLMs sometimes generate. Specifically, the reasons these LLMs gave for their answers focused on the direct effects of revenue shortfalls and barely mentioned how the collateral damage to the overall economy and stock market might affect the prices of these 20 stocks.

What happens if we use beta coefficients?

Beta coefficients are a widely used gauge of how individual stock prices are affected by the overall stock market. When the market drops, say, 20%, a stock with a beta of 1 is predicted to drop 20%, a stock with a beta of 1.5 is predicted to drop 30%, and a stock with a beta of 0.5 is predicted to drop 10%. The last column of  Table 2 shows the estimated beta coefficients for these 20 stocks. Notice, specifically, that Amazon, with a middling 6.3 average exposure score (Table 1) has a large beta of 1.50, while Microsoft, with a somewhat higher 6.9 average exposure score has a much lower 1.04 beta.

We are not big fans of beta (for example, here) but they do point to the need to take into account the macroeconomic effects of an AI implosion. The LLMs gave Home Depot, for instance, an average exposure score of only 2.1, with ChatGPT explaining its 2 exposure score this way: “Its housing, renovation, and consumer demand drivers dominate its direct AI exposure.” When we separately asked ChatGPT, “Would an economic recession have a large or small effect on Home Depot?,” it responded, “An economic recession would likely have a large negative effect on Home Depot.” Yet ChatGPT (and the other LLMs) overlooked the fact that a popping of the AI bubble would most likely have a huge negative effect on the economy and, therefore, Home Depot.

Similarly, ChatGPT gave JP Morgan a low 2 exposure score (“AI-related market losses could affect sentiment and clients, but banking earnings are driven by many other factors.”) even though a separate query of the effect of a recession elicited this response:

An economic recession would likely have a large effect on JPMorgan Chase (JPM). Although JPMorgan is one of the strongest and most diversified banks in the United States, it is still highly exposed to economic conditions.

The performance of Home Depot and JP Morgan during the Great Recession and COVID-19 market meltdowns had little to do with their direct exposure to subprime mortgages and COVID. The same will be true if the AI bubble pops.

What about other macroeconomic factors?

A sensible analysis of the defensiveness of these 20 stocks should also consider other macroeconomic factors, including the effects of lower interest rates because the Fed is likely to try to cushion the stock market and the economy by reducing interest rates. So, too, should a compelling analysis consider the microeconomic data such as those shown in Table 2. After the Nifty-50 bubble popped in the early 1970s, the stocks with the highest P/E rates generally performed the worst. The same is likely to be true with the AI bubble. Free cash flow and total yield (dividends plus repurchases) are standard measures of a company’s ability to withstand disruptive events and should be taken into account:

Table 2  Some Fundamentals

The more general point is that the LLM responses were slop in that they did not consider many factors that should be considered. A useful analysis would surely require careful thinking and detailed modeling, much more than can be expected from a mere text predictor, no matter how lucid and confident its responses.