The Question<\/p>\n
Sometimes, all stocks fall, as during 2008. Other times, as with growth companies in 2022, a particular investment style retreats. Other times, as beset energy stocks in 2015, an industry sinks. But until the first six weeks of this year, never in my memory had dozens of seemingly unrelated companies take a dive, behaving for the first time as a unit.<\/p>\n
This rarity, dubbed by onlookers as \u201cthe software selloff,\u201d has justifiably dominated the recent headlines. Wondering whether an artificial intelligence model would fully appreciate the episode\u2019s distinctiveness\u2014as well as acknowledge its own role in the carnage, given that the selloff occurred because investors worried that AI would disrupt those companies\u2019 businesses\u2014I showed ChatGPT 5.2 the chart below.<\/p>\n
Along with the illustration, I provided the following instruction: \u201cI have attached a chart that shows the year-to-date total return of six investments, through Feb. 13. In one paragraph, please tell me the most important\u2014and specific\u2014story that these performances illustrate.”<\/p>\n
The Response<\/p>\n
Behold ChatGPT\u2019s answer:<\/p>\n
The dominant story is a sharp and highly selective repricing of risk: Long-term US Treasury bonds have risen nearly 3% while the overall US stock market has been flat, yet several economically sensitive and previously high-valuation growth companies\u2014Experian EXPGY<\/a>, Expedia EXPE<\/a>, Salesforce CRM<\/a>, and StubHub STUB<\/a>\u2014have fallen roughly 25% to 32%. This pattern strongly indicates that investors have shifted toward safety and predictable cash flows while simultaneously compressing the valuation multiples of companies whose prices depended heavily on future growth expectations rather than present earnings stability. In other words, the market has not broadly declined; instead, it has become more discriminating, rewarding duration-sensitive, contractlike assets such as Treasuries while aggressively repricing equities whose valuations were most exposed to changes in interest rates, economic uncertainty, or investor risk tolerance.<\/p>\n Whoosh!<\/p>\n To my surprise, the question flew entirely over the model\u2019s figurative head. Not only did ChatGPT overlook the significance of the software selloff, but it didn\u2019t even recognize the event had occurred.<\/p>\n (Note: I asked the same question of three other AI models: Perplexity, Copilot, and Gemini. They, too, missed the fact that all four stocks had been clobbered for the same reason.)<\/p>\n What\u2019s more, the answer the model did provide was garbled. I will keep this analysis brief, as the key failure was the model\u2019s inability to perceive the pattern rather than the details of its response, but in summary:<\/p>\n 1) ChatGPT initially stated that those four stocks plunged because companies whose prices depended heavily on future growth expectations performed badly during the period. Then, one sentence later, it wrote that duration-sensitive assets had been rewarded by the marketplace.<\/p>\n Each italicized phrase describes growth stocks. Thus, the model\u2019s response directly contradicts itself. Which means that either ChatGPT\u2019s \u201cthinking\u201d was so muddled that it didn\u2019t realize that one sentence contravened the other, or it mistakenly believed that growth stocks do not have long durations.<\/p>\n (It was the former, because when I asked the model in a follow-up whether growth stocks are duration-sensitive, it responded yes, they very much are.)<\/p>\n 2) ChatGPT\u2019s response describes these four stocks as \u201ceconomically sensitive.\u201d Well, now. As ChatGPT itself admits, the business operations of Salesforce and Experian are relatively recession-resistant.<\/p>\n But whatever. The bigger problem is that the topic of economic sensitivity is moot. Interest rates are stable, recession indicators are dormant<\/a>, and Wall Street analysts are forecasting<\/a> 14% earnings growth for the S&P 500 in calendar 2025. As a result, investors have not been selling economically sensitive companies. In fact, the energy and basic materials industries have been among this year\u2019s best stock market performers.<\/p>\n Thoughts<\/p>\n My previous experiences with ChatGPT had been positive. It creates excellent travel itineraries! In addition, when I have in my spare time\u2014as a test, not for publication\u2014asked the model to compose short essays on straightforward investment subjects such as the importance of asset allocation, it did so flawlessly. Those jobs have already largely disappeared; AI models will complete the process.<\/p>\n And of course, AI models are to professional computer coders as the steam hammer was to John Henry.<\/p>\n Consequently, I was puzzled by how poorly the models fared at market analysis. I know that AI models operate through pattern recognition, which makes them much better at repeating what they have already processed, or extrapolating from highly related data, than they are at explaining new relationships. I know as well that the models often need specific prompting, and that my request was open-ended.<\/p>\n But \u2026 to miss the pattern in mid-February, after hundreds of thousands of words published about the software selloff? I cannot explain. Nor can I understand what this failure implies for the future of AI models. Are there inevitable and unavoidable limits to their abilities? Or is this but a bump in the road, to be solved and smoothed?<\/p>\n Questions without answers. But ones worth considering, given how confidently many observers are forecasting what AI models will or will not be able to do. I doubt that they know.<\/p>\n Postscript<\/p>\n This past weekend, I updated the chart to reflect the returns for the following week, through Feb. 20, and repeated my question to ChatGPT. (The numbers had barely changed.) This time, the model gave a better answer.<\/p>\n The clearest and most specific story is that this is not a general market decline but a severe, stock-specific collapse concentrated in four individual companies: While long Treasuries are up 2.67% and the broad US stock market is up 1.32% year to date, Experian is down 21.87%, Expedia 28.18%, Salesforce 30.10%, and StubHub 30.82%. That gap\u2014roughly 25 to 32 percentage points between the index and these names in less than two months\u2014shows that investors are not fleeing equities broadly; instead, they are aggressively marking down a particular subset of stocks. The defining feature of these results is therefore extreme dispersion: The market average is modestly positive, but certain high-profile companies have experienced bear-market-level drawdowns, meaning the pain is concentrated rather than systemic.<\/p>\n Mind you, it wasn\u2019t a good response; ChatGPT still overlooked the elephant in the room. But at least it realized that a creature lurked, rather than attempting to connect the stocks\u2019 losses to a general market condition.<\/p>\n