Warning: the AI Theme is Officially Dead
The bull market that began in mid-2022, and which accelerated starting in late-2023, has been largely driven by the Artificial Intelligence (AI) revolution.
Since late-2023, AI-related companies have accounted for 75% of S&P 500 market gains, 80% of corporate profits, and 95% of capital expenditures. To be clear, this is largely a Big Tech story: profits at the MAG-7 companies (Nvidia, Microsoft, Apple, Meta, Amazon, Alphabet, and Tesla) have doubled in the last three years. And thanks to investor enthusiasm pertaining to the potential of AI, these companies’ market capitalizations have tripled driving the S&P 500 higher (the MAG-7 are the largest, most profitable companies in the index).
Again, much of these results have been driven by the perceived potential of AI. But unfortunately, this potential may not actually manifest any time soon, if at all.
Let me explain…
From a philosophical perspective, the potential of AI hinges on whether this technology will ever be able to actually “think,” as opposed to simply organizing material and presenting it based on patterns/ probabilities.
If AI can actually “think,” then it is indeed a potential boon to corporate productivity and profits. However, if AI cannot think, but is only capable of organizing material in a cohesive manner, then it has major issues. Specifically, AI will not “know” whether its answers are correct or wrong, nor will it be able to come up with creative solutions to the problems users present it with.
I want to emphasize that this debate has been happening at the highest levels of the AI-technological revolution. Yann LeCun is a Turing award winner (the Turing award is considered the “Nobel Prize” of computing) who ran Meta’s AI Research Lab from 2013 to 2025. According to reports, he and META Founder/CEO Mark Zuckerberg were at odds as to whether AI-based Large Language Models (LLMs) would ever be able to reason like humans. Zuckerberg believed they will. LeCun believed they would not. And in 2025, LeCun left Meta, suggesting that he has decided Meta is going in the wrong direction as it pursues a goal of “thinking” AI.
I bring this up because the recent spate of data concerning the actual efficacy of AI models suggests LeCun is correct: AI cannot think nor can it come up with creative solutions.
Major Problem #1: AI Hallucinates
One of the issues emerging in the AI story is the presence of AI “hallucinations.”
LLMS work by predicting the most statistically plausible next word or phrase given a prompt, based on patterns learned from training data. They don’t have a fact-checking mechanism built in, meaning they are not “looking things up” in a verified database. So, when asked about something obscure, ambiguous, or outside their training, they tend to generate answers that are fluent, confident-sounding but wrong.
Some of the more common hallucinations include:
- Fabricating citations — inventing plausible-sounding academic papers, authors, or URLs that don’t exist
- Getting biographical details wrong about real people
- Misremembering statistics or dates
- Inventing legal cases, historical events, or product specifications
- Confidently describing a process incorrectly
As the below table illustrates, this issue is endemic in LLMs with hallucination rates ranging from 15% to 52%!
In this context, the only people who could use AI effectively in a corporate setting would be those individuals who are experts are on the subject matter they are discussing with the AI model. After all, these would be the only people capable of determining when an LLM is providing a sound insight/ idea as opposed to fabricating key aspects of its answer!
This raises countless issues as to the true effectiveness of AI. If it is simply organizing information as opposed to “thinking” it isn’t actually providing insights or solutions, but simply arranging information in a pattern that seems as if it makes sense.
Unfortunately, this is not the only major issue pertaining to AI models.
Major Problem #2: AI Is NOT Creative
Researchers at Stanford recently performed a study through which they asked multiple AI models open ended queries (questions that do not have a single correct answer). And they didn’t do this once or twice, but 26,000 times.
Initially, the researchers expected two things:
- A specific AI model to come up with different answers each time it was asked an open-ended query.
- Different AI models to come up with different answers based.
Instead, they found the exact opposite happened:
- An AI model would answer the same answer to an open-ended query every time.
- Different AI models ended up coming up with the same solutions.
The paper concluded that AI actually suffers from “Hive Mind” or homogenous thinking. And most worrisome this was the case across numerous AI models. Put simply, instead of providing diverse, creative answers, AI, even across different models, was quite boring and repetitive.
This is a MAJOR problem. Remember, the MAG-7 hyperscalers are spending upwards of $400 BILLION per year building out AI-infrastructure. But thus far, this technology is proving that A) it can’t think and B) it’s “thinking” is mundane, repetitive, and rife with hallucinations.
And this is the technology that is supposed to boost corporate productivity while reducing corporate profits!?!?
Given the significance of AI to this bull market (again, AI-related companies have accounted for 75% of S&P 500 market gains) it’s only a matter of time before the stock market begins to discount that AI, at least in its current form, isn’t going to be generating anywhere near the perceived potential investors are hoping for.
This opens the door to a MAJOR repricing of AI-related stocks to the downside. If stocks simply reprice based on this risk, we can expect a 10% correction to unfold bringing the S&P 500 down to the 6,300s.
But if AI proves to be something of a dud while the U.S. economy experiences another wave of inflation (thereby negating any hope of additional Fed easing) or worse still, a recession, then we can expect a full-scale bear market with the S&P 500 declining to the mid-to low-5,000s.
In terms of preparing for a market collapse, I rely on a proprietary indicator that has triggered before every major meltdown in the last 50 years. This signal caught the 1987 crash, the Tech Crash, the Great Financial Crisis and more.
We detail this trigger, how it works, and what it’s saying about the markets today in How to Predict a Crash.
Normally we’d sell this report for $499, but in light of its recent warning, we’re making 99 copies available to the investing public.
To pick up one of the last copies…
Graham Summers, MBA
Chief Market Strategist
Phoenix Capital Research