Here’s a quick warning for Harrisburg, City Hall, and Main Street. If you think your budgets are challenged now, get ready: The AI crash is coming.

Many an observer has dismissed the idea of yet another tech crash and offered the hopeful claim that “this time is different!” But an AI correction is inevitable.

Will it be this year, or next? It’s impossible to say, since the stock market routinely stays irrational for longer than seems possible. As a friend who lost his shirt short-selling high-flying internet stocks in the late 1990s said: “The market can stay crazy for longer than you can fund your short.”

I have no doubt that generative AI will be revolutionary, just not in the short time frame that bandwagoners believe. Technological revolutions — like the internet itself — are known for delivering less over the short term (say, five years) and more over the long term (say, 20 years) than anyone can foresee.

The early hype on artificial intelligence has been rhapsodically embraced by the media and investors and has fueled one of the biggest technology booms in memory — with soaring valuations, massive spending on dazzlingly advanced chips and data centers, and a rush to build capacity for the voracious electric power needs of AI. This has even spurred a wholly unexpected resuscitation of nuclear power. Some have estimated that well over half of our current GDP growth is coming from this AI-related tech boom.

We can hardly read a newspaper or blog without stumbling over another story about the unending need for yet another multi-billion-dollar data center or the imminent attainment of the nirvana of Artificial General Intelligence. Really? If you buy into this hype, I’ve got a bridge in Manhattan I want to sell you. The truth is that with every great technological surge, over-investment comes first and a correction follows.

AI has a long way to go

Today, generative AI still has a very long way to go, and the gap between actual AI performance and AI hype is colossal. The opening sentence of MIT’s recent STATE OF AI IN BUSINESS 2025 report states that “Despite $30 to 40 billion in enterprise investment into GenAI, this report uncovers a surprising result in that 95 percent of organizations are getting zero return.”

Further, even though increasing numbers of software developers are getting genuine benefits from AI tools, METR, an industry nonprofit whose mission is to develop scientific methods to assess AI risks, reports that “when developers are allowed to use AI tools, they take 19 percent longer to complete issues — a significant slowdown that goes against developer beliefs and expert forecasts.” A key reason is that generative AI models, even the very latest ones, often get things wrong, which requires considerable human oversight to correct.

An AI correction will do no worse than the internet 1.0 crash, and perhaps better, but in no case will bring a deep plunge in GDP such as we saw in the Global Financial Crisis.

Veteran reporter Thomas Claburn of the UK’s widely-read industry gadfly, The Register, is even more critical. He writes that even though some forms of AI have been deeply and usefully embedded in business and manufacturing processes for some time, “AI agents get office tasks wrong around 70 percent of the time.” To date, AI’s primary achievement has been to increase productivity in “low-skill” tasks, such as taking meeting notes or providing customer service. For higher skilled jobs where complexity is vastly increased and accuracy is essential, AIs (even cutting-edge ones) make errors so frequently that the extensive human oversight often makes the entire effort less productive than not using AI at all.

The question is not whether generative AI will be revolutionary. It will. A prominent biotech researcher is using AI to detect large-scale disease patterns in the trillions of cells of the human body and states flatly that this task would simply be impossible without AI. Instead, the question is whether — at this moment — the astronomical AI-related stock values are justified by growth projections for the industry. And the answer is no.

So the question now isn’t if there will be a bust, but when and by how much. The U.S. stock market, including its tech sector, is now valued at an all-time high by almost every measure, and the related factors of margin debt and tech sector debt are at all-time highs as well. Given that a bust is inevitable, it makes sense to examine financial history for clues about how an AI crash might unfold.

Remember the internet bubble?

The most relevant example when thinking about the AI boom is the internet 1.0 crash of 2000 to 2002 that came with the internet revolution of the late 1990s and was accompanied by a frantic telecom and fiber-optic cable boom. In it, the tech-heavy NASDAQ lost 78 percent of its value in less than two years and it took 14 years for that index to attain its previous high. The broader S&P 500 collapsed by 49 percent in that same period (though even with that collapse, it was still above its record level of the late 1990s, which speaks to how overblown valuations had become.)

Remember AOL, Yahoo and Worldcom? Four years after this crash, reportedly 90 percent of the fiber-optic internet networks laid in that boom remained unused — though over the coming years, after having been written off, they provided a cheap and useful avenue for growth.

It may surprise readers that stock market corrections of 10 to 20 percent are relatively frequent, and their aftermath is often short-lived. Five such corrections have happened since the Global Financial Crisis of 2008, and after those, the stock market took only seven months on average to regain its former high. So the question for us is not whether a correction will occur, but whether it will be an “ordinary” correction of 10 to 20 percent, or instead one closer to the disastrous 78 percent the NASDAQ lost in 2000 to 02.

With every great technological surge, over-investment comes first and a correction follows.

That crash was actually a pile-up of crashes: the Internet 1.0 crash from March 2000 to October 2002, the 9/11 2001 terrorist attack-related crash, the Enron scandal and crash from October to December 2001, and the Worldcom crash of July 2002. With this multi-layered stock market crash, short-term interest rates collapsed from 6.5 percent to 1.75 percent, and longer-term rates from 8.5 percent to 6.5 percent, which helped significantly in the recovery.

But the key thing here is that GDP itself — that is, spending and income — held up fairly well despite these multiple shocks. Though economists labeled it a recession at the time, the economy during the internet 1.0 crash never entered a true recession, since real GDP, meaning nominal GDP adjusted for inflation, did not decline for two consecutive quarters, which is the broadly accepted definition of a recession. More importantly, annual GDP growth never declined at all and always remained positive. By comparison, in the Global Financial Crisis of 2008, annual GDP growth plunged by a brutal 3 percent.

Further, unemployment, which had been in the 4 to 4.5 percent range prior to 2001, generally did not exceed the 5 to 6 percent range in this period, and never exceeded 6.3 percent, as compared to the 10 percent unemployment experienced in the Global Financial Crisis.

So the calamitous stock market collapse of 2000 through 2002 was accompanied by a mild and short-lived economic slowdown, reminding us that the stock market is not the economy. That bears repeating: The stock market is not the economy. We need to look past the stock market to deeper economic fundamentals such as the real estate sector to assess how deep a recession might be.

An AI correction will do no worse than the internet 1.0 crash, and perhaps better, but in no case will bring a deep plunge in GDP such as we saw in the Global Financial Crisis. The vital point is that all of history’s worst financial crises have come from the profligate overbuilding of real estate, which dwarfs the tech sector in size and is the economy’s largest non-government sector. In 2008, the real estate sector had been egregiously overbuilt — by some measures it was twice its appropriate size. Today, if anything, the real estate sector is slightly underbuilt and will surge if interest rates decline which will be our deliverance in any recession.

A caveat: There are two factors that were not present in 2002 that may well exacerbate an AI-led slowdown. First, the shutdown in immigration, however wise or unwise from a policy perspective, denies the economy of the significant spending brought by those immigrants. Second, tariffs, which may well serve the national interest longer-term, could make GDP trends worse in the immediate term than they were in 2001 and 2002 since they increase costs to households. But I would nevertheless expect this AI slowdown to be more like 2002 than 2008.

So fasten your seatbelt and take steps necessary to tighten up your own financial situation. To put a useful twist on an old saying: “Caution is the better part of valor.”

MORE ON THE AI TECH BOOM