How serious would an AI bubble be?
The following contribution is from The Atlantic, one of the most prestigious news and general information outlets in the United States.
The author is Rogé Karma, who is part of the editorial team.
The entire US economy is being driven by the promise of productivity gains that seem far from being realized.
If there’s one field where the rise of AI is said to be making humans obsolete—where the dawn of superintelligence is already upon us—it’s programming. This makes the results of a recent study truly surprising.
In the study, published in July, the think tank Model Evaluation & Threat Research randomly assigned a group of experienced software developers to perform coding tasks with or without AI tools.
It was the most rigorous test to date of AI’s real-world performance. Because coding is one of the skills largely mastered by existing models, nearly all participants expected AI to generate huge productivity gains.
In a pre-experiment survey of experts, the median prediction was that AI would speed up developers’ work by nearly 40%.
Afterward, study participants estimated that AI had sped them up by 20%.
But when the METR team analyzed employees’ actual work performance, they found that developers completed tasks 20% slower when using AI than when working without it.
The researchers were stunned. «No one expected that result,» Nate Rush, one of the study’s authors, told me. «We didn’t even consider a potential slowdown.»
No single experiment should be considered the final word. But the METR study is, according to many AI experts, the best we have, and it helps to understand an otherwise paradoxical moment for AI.
On the one hand, the United States is experiencing an extraordinary economic boom driven by AI: the stock market is soaring thanks to the sky-high valuations of AI-related tech giants, and the real economy is being boosted by hundreds of billions of dollars in spending on data centers and other AI infrastructure.
Underpinning all this investment is the belief that AI will make workers vastly more productive, which in turn will boost corporate profits to unimaginable levels.
On the other hand, there is growing evidence that AI is not delivering the expected results in the real world.
The tech giants that invest the most in AI are far from recouping their investments.
Research suggests that companies trying to incorporate AI have seen virtually no impact on their bottom lines. And economists looking for evidence of job losses replaced by AI have found virtually nothing.
None of this means that AI cannot, over time, be as transformative as its biggest proponents claim. But, over time, it could take a long time. This raises the possibility that we are experiencing an AI bubble, in which investor enthusiasm has far outweighed the short-term productivity benefits the technology offers. If this bubble bursts, it could dwarf the dot-com crash, and tech giants and their Silicon Valley backers won’t be the only ones hurt. Almost everyone agrees that programming is the most impressive use case for AI technology today. Prior to its most recent study, METR was known for a March analysis showing that the most advanced systems could handle programming tasks that would take a typical human developer nearly an hour to complete. So how might AI have reduced developer productivity in its experiment?
The answer has to do with the «capacity-reliability gap.»
Although AI systems have learned to perform an impressive array of tasks, they struggle to complete them with the consistency and accuracy demanded by real-life situations.
The results of the March METR study, for example, were based on a 50% success rate, meaning the AI system was only able to complete the task reliably half the time, rendering it virtually useless on its own.
This shortcoming hampers the use of AI in the workplace. Even the most advanced systems make small mistakes or slightly misinterpret instructions, requiring a human to carefully review their work and make necessary changes.
This appears to be what happened during the most recent study.
Developers ended up spending a significant amount of time reviewing and reworking the code the AI systems had produced, often more time than it would have taken them to write it themselves.
One participant later described the process as the «digital equivalent of looking over the shoulder of an overconfident junior developer.»
Since the experiment was conducted, AI coding tools have become more reliable.
The study focused on expert developers, while the greatest productivity gains could come from enhancing or replacing the skills of less experienced workers.
But the METR study could easily be overestimating the productivity benefits associated with AI.
Many cognitive labor tasks are harder to automate than coding, which benefits from massive amounts of training data and clear definitions of success.
«Programming is something that AI systems typically do extremely well,» Tim Fist, director of Emerging Technologies Policy at the Institute for Progress, told me.
«So, if it turns out they’re not even making developers more productive, that could dramatically change the landscape of how AI could impact overall economic growth.»
The gap between capability and reliability could explain why generative AI has so far failed to deliver tangible results for the companies that use it.
When MIT researchers recently analyzed the results of 300 publicly disclosed AI initiatives, they found that 95% of the projects failed to increase profits.
A March report from McKinsey & Company found that 71% of companies reported using generative AI, and more than 80% of them reported that the technology had no «tangible impact» on profits. Given these trends, Gartner, a technology consultancy, recently declared that AI has entered the «disappointment» phase of technological development.
Perhaps AI’s progress is just experiencing a temporary speed bump.
According to Erik Brynjolfsson, an economist at Stanford University, every new technology experiences a «productivity J-curve»: Initially, companies struggle to implement it, causing productivity to drop.
However, over time, they learn to integrate it, and productivity soars.
The classic example is electricity, which became available in the 1880s but didn’t begin generating large productivity gains for businesses until Henry Ford reinvented industrial production in the 1910s. Some experts believe this process will unfold much faster for AI.
«With AI, we’re in the initial, downward phase of the J-curve,» Brynjolfsson told me.
«But by the second half of the 2020s, it will really take off.» Dario Amodei, CEO of Anthropic, has predicted that by 2027, or «not much later,» AI will be «better than humans at almost everything.»
These forecasts assume that AI will continue to improve as rapidly as it has in recent years.
This is not a given. Newer models have been plagued by delays and cancellations, and those released this year have generally shown fewer significant improvements than earlier models, despite being much more expensive to develop.
In a March survey, the Association for the Advancement of Artificial Intelligence asked 475 AI researchers whether current approaches to AI development could produce a system that matches or surpasses human intelligence; more than three-quarters responded that it was «unlikely» or «very unlikely.»
OpenAI’s latest model, GPT-5, was released early last month after nearly three years of work and billions of dollars in investment. (The Atlantic signed a corporate partnership with OpenAI in 2024.) Before its launch, CEO Sam Altman stated that using it would be equivalent to having «a PhD expert in any field» on hand.
In some areas, including programming, GPT-5 represented a significant leap forward. However, based on the most rigorous measurements of AI performance, GPT-5 turned out to be, at best, a modest improvement over previous models.
The prevailing view in the industry is that it’s only a matter of time before companies find the next way to drive AI progress. This could be true, but it’s not guaranteed.
Generative AI wouldn’t be the first tech fad to experience a wave of hype.
What’s distinctive about the current situation is that AI appears to be driving the US economy as a whole.
More than half of the S&P 500’s growth from 2023 onward will come from just seven companies: Alphabet, Amazon, Apple, Meta, Microsoft, Nvidia, and Tesla. These companies, collectively known as the Magnificent Seven, are considered especially well-positioned to prosper from the AI revolution.
That prosperity has yet to materialize anywhere except in their stock prices. (The exception is Nvidia, which provides the crucial inputs—advanced chips—that the rest of the Magnificent Seven are buying.)
According to The Wall Street Journal, Alphabet, Amazon, Meta, and Microsoft have seen their free cash flow decline 30% over the past two years.
By one estimate, Meta, Amazon, Microsoft, Google, and Tesla will have invested $560 billion in AI-related capital investments by the end of this year since the beginning of 2024 and will have generated only $35 billion in AI-related revenue.
OpenAI and Anthropic are generating large revenues and growing rapidly, but they are still far from profitable.
Their valuations—roughly $300 billion and $183 billion, respectively, and rising—far exceed their current revenues. (OpenAI projects about $13 billion in revenue this year; Anthropic, between $2 billion and $4 billion.)
Investors are betting heavily that all this spending will soon generate record profits.
However, if that belief collapses, investors could start selling en masse, triggering a dramatic and drastic correction in the market.
During the internet revolution of the 1990s, investors poured money into virtually every company with a «.com» in its name, convinced that the internet was about to revolutionize business.
However, by 2000, it became clear that companies were wasting their money without making much in the way of profits, and investors responded by dumping the most overvalued technology stocks.
From March 2000 to October 2002, the S&P 500 fell nearly 50%. Over time, the internet transformed the economy and gave rise to some of the most profitable companies in human history. But that didn’t stop many investors from going bankrupt.
The dot-com crash was severe, but it didn’t trigger a crisis. The collapse of the AI bubble could be different.
AI-related investments have already surpassed the level reached by telecommunications at the peak of the dot-com boom as a percentage of the economy.
In the first half of this year, business spending on AI contributed more to GDP growth than all consumer spending combined.
Many experts believe that one of the main reasons the US economy has been able to weather tariffs and mass deportations without a recession is that all this AI spending acts, in the words of one economist, as a «massive private-sector stimulus program.»
An AI collapse could lead, overall, to lower spending, fewer jobs, and slower growth, potentially dragging the economy into a recession.
Economist Noah Smith argues that it could even trigger a financial crisis if the unregulated «private credit» lending that finances much of the industry’s expansion goes bust at once.
Rogé Karma: Does the stock market know something we don’t?
If we do end up in an AI bubble, the silver lining would be that fears of a sudden AI-driven job displacement are overblown.
In a recent analysis, economists Sarah Eckhardt and Nathan Goldschlag used five different measures of AI exposure to estimate how the new technology might be affecting a range of labor market indicators and found virtually no effect on any of them.
For example, they note that the unemployment rate for workers least exposed to AI, such as construction workers and fitness trainers, has risen three times faster than that of workers most exposed, such as telemarketers and software developers.
Most, but not all, other studies have reached similar conclusions.
But there is also a more unusual, intermediate possibility. Even if AI tools don’t increase productivity, the hype surrounding them could drive companies to continue expanding their use.
«I hear the same story over and over again from companies,» Daron Acemoglu, an economist at MIT, told me.
«Middle and senior managers are told by their bosses that they need to use AI for X percent of their work to satisfy the board.»
These companies might even lay off workers or slow down their hiring because they’re convinced—like the software developers in the METR study—that AI has made them more productive, even when it hasn’t.
The result would be an increase in unemployment that would not be offset by real improvements in productivity.
Although it seems unlikely, a similar scenario occurred in the not-so-distant past. In his 2021 book, «A World Without Email,» computer scientist Cal Newport points out that, beginning in the 1980s, tools such as computers, email, and online calendars allowed knowledge workers to manage their own communications and schedule their own meetings.
In turn, many companies decided to lay off their secretaries and typists. In a perverse result, the most skilled employees began spending so much time emailing, writing notes, and scheduling meetings that they became far less productive at their actual work, forcing companies to hire more employees to accomplish the same amount of work.
A subsequent study of 20 Fortune 500 companies found that those with computer-induced «staffing imbalances» spent 15% more on salary than necessary.
«Email was one of those technologies that made us feel more productive, but it actually had the opposite effect,» Newport told me.
«I worry that we’re going down the same path with AI.»
On the other hand, if the alternative is a stock market crash that precipitates a recession or financial crisis, that scenario might not be so bad.
This project received support from the William and Flora Hewlett Foundation.
Tech guru Erik Gordon says investors will suffer far more from the rise of AI than from the dot-com crash.
The following contribution is from the prestigious Business Insider website and is written by Theron Mohamed, a correspondent on Business Insider’s Trends team, based in London. His coverage covers finance, investing, wealth, markets, and economics.
Theron joined Business Insider in 2019 as a Markets Insider reporter and rose to correspondent before joining the Trends team in 2024. Previously, he covered technology, media, and telecom stocks for Investors Chronicle magazine and briefly contributed to the Financial Times’ Data team. He interned at the Wall Street Journal in New York, where he primarily wrote for Heard on the Street.
Business professor Erik Gordon commented on Pets.com, which went bankrupt during the dot-com crash: «The loss was minimal compared to what we might see with AI.» The AI boom will fail and dwarf the dot-com crash due to its larger scale, Erik Gordon said.
AI startups like CoreWeave threaten greater losses for investors than companies like Pets.com, he added.
The business professor has previously warned that AI is an «overvaluation bubble of considerable magnitude.»
The fall in one company’s stock shows how the financial consequences of a stalled AI boom will be far greater than those of a dot-com bust, tech guru Erik Gordon told Business Insider.
Gordon, an entrepreneurship professor who researches financial markets and technology at the University of Michigan’s Ross School of Business, has previously called the AI boom an «overvaluation bubble of considerable magnitude.»
Some investors have claimed that tech stocks soaring on AI optimism will collapse like the dot-com companies of the early 2000s, but others, like Kevin O’Leary, have dismissed the comparison.
