In an era of exceptional political partisanship and factionalism, policymakers across the political spectrum seem to agree on one thing: that artificial intelligence (AI) is a revolutionary technology poised to transform our world. From industrial policy to national security to health care, policy decisions are increasingly being made based on the assumption that AI will fundamentally reshape our economy and society. Some policymakers are critical of AI’s effects on the electric grid, kids’ mental health, or other matters of public concern, but they are critics, not skeptics.
Meanwhile, a historically large speculative AI bubble continues to inflate, and policymakers are missing the dangers welling up beneath them. To safeguard against economic catastrophe, policymakers should be confronting and resisting AI industry hype. Instead, even those known for their opposition to Wall Street and Silicon Valley’s self-serving narratives are parroting the notion that an AI revolution is nigh.
Take Senator Bernie Sanders (I-VT), who released a report last fall warning that AI could “replace up to 97 million jobs in 10 years.” The report’s message was that the economic impacts of AI could be much more far-reaching than most people expect. But the key source the report cited for its “97 million jobs” statistic was ChatGPT, which Sanders’s staffers asked to analyze federal job descriptions and estimate which occupations are most vulnerable to replacement. That incident could have been chalked up to overeager committee staffers, except that Sanders again incorporated generative AI into his policy work in March when he released a bizarre YouTube video of him “interviewing” Anthropic’s Claude to explore AI-related data privacy risks.
Sanders’ fears apparently extend beyond a labor apocalypse to an actual apocalypse. In February, he gave a floor speech claiming that “some very knowledgeable people” believe that superintelligent AI — that is, a hypothetical AI system far smarter than any human in every way — could “pose an existential threat to the human race.” Shortly thereafter, Sanders met with one of these “knowledgeable people:” Eliezer Yudkowsky. Yudkowsky’s latest book is titled If Anyone Builds It, Everyone Dies, and he has a longstanding relationship with Peter Thiel, the libertarian tech billionaire and Trump megadonor whose financial backing of Yudkowsky’s depressingly well-funded think tank was a seminal event in the strange, heavily online world that birthed the tech industry’s current obsession with AI.
I do not mean to pick on Sanders, who is far (far, far) from the biggest problem when it comes to AI policy. He undoubtedly views his warnings about massive AI-induced job losses as consistent with his decades of support for workers and sees potential for an alliance of convenience with Yudkowsky, who, like Sanders and Rep. Alexandria Ocasio-Cortez (D-NY), supports a moratorium on data center construction. But Sanders’s flirtations with an AI doom-monger, his willingness to rely on chatbots in formulating policy, and his belief that those same chatbots can replace tens of millions of American workers and perhaps trigger an apocalypse reveals something else: that even policymakers who are skeptical of the tech industry have internalized the industry’s core narrative of imminent AI transformation.
There is indeed a high risk of an AI-driven crisis, but it is one that will result from people overestimating AI’s capabilities and impacts rather than underestimating them. And unfortunately, the policymakers best-positioned to sound the alarm are ignoring the rapidly growing AI bubble because they are too distracted by Silicon Valley’s sci-fi-esque prophecies.
To understand why fears of mass job loss and other catastrophic effects of advanced AI are dubious, it helps to understand how the models that drive modern AI work. University of Washington scientists Carl Bergstrom and Jevin West have aptly dubbed large language models (LLMs), which underpin all of the most advanced AI systems available or in development today, “The Bullshit Machines,” using the technical definition of “bullshit” that philosopher Harry Frankfurt established in a 1986 essay.
What Bergstrom and West mean is that LLMs make statements without regard to whether they are true. Humans who make such statements often know better, or at least they should. LLMs, by contrast, produce bullshit precisely because they do not understand anything about the world, including the difference between what is real and what is not. The one and only thing LLMs do is predict, one small chunk of text at a time, what the user is expecting to hear based on the information in the AI’s training data. The errors (or “hallucinations”) that plague ChatGPT and other LLM-based chatbots are not bugs that can be worked out over time, but an intrinsic and defining feature of machines designed to produce text that is plausible rather than accurate or reliable.
Perhaps the more sophisticated “agentic” AI systems are better in these respects? Not at all. “AI agents” and “agentic AI” are, in reality, simply LLMs that have been linked together or given access to other applications. Rather than catching and correcting errors, such systems often repeat and amplify them.
As a result, LLM-based systems are often counterproductive, if not downright harmful, in settings where accuracy or reliability matter. That is the reason one should never use LLMs when formulating public policy. It is also the reason that fears of an AI-induced job apocalypse are just as speculative as fears of AI-induced extinction.
The basic mechanism through which automation leads to job losses is by increasing productivity; if each worker can do more, companies don’t need as many workers. That is not happening with AI. At the macroeconomic level, there have been no measurable productivity improvements from AI. Zoom in, and the picture is much the same. Several independent studies and surveys show that the vast majority of companies are failing to see any return on their investments in generative AI. This may be because low-quality AI-generated “work slop” frequently slows workers down rather than making them more productive. Last year, a study showed that using generative AI reduced programmers’ productivity by 19% even though the programmers thought it increased their productivity by roughly the same margin — and coding is considered one of the most promising commercial applications of generative AI.
While headlines often feature companies claiming to be laying off workers and replacing them with AI, many (if not most) appear to be instances where companies, particularly those that had over-hired during the pandemic or experienced plummeting share prices, used AI as an excuse for layoffs that they would have made anyway. This phenomenon has become so common that there’s a term for it: “AI-washing.”
Corporations are also using the threat of AI replacement to bully workers into accepting more intensive work, lower pay, and greater precarity. Some companies that claimed to have replaced workers with generative AI later reversed themselves by quietly re-hiring the fired workers or hiring gig workers to perform the work.
Perhaps the companies that engaged in those tactics genuinely believed that AI could replace their workers. Or perhaps they were simply invoking AI as a convenient excuse to convert these positions into more precarious roles. Either way, it is the dominant, hype-driven narrative about AI that enables these tactics — which makes it all the more important for policymakers to reject that narrative.
In fact, the true risk to millions of Americans’ jobs comes not from AI-driven automation but from the economic bubble inflated by AI hype. The U.S. economy has become an enormous unhedged bet on rapid AI transformation. The eight most valuable U.S. companies, which together account for nearly 40% of the S&P 500’s total value, are all tech companies deeply invested in and exposed to the AI bubble. The stock market is more concentrated at the top than it has been since at least the 1920s, more concentrated in a single sector than it has been since the turn of the 20th century, and larger relative to the total economy than it has ever been. With more Americans invested in the stock market than ever before, an AI-driven stock market crash would do immense damage.
The risks extend beyond a stock market crash. Big Tech companies are expected to spend $650 billion on AI infrastructure this year, up from $410 billion in 2025. Increasingly, that spending is financed through massive amounts of debt, particularly from the largely unregulated and already struggling private credit markets. Workers and retirees are now directly exposed to those opaque risks because the Trump administration recently stripped away regulations that protected workers’ retirement accounts from the tentacles of private fund managers.
Crucially, the risk of a crash is high even if AI turns out to be transformative in the long run. AI is currently, in the words of tech writer Ed Zitron, a “cash incinerator,” with all the leading model developers losing billions of dollars per year on the technology. But many AI data centers are financed through loans that will mature within the next few years. A crash is thus increasingly likely even if an AI revolution merely unfolds more slowly than markets expect. And if it never arrives, the fallout from a crash will be much, much worse.
There is a final strategic reason to vocally reject AI hype. By accepting the premise of imminent AI transformation, policymakers narrow the debate to the conclusions that flow from that premise. That tilt favors the tech industry and ensures policy is debated on its terms. Such deference would make sense if the premise had merit, but it does not. Accepting it is therefore unnecessary and counterproductive.
Policymakers should continue to call out the very real harms created by the tech industry’s rush towards AI. But to address the true nature and extent of those harms, policymakers must also explicitly confront and reject the narratives inflating the AI bubble. That means calling bullshit on AI itself.
