The Bosses’ Oldest Dream

The English inventor Charles Babbage is best remembered for his work designing the foundations for all digital computers with Ada Lovelace in the nineteenth century. What’s lesser known is why he wanted to build them. In the decades leading up to the British Empire’s abolition of slavery in the West Indies in 1833 (which Babbage supported), there was a serious question of how to get free and rebellious industrial workers to produce enough to sustain the empire. Babbage’s answer was to surveil workers and divide their efforts into standardized, measurable tasks — a project he saw as a necessary step on the way to automating labor entirely.

Babbage didn’t invent these methods, as Meredith Whittaker, an artificial intelligence scholar who coled Google’s 2018 walkout, writes in a Logic(s) essay. Similar technologies of labor control were developed first on plantations to extract maximum labor from enslaved people, with management guides routinely circulating among British capitalists. “Babbage’s proto-Taylorist ideas on how to discipline workers are inextricably connected to the calculating engines he spent his life attempting to build,” Whittaker writes. Automation gave bosses a tool to replace striking workers and provided a disciplinary check, in Babbage’s words, “against the inattention, the idleness, or the dishonesty of human agents.”

There’s a neat alpha and omega quality to computing: its ultimate effort, the Obsoleting Project — the industry’s drive to build machines that can do all our jobs — rhymes with its earliest. Should they work as intended, those machines would act as the tireless, pliant, cheap workers that bosses could previously only dream of. From Babbage to Bezos, managers have always resented the various ways that humans aren’t optimized for output — and therefore, their bottom line.

For bosses, cost savings alone are only part of the attraction. Automation also weakens workers’ bargaining power and is used to make organizing even harder. After workers alleged Amazon used algorithmic management to surveil and stifle protected labor organizing, Stanford Law Lecturer Seema N. Patel told The Guardian that AI lets employers discipline hundreds of thousands of workers “in a way that no human manager or group of managers could even do.”

“You feel like you’re in prison,” Wendy Taylor, an Amazon packer, said. “They know every move you make, when you’re working, when you’re not working.”

Of course, Amazon is also leading the charge on replacing workers entirely.

Yoshua Bengio is the most-cited living scientist thanks to his foundational contributions to deep learning. “I’m talking to a lot of CEOs around the world,” he told me, “and they’ll automate [jobs] as fast as they possibly can.” Indeed, a 2025 World Economic Forum survey of more than one thousand large employers found more than 40 percent planned to reduce their workforce as a result of automation by 2030.

Companies are also creating more wealth with fewer people than ever. Per employee, the world’s most valuable companies in recent decades were worth anywhere from $3.5 million to as much as $30 million. Nvidia? More than $100 million. Pure AI firms like OpenAI, Anthropic, and xAI have pushed the trend even further, with ratios between $140 million and $167 million.

This surge in AI-driven wealth concentration builds on a decades-long rise in inequality. As the economists Emmanuel Saez and Gabriel Zucman have shown, inequality levels in the United States are the highest they’ve been since the 1920s. If machines can truly substitute for human labor across the board, the result could look like a worldwide game of musical chairs, where your economic position is determined by whatever wealth you happened to have at the moment your work stopped being worth paying for.

Looking at this data, AI appears less a bolt from the blue than the next — and perhaps final — frontier in the endless pursuit of profit.

Can They Really Do It?

It probably comes as little surprise to you that many capitalists want to automate their workers away. But can they?

The question is difficult to answer definitively. Anthropic CEO Dario Amodei sounded the alarm in mid-2025, predicting what Axios dubbed a “white-collar bloodbath.” AI could eliminate half of all entry-level white-collar jobs and spike overall unemployment to 10 to 20 percent in the next one to five years, he told the publication. Of course, to skeptics, the CEO of an AI company warning that its technology would be so effective it would soon replace human workers was nothing more than a savvy earned media campaign. But that’s hard to square with more recent reporting that OpenAI blocked its economists from publishing research that showed its technology was displacing jobs.

By 2025, signs of AI’s deleterious effects on the workforce appeared to be everywhere. Rising unemployment for college graduates, dips in job postings, and dramatic Big Tech layoffs were all pointing the same way, and software engineer hiring in particular cratered. People were quick to attribute the drop to AI.

However, as keen observers noted, the plunge in hiring coincided more neatly with the Fed’s increases to interest rates than with the rise of ChatGPT, which didn’t become much of a coder until after net hiring had (barely) resumed. Generative AI rose to prominence at around the same time as today’s turbulent macroeconomic and political conditions, making it fiendishly difficult to explain changes. Anecdata in the form of news stories about young people struggling to find programming jobs won’t cut it.

I hate to say it, but this sounds like a job for economists.

In 2024, when Daron Acemoglu — who would go on to share that year’s Nobel Prize in economics — predicted the labor market effects of AI over the next decade, people took notice. Against Goldman Sachs and McKinsey predictions that AI would add trillions to GDP, he expected a nothingburger, forecasting that US GDP would only be around 1 percent higher in ten years due to AI.

Economists mostly don’t believe there’s any real chance AI could put most people out of work anytime soon, Stanford economics postdoc Phil Trammell told me, but Acemoglu’s view is even “more pessimistic than the mainstream.”

On the opposite end of the spectrum is Anton Korinek, a University of Virginia economist on leave at Anthropic since March, who uses standard economic models to look at what might happen if firms can effectively turn capital into more labor. Assuming full automation is reached in twenty years (“baseline AGI”), he and his PhD student Donghyun Suh predict the economy would double over the subsequent decade. If it’s possible in five years (“aggressive AGI”), they predict a tripling of GDP by 2034 — ten times faster growth than the business-as-usual scenario. “However,” they write, “wages collapse as the economy approaches full automation.”

“Economists have been telling the rest of the world for the past 200 years that we shouldn’t worry about automation,” Korinek told me, “because if you destroy some jobs, the economy always creates new ones.” That was true in the past, he acknowledged, but only because “there were always things that only humans could do.” Indeed, there are more people than ever before, with more jobs than ever before, even after more than two centuries of automation. However, we’ve never had a machine that can actually make labor. (Korinek has separately clarified he’s not predicting imminent AGI himself, but rather using traditional economics to show what would happen if the predictions of AI experts like Geoffrey Hinton actually came to pass.)

Economic forecasts are notoriously sensitive to their underlying assumptions. Acemoglu’s 2024 paper uses mainstream ways to measure a task’s automation potential. But when it comes to AI’s capabilities, he takes a distinctly conservative approach, assuming they won’t significantly advance beyond GPT-4’s, which was the state of the art at the time.

Trammell told me that the difference between the Acemoglu and Korinek camps comes down to “whether the economic impacts of AI over the next ten years will be mainly about the diffusion of what AI could already do as of a year ago versus the things it becomes able to do over the next ten years.”

That’s not really an economics question, Trammell said. “It’s a question about technological forecasting. I think reasonable people can disagree about it. But it’s also not Daron’s wheelhouse.”

In September 2025, I interviewed Acemoglu for a Nobel Foundation panel during the United Nations General Assembly. To my surprise, he said he was going to “stick to his guns.” In a late 2024 interview, he had argued that the productivity story depends on the technology creating genuinely new tasks and new products, and he doesn’t see the industry’s current focus delivering much of that within a decade. He expects near-term gains to concentrate in relatively “easy-to-learn” white-collar tasks, while jobs that rely on tacit judgment, social interaction, and hard-to-verify outcomes remain far less exposed.

It is true that, even years after the release of ChatGPT, it’s been hard to find evidence that AI is actually killing jobs. An October 2025 Yale study using the standard dataset for labor economists, the US Current Population Survey (CPS), found that “currently, measures of exposure, automation, and augmentation show no sign of being related to changes in employment or unemployment.” Also in October, when I asked Korinek about the data to date, he admitted that we only have “tiny little bits of empirical evidence that are not rising to the level of transformativeness.” So, “if you say I’m an honest evidence-bound economist, it’s still fair to say that the evidence doesn’t suggest that this is different from any prior technological revolution.”

But focusing on whole occupations may miss the early signs of automation. This is where three Stanford economists stepped in with a groundbreaking paper in August 2025. Rather than rely on traditional government sources like CPS, which offer a useful — but bird’s-eye — view of employment changes, they got their hands on data from a payroll processor, allowing them to zoom in on trends too subtle to spot otherwise. Spoiler: the anecdotes were a real sign that the white-collar bloodbath was already well underway, and like the 1918 flu, it claimed young people most.

With payroll data going back to 2021, the authors split professions into six age cohorts, starting with twenty-two- to twenty-five-year-olds, and bucketed jobs by AI exposure using the same metric as Acemoglu. The most exposed were in fields like programming and customer service.

After controlling for interest rates, pandemic-era overhiring, and other factors, the data tells no clear story at first. Even for the youngest, most AI-exposed workers, relative employment bounced around for years, declining as much as 6 percent in 2022 before recovering to zero in mid-2024. But then it falls off a cliff, plummeting 16 percent in just over a year, with little sign of stopping.

By late 2025, the macro data looked consistent with this story. The US economy grew at 4.3 percent in the third quarter — the fastest in two years — while job creation averaged an anemic 51,000 per month. Joseph Brusuelas, the chief economist at the accounting firm RSM, writes, “The great decoupling of jobs and growth will take some explaining to the American public.” The Stanford paper offers one compelling explanation.

Who Else Stands to Lose?

Korinek worries that we’re heading toward something more fundamental than a rough patch for entry-level coders. “The main economic risk under AI,” he told me, “is that it will fundamentally undermine labor, that it will lead to essentially a large part of the population becoming losers of this technology.” He observed that wage labor was “basically a construct of the industrial age” — one not likely to survive long term under AGI, where, by default, he expected labor to “experience a very significant devaluation.”

The rise of automation in the United States is happening after decades of declining union density, with just 6 percent of private sector workers belonging to one. Currently, labor automation mostly doesn’t look like mass layoffs; instead the Stanford paper suggests firms will increasingly avoid hiring anyone they might be able to automate — a group that will grow larger as AI improves.

And since AI labor is typically orders of magnitude cheaper than the human variety, substitution doesn’t require that AI be as good, just good enough.

Collective Power

After years reporting on the Obsoleting Project, I’ve found that lawmakers won’t stop it unless we make them. The industry has captured them, and they were already catastrophically insulated from what the public wants. What has slowed automation down, where anything has, is workers organizing.

One place where bosses tried — but largely failed — to use automation to reduce labor power was in scriptwriting. To better understand how, I spoke to Ellen Stutzman, the chief negotiator for the Writers Guild of America (WGA) 2023 strike. She told me that AI concerns emerged organically from what members were already experiencing: a speedup. As streaming replaced broadcast television over the preceding fifteen years, studios shifted from fewer shows with twenty-two-episode seasons to more shows with just eight-episode seasons, asking writers to do more work in less time for less money. The writers were already going into the strike seeking contractual minimums on staffing levels and time to do jobs. So when ChatGPT arrived, it looked like the natural progression of that trend. “It’s like one writer with an AI system doing an entire season,” Stutzman said, “or eventually no writers.”

The WGA didn’t go into bargaining demanding a ban on AI. Instead, it wanted restrictions on how the technology was used, namely for AI-generated material to be treated the way Wikipedia articles were — as research, not writing. If a studio hands a writer AI-generated material, the writer should still get full credit and compensation. Whatever the machine spits out doesn’t count as literary material under the contract. And writers shouldn’t be forced to use AI tools. That last one mattered, she said, because members’ views ran the whole spectrum: “People like David Simon, who was on the negotiating committee, who I think said, ‘I would rather put a gun in my mouth than use AI.’”

It was only on the final weekend before the contract deadline, Stutzman said, that the studios were willing to say anything at all about AI. At that point, they finally pointed to language in the existing collective bargaining agreement saying that a “writer” is a human being and told the guild that, at least on paper, AI could not be credited as the writer. But they stressed that they did not want to limit how they might use the technology in practice.

This coincided with concerns over studios using AI to scan actors to create “synthetic performers” for background shots. The Screen Actors Guild–American Federation of Television and Radio Artists (SAG-AFTRA) launched its own strike a few months later, in which the use of digital replicas became a key sticking point.

It took a five-month strike to get the protections writers wanted, and Stutzman is clear-eyed about the limits of what they achieved: the WGA has no power over any AI company. “What we can do in a collective bargaining agreement is regulate our employers,” she said, “but OpenAI is not our employer.” The companies that scraped writers’ work to train their models remain beyond the guild’s direct reach — a problem the WGA is now trying to address through litigation and legislation.

Against this bleak backdrop, I asked Stutzman what other automation-exposed workers could learn from their experience. She didn’t hesitate. “It’s the same lesson I have on everything, which is about collective power,” she said. “If you want to push back, if you want to regulate AI, if you want to make it a tool that you use versus one you’re being replaced with — it’s all about collective action.”

This article originally appeared in Jacobin, a democratic socialist magazine publishing long-form essays and analysis on politics, economics, and culture. Subscribe to the print edition for $20 a year.