Dario Amodei, the CEO of Anthropic, spent the last year telling anyone who would listen that AI would eliminate half of all entry-level jobs within five years and become a ‘general labor substitute for humans.’ Then his own company published a report that quietly contradicted him. That tension is the most honest summary of where the AI employment debate actually stands right now.
As reported by The Guardian, Anthropic’s internal analysis found no systematic increase in unemployment among workers most exposed to AI since late 2022. Claude, Anthropic’s flagship model, covers just 33% of tasks in the computer and math category, despite being theoretically capable of handling close to 100%. Deployment, in other words, is lagging far behind what the technology could do. And labor productivity during the first three years of the AI era has actually grown more slowly than it did during the 1990s IT boom. That’s a damning comparison.
Even Sam Altman has walked back the apocalyptic framing. ‘I don’t think we’re going to have the kind of jobs apocalypse that some of the companies in our space advocate or talk about,’ he said in May. Whether that’s genuine recalibration or, as MIT economist David Autor put it, a recognition that ‘the world is not changing as fast as they predicted,’ the shift in tone from the loudest AI voices is real.
One useful framework for understanding why AI hasn’t replaced human workers en masse is what economists call the O-ring argument, named after the rubber component whose failure destroyed the Space Shuttle Challenger in 1986. The idea is that as long as AI can’t perform every task reliably, the tasks it can’t do become more valuable, not less. Depending on where AI takes over in a workflow, it can actually increase demand for skilled workers freed from low-end tasks, or create openings for less experienced workers by automating the more expert-level components. One recent study found that ‘despite strong substitution at the task level, overall employment effects are modest, as reduced demand in exposed occupations is offset by productivity-driven increases in labor demand at AI-adopting firms.’
None of this means the risk is gone. Adoption is expanding fast, models are improving, and researchers like Autor are careful to note that AI hasn’t obviously hit a ceiling yet. Amodei’s one-to-five year window still has time left. Nobel laureate Daron Acemoglu points out that insiders remain as confident as ever in artificial general intelligence arriving soon. And Elon Musk still believes AI plus robots will make human work optional.
But clouds are gathering from a different direction, one that has less to do with jobs and more to do with whether the whole AI buildout is economically sustainable. Seven in ten Americans now oppose datacenter construction in their area, driven partly by energy costs but also by general distrust of what AI is supposed to do to society. The Nasdaq, which had been riding AI-related stocks almost exclusively, dropped roughly 8% from its early June peak.
The cost picture is the part that doesn’t get enough attention. The International Energy Agency estimates datacenter power demand will more than double by 2030, reaching around 945 terawatt-hours, more than Japan’s entire energy consumption. And AI models depreciate fast. Last quarter’s frontier model is this quarter’s obsolete product. Acemoglu is blunt about this: companies developing AI models ‘are never going to make money. They are losing hundreds of billions of dollars every year.’
There’s also a more fundamental limit. As Autor notes, ‘not everything is a computational problem.’ AI is good at replicating language patterns, but connecting language to physical reality, exercising reliable judgment in high-stakes contexts, and avoiding critical errors remain genuine weak spots. Those aren’t bugs that will obviously be patched in the next release cycle.
The honest read right now is this: the job apocalypse narrative was always more useful for fundraising and media coverage than it was accurate. The real story is slower, messier, and ultimately more interesting. AI is spreading, productivity gains will probably come, but the timeline is longer and the economics are harder than anyone hyping this technology in 2023 wanted to admit.
