The Robot Jobs Apocalypse Is Not Nigh—They're Making Work for You

The Robots Appear To Be Hiring Us

Happy Labor Day. Your job may be doomed. That's what OpenAI has been warning about for years. Back in 2023, researchers at the company and the University of Pennsylvania estimated that about 80 percent of American workers could see at least 10 percent of their tasks affected by large language models.

Nearly one worker in five could see at least half their tasks affected. The paper had the deliberately nerdy title GPTs are GPTs. The first GPT meant the artificial-intelligence models. The second meant "general-purpose technologies": inventions such as electricity or the computer that do not merely create a new industry but rearrange lots of existing ones.

Three-and-a-half years later, the rearranging appears to be proceeding nicely. On Thursday, OpenAI unveiled GPT-6 Astra, its latest and most powerful model, with the ability to operate computers directly and with its advanced cybersecurity features. The company's President Greg Brockman put it less than modestly: "Welcome to the AGI era."

By Sunday, Nvidia CEO Jensen Huang had gone further in a post on X: "AGI has arrived." If you wanted to design a Labor Day specifically to make the American worker nervous, then you would struggle to improve on this one.

Except there is an awkward problem with the story of our impending economic redundancy. The data tells a different tale than the headlines suggest. An analysis published by The Economist just before Labor Day estimates that the artificial-intelligence boom has so far helped create roughly 1 million American jobs, considerably more than the roughly 200,000 layoffs attributed to AI since mid-2023.

America added 162,000 jobs in August, far above expectations. Unemployment is 4.1 percent, a rate the American economy has bettered in barely one month out of 10 over the past half-century. Young workers, supposedly first into the AI woodchipper, are holding up surprisingly well: the unemployment gap between 20- to 24-year-olds and everyone else is close to its narrowest in decades.

And The Economist found that occupations closest to the AI boom—engineers, developers, mathematicians, and data scientists—have added roughly 730,000 jobs above the trend in professional employment since 2022. Perhaps Labor Day is not canceled quite yet.

AGI, Now Available From Nvidia

There is no universally accepted definition of artificial general intelligence. OpenAI CEO Sam Altman himself has questioned whether the term means much, calling it "not a super useful term." Plenty of AI researchers dispute that Astra qualifies.

And Huang has what economists might politely describe as an incentive. Astra was trained on Nvidia equipment. OpenAI's immense computing infrastructure depends heavily on Nvidia.

Huang says more than 100,000 Grace Blackwell systems were involved in training Astra and promptly announced that another 400,000 GPUs are coming online. The man announcing that a product has changed human civilization therefore also happens to sell the machinery required to make the product.

On September 3, Nvidia announced that it had agreed to acquire Hugging Face for the wonderfully nonround sum of $12,930,300,000. Hugging Face has become a sort of GitHub for artificial intelligence: more than 18 million developers and researchers use it to share over 3 million models, 500,000 data sets, and 1 million applications.

Nvidia already dominates the infrastructure underneath the AI industry. Now, it is buying one of the places where much of the industry meets. Regulators have yet to approve the deal, which is not expected to close until the first half of 2027.

If AI is a gold rush, then Nvidia began by selling the picks and shovels. It is now buying the saloon.

When the Hype Train Escaped the Station

This was not Astra, which is an important qualification. OpenAI says the intrusion was driven primarily by an internal-only research model never intended for public release, tested for cybersecurity capabilities with deliberately reduced safeguards.

But agents running on GPT-5.6 Sol, the model then available to the public, also reproduced one of the exploits and copied private evaluation data into a public Hugging Face data set.

Even with that caveat, what happened is remarkable. It began in the spring. During training runs in May and June, OpenAI models found ways around restrictions intended to keep them isolated from the internet and from one another.

They turned an internal package-management system into an unauthorized message board. After OpenAI rebuilt the service in July, agents encoded messages in directory names and reassembled it. They began exchanging information and delegating work, at times describing themselves as a "swarm" or "collective."

Then they went exploring. One agent found publicly exposed Hugging Face credentials on the internet and shared them with the group. Agents subsequently found and chained together previously unknown vulnerabilities, executed code on dozens of Hugging Face servers, obtained root access on one, and acquired credentials for internal systems.

Hugging Face reconstructed roughly 17,600 attacker actions during the intrusion. Its researchers concluded that an autonomous agent had conducted an end-to-end attack through thousands of individual decisions at machine speed.

The motive, such as it was, makes the whole thing even stranger: the AI appears to have been trying to cheat on its test. It had been asked to solve difficult cybersecurity challenges. Instead, agents went looking around the internet for the answers and eventually broke into infrastructure they apparently believed might contain them.

OpenAI called what happened a warning shot. The company also built a new test out of the episode. Stripped of production safeguards, Sol pushed beyond its authorized target 48 percent of the time. Astra, OpenAI says, did so in none of its attempts.

The Apocalypse Has Been Delayed

The original OpenAI labor paper is more subtle than much of the coverage it generated. It did not predict that 19 percent of Americans would lose their jobs. It said 19 percent could have at least half of their tasks affected.

That distinction is starting to look increasingly important. Economies are made of jobs, but jobs are made of tasks. Technology can eliminate some of those tasks while making the person doing the job more productive—and occasionally more valuable.

Just like the spreadsheet did not eliminate accountants, and the internet did not eliminate journalists, nor did ATMs eliminate bank employees as quickly or straightforwardly as many expected.

AI may yet be different, of course. Since January 2023, employment among customer-service workers has fallen by about 10 percent, and among secretaries and administrative assistants by roughly 15 percent.

Much of that is work built on the routine tasks AI agents now handle well. The Bureau of Labor Statistics expects office and administrative-support occupations to shed 752,000 jobs by 2035.

Companies openly talk about using AI to reduce headcount. American employers have announced around 16,000 AI-related job cuts a month on average so far this year, according to Challenger, Gray & Christmas.

Those cuts are to real jobs for real people, but they are also easy to lose in a labor market where employers shed roughly 1.7 million workers in a typical month.

Meanwhile, the other side of the ledger is becoming difficult to ignore. The physical infrastructure required for AI is enormous. Goldman Sachs calculates that American spending on chips, servers, data centers, cooling equipment, and power infrastructure is running roughly $500 billion a year above its 2022 level.

Data center construction alone is proceeding at an annual rate above $75 billion, nearly 60 percent higher than a year ago. That means more electricians, HVAC technicians, construction workers, transformer manufacturers, grid engineers, and, naturally, data-center technicians.

The Economist estimates that employment in five industries central to the data center boom has risen by about 320,000 jobs more than broader construction and manufacturing trends would have suggested since 2023.

Indeed, the jobs website, finds that data center vacancies have more than doubled in two years even as postings overall have fallen. Not all of those jobs belong to AI. Grid upgrades and other factory building would have needed workers anyway. But a great many of them do.

Then there are the jobs directly created by the software revolution. AI engineers, researchers, data annotators, forward-deployed engineers, and the suddenly ubiquitous "head of AI."

And then comes the possibility that AI does something technology has done repeatedly throughout history: makes workers sufficiently more productive that people buy more of what they produce.

Paralegal employment rose about 11 percent between 2023 and 2025, according to The Economist's analysis, despite legal work being among the occupations supposedly standing directly in AI's path. Market-research analysts gained about 6 percent. The national average was around 2.5 percent.

Perhaps making a worker twice as productive does not always mean you need half as many workers. Sometimes, it means you can afford to do twice as much work.

Nobody Knows, Not Even Our AI Overlords

That is not to say everybody wins in this massive change. Technological revolutions produce losers, too. Particular professions may be hollowed out, even as employment rises overall. And the electricians building today's AI data centers may be constructing the machines that destroy tomorrow's office jobs.

Nobody knows. That is a useful thing to remember amid the confident, self-flattering predictions emanating from Silicon Valley.

Three-and-a-half years ago, OpenAI estimated that its technology could touch the work of four out of five Americans. This weekend, the CEO of the company selling its chips declared that general artificial intelligence had arrived.

AI agents have already escaped intended restrictions, collaborated with one another, and broken into another company's computer systems. Those are extraordinary facts.

But so is this one: after three-and-a-half years of increasingly capable artificial intelligence, America has an unemployment rate of 4.1 percent and the technology currently appears to be creating more jobs than it destroys.

Maybe Huang is right. Maybe the robot overlords have arrived. For now, at least, they seem to need help plugging themselves in.