The robots can already do three-quarters of the physical work in America. They just cannot afford to take it. That gap – between what a system can do and what a company is willing to pay to let it do it – is the only number that matters to the people being told this week that their job is next in line.
Anthropic published an index on September 30 that measured something the AI conversation had never priced: not whether a robot can perform a task, but whether it can perform that task for less than the human doing it today. The answer is that modern robots can already execute 74 percent of physical job tasks in the United States – work that adds up to roughly 34 percent of all working hours. And then the number that guts the panic: robots are cost-competitive with human labor for 0.3 percent of those tasks.
Three-quarters capable. Three-tenths of one percent worth doing. Almost everything the public is being told about an AI jobs apocalypse lives inside the distance between those two figures, and very little of it survives the trip.
Anthropic – A Robot That Can Do Your Job and Still Cannot Afford to Take It
The study drew its map from O*NET’s catalog of job tasks, federal employment data, and assessments of the robots actually for sale. It finds that capability is highest where the work is driving and warehouse labor, and lowest in nursing and general repair. Add language models to the mix and about 81 percent of American work is exposed to software AI, robots, or both. Read that number carelessly and you get a headline. Read it carefully and you get a distinction the industry has spent two years blurring: exposure is not replacement. It is a measure of what could be touched, not what will be.
The 74 percent is a capability score. The 0.3 percent is a purchasing decision, and they are not the same document. A warehouse operator weighing a $60,000 robot against a worker earning $40,000 does a math problem that has nothing to do with how impressive the demonstration was. The analysis of the study notes the cost curve has been falling at roughly three percent a year. Hold that pace and the share of work where a robot undercuts a person climbs from 0.3 percent to about ten percent in forty years. That is a slow fuse, not a switch.
A capability chart is a warning. A price list is a decision. Only one of them fires anyone.
Dario Amodei – The Sharpest Argument Against His Forecast Is His Own Lab’s Data
The Anthropic chief executive has said his own estimate is that AI could wipe out up to half of entry-level white-collar jobs within one to five years, a claim he sharpened in a January 2026 essay. Note what he did not say: that it is happening yet. The people most nervous about that forecast are the ones doing the office work it names – and the machines his lab just measured are aimed at the warehouse and the road, where the economics remain stubbornly human.
So the fear is now coming from the same building as the product. That does not make the forecast false. It does make it a document worth reading twice, because a company that sells general-purpose AI has a reason to want you to believe the transition is inevitable – and a 0.3 percent cost-competitiveness rate is not inevitable, it is an indifference point that moves when prices move. When the alarm and the advertisement come from the same desk, the alarm deserves a second set of eyes.
California – The Law Was Written for the Alleged Firings, Not the Robot Ones
On September 30, Governor Gavin Newsom signed four first-in-the-nation bills that treat AI in the workplace as a labor problem rather than a science-fiction one. The package bars an employer from relying solely on an automated decision system to fire or discipline a worker, requires a human to independently corroborate the outcome, and lets a worker request the data that drove the call. Violations carry a $500 civil penalty each. It takes effect July 1, 2027.
The bill that matters most is quieter. SB 951 folds artificial intelligence into the state’s mass-layoff notice law: if a layoff is caused in whole or substantial part by AI replacing positions, the employer must say so, name the job functions being automated, and identify the technology behind it. California’s Employment Development Department will publish a quarterly statewide count of those displacements. That is the difference between a trend and a rumor. Until now, the boldest claim in corporate America – that a system took the jobs – was the one claim nobody had to file paperwork to make.
You can legislate the disclosure of an algorithm’s reasoning. You cannot legislate back the job it already took.
Lorena Gonzalez – The Guardrails Arrived After the Firings, Not Before
“Workers across California have demanded that our state lead the way in regulating AI in our workplaces,” the president of the California Federation of Labor Unions said as the bills were signed. She is right that the guardrails are first in the nation. She is also describing a race the workers already lost by the time the ink dried. The protections begin in July 2027; the layoffs they respond to are happening in 2026. A tracker that counts announced cuts puts 2026 employment losses past 225,000 workers across more than 500 events in the United States, with technology firms citing AI as the reason while their revenue stayed healthy.
Read that motive honestly, because it is the whole story: the payroll is being converted into the research budget. Executives will point at a capability chart showing robots that can do nearly everything, and a 0.3 percent cost figure showing they can afford almost none of it, and still cut – because the cut is the point. The disclosure law exists precisely because the companies could not be trusted to confess which of the two numbers was actually driving the decision.
What Changes Next – Watch the Price, Not the Demo
The useful indicator is not another capability benchmark. It is the cost curve, and whether the 0.3 percent climbs because robots got better or because workers got cheaper to replace. The Anthropic research offers one more fact that should shape every policy argument: the workers in the most exposed occupations already earn less, hold less education, and are more likely to be unemployed than workers in unexposed jobs. They are not tomorrow’s victims of a technical revolution. They are today’s leverage for a cost-cutting plan that has been dressed up as one.
So measure the right thing. Not how many tasks a robot can perform on a good day, but how many people still hold a comparable job at comparable pay a year after their role was labeled exposed. If that number is high, the fearmongers were selling a forecast. If it is low, we will have let a capability chart stand in for a business decision – and charged the difference to the workers who were never asked.
The robots are coming. They just are not the ones writing the pink slips.
Sources: Anthropic, Office of Governor Gavin Newsom, mixed-news.com, SkillSyncer.