On March 5, 2026, Anthropic published a research paper with a title dry enough to ignore: "Labor market impacts of AI: A new measure and early evidence." The authors were Maxim Massenkoff and Peter McCrory, two economists working inside the company that builds Claude, one of the most powerful AI systems on the planet.
The paper was not dry. Buried in the analysis was a scenario the researchers named explicitly: a "Great Recession for white-collar workers." Not a metaphor. A direct comparison to 2007-2009, when the U.S. unemployment rate doubled from 5% to 10% and millions lost their homes.
The scenario works like this. If unemployment among workers in the most AI-exposed occupations doubled from its current 3% to 6%, the economic impact would mirror what happened during the financial crisis. Except this time, the people losing their jobs would not be construction workers and bank tellers. They would be programmers, financial analysts, customer service representatives, and marketing specialists. The college-educated. The high earners. The people who thought automation was something that happened to other people.
This is an AI company publishing a roadmap to the displacement of the workers who use its own product.
For context: This article builds on our earlier coverage of the AI productivity paradox and the growing pattern of AI reshaping developer workflows.
What Anthropic Actually Measured
The study introduces a metric the researchers call "observed exposure": a measure that compares what AI could theoretically do against what it is actually doing in professional settings right now.
Here is the methodology. Anthropic used a privacy-preserving analysis tool called Clio to map millions of real Claude conversations onto roughly 20,000 specific work tasks in the U.S. Department of Labor's O*NET database. They then compared these real-world usage patterns against theoretical capability estimates from earlier research by Eloundou et al. (2023), which assessed whether large language models could double the speed of a given task.
The result is a two-layer picture of every occupation in America. One layer shows what AI could do. The other shows what AI is doing.
| Occupation Category | Theoretical Capability | Actual Claude Usage | Gap |
|---|---|---|---|
| Computer & Mathematical | 94% | 33% | 61 points |
| Office & Administrative | 90% | ~15% | ~75 points |
| Business & Financial | ~85% | ~20% | ~65 points |
| Legal | ~80% | ~18% | ~62 points |
| Management | ~75% | ~12% | ~63 points |
The gap tells a story. AI is barely scratching the surface of what it could automate. The researchers describe actual usage as being "dwarfed" by theoretical possibility. Ninety-seven percent of observed Claude usage falls into task categories that are theoretically feasible to automate. The capability is there. The adoption is not.
Not yet.
The Jobs Most Exposed Right Now
The study ranks occupations by observed exposure. The top ten are exactly what you would expect, and exactly what should concern anyone in a white-collar career.
| Rank | Occupation | Observed Exposure |
|---|---|---|
| 1 | Computer Programmers | 75% |
| 2 | Customer Service Representatives | 70% |
| 3 | Data Entry Keyers | 67% |
| 4 | Medical Record Specialists | 67% |
| 5 | Market Research Analysts / Marketing Specialists | 65% |
| 6 | Sales Representatives | 63% |
| 7 | Financial and Investment Analysts | 57% |
| 8 | Software Quality Assurance Analysts | 52% |
| 9 | Information Security Analysts | 49% |
| 10 | Computer User Support Specialists | 47% |
At the other end of the spectrum: 30% of American workers have zero AI exposure. Cooks. Motorcycle mechanics. Lifeguards. Bartenders. Dishwashers. Jobs that require physical presence, and that no language model can replicate.
The demographic profile of the most exposed workers inverts every assumption about who automation threatens. Workers in the top exposure quartile earn 47%** more** on average than those in the bottom. They are 16 percentage points more likely to be female. They hold graduate degrees at nearly four times the rate of unexposed workers. They are 11 percentage points more likely to be white and nearly twice as likely to be Asian.
This is not blue-collar disruption. This is the professional class.
The Hiring Freeze Has Already Started
The study's most consequential finding is not about layoffs. It is about the jobs that never get created.
Massenkoff and McCrory found no systematic increase in unemployment among AI-exposed workers since ChatGPT launched in late 2022. That sounds reassuring until you read the next line. Among workers aged 22 to 25 entering high-exposure occupations, the monthly job-finding rate has dropped by roughly 14% compared to pre-ChatGPT levels.
The researchers are careful to note this finding is "just barely statistically significant." But the direction is clear. Young people are not being fired. They are not being hired.
A 16% fall in employment among workers aged 22-25 in AI-exposed fields tells a story that no unemployment statistic captures. The entry-level pipeline is narrowing. And the people affected have no work history to fall back on.
This pattern is showing up in the broader economy. The February 2026 jobs report, released the day after Anthropic published its study, showed the U.S. economy shed 92,000 jobs. Unemployment ticked up to 4.4%. Long-term unemployment hit its highest level since December 2021, with the average duration stretching to 25.7 weeks.
How It Unfolded
The Irony No One Can Ignore
There is something deeply strange about this study, and everyone covering it has noticed.
Anthropic is the company that builds Claude. Claude is the product that is automating these jobs. And Anthropic is the company publishing the research showing, in precise statistical detail, exactly how that automation is progressing and where it could lead.
This is not an outside critic sounding alarms. This is the manufacturer publishing the safety recall.
Dario Amodei, Anthropic's CEO, set the tone for this conversation in May 2025 when he told Axios that AI "could disrupt half of entry-level white-collar work" within one to five years. He said something else that day that few companies in his position would say: "We, as the producers of this technology, have a duty and an obligation to be honest about what is coming."
In January 2026, Amodei published a 20,000-word essay titled "The Adolescence of Technology," using starker language and shorter timelines than he had in the past. The essay warned of a "country of geniuses in a datacenter."
Mustafa Suleyman, Microsoft's AI chief, went further in a February interview with the Financial Times: "White-collar work, where you're sitting down at a computer, being a lawyer, or an accountant, or a project manager, or a marketing person, most of those tasks will be fully automated by an AI within the next 12 to 18 months."
The predictions keep getting bolder. The data keeps getting more specific. And the people whose jobs are in the crosshairs are the ones generating the data by using the products.
Understanding how large language models actually work makes the pattern clear: these systems excel at exactly the text-heavy, pattern-matching, structured-output tasks that define most white-collar work.
The Counterarguments Are Substantial
The catastrophe scenario is not the only reading of this data. And the people pushing back are not handwaving.
The adoption gap is the story, not the displacement. The study's own numbers show that actual AI usage sits far below theoretical capability in every single occupation category. Legal constraints, model limitations, integration requirements, and the need for human review all slow adoption. The gap between 94% capability and 33% usage is not a ticking time bomb. It might be a permanent friction.
History says these predictions are wrong. The researchers themselves note that past approaches to measuring "job offshorability" identified roughly a quarter of U.S. jobs as vulnerable to outsourcing. A decade later, most of those jobs maintained healthy employment growth. The Anthropic team writes that "the track record of past approaches gives reason for humility."
The February jobs report has other explanations. The 92,000 jobs lost in February were driven by severe winter weather, a major healthcare strike, and significant downward revisions to prior months. Not AI. Economists at multiple outlets described the report as a weather-driven anomaly.
The NBER found almost nothing. A February 2026 survey of 6,000 C-suite executives across four countries found that more than 80% said AI had had no impact on productivity or employment at their business. Only one-third of leaders personally used AI, averaging just 1.5 hours per week. The authors invoked Robert Solow's famous 1987 observation: "You can see the computer age everywhere but in the productivity statistics."
McKinsey projects net job creation. Their modeling suggests AI will displace roughly 3.5 million jobs through direct and indirect effects, but those losses would be offset by 4.2 million new jobs driven by capital expenditure, rising incomes, and healthcare spending. The net effect: more jobs, different skills.
Daron Acemoglu, the MIT economist who won the 2024 Nobel Prize for his work on institutions and prosperity, cautions that this transition could favor skilled workers and widen inequality if retraining does not keep pace. The jobs may come, but they may not come to the same people.
The Bottom Line
An AI company analyzed its own product's usage data, mapped it against every occupation in America, and told the world that the gap between what AI can do and what it is doing is enormous. The theoretical ceiling sits at 94% of computer and math tasks. The current floor sits at 33%. Everything between those two numbers is the territory where the next decade of white-collar work will be decided.
The study found no employment crisis today. But it found something that may matter more: a 14% drop in job-finding rates for workers aged 22 to 25 in the most exposed fields. Not firings. A quiet narrowing of the pipeline. Entry-level positions disappearing before anyone applies for them. If that trend accelerates as adoption closes the gap between theoretical capability and actual usage, the researchers' own scenario applies: a Great Recession concentrated entirely in the professional class.
What makes this study different from every other AI-and-jobs forecast is who published it. This is not a think tank or a consulting firm or a viral blog post. This is Anthropic, the company building the technology, telling the world what its own data shows. Amodei said it plainly: "We have a duty and an obligation to be honest about what is coming."
Whether honesty arrives in time to change the outcome is a question the data cannot answer.
Sources
- Anthropic: Labor Market Impacts of AI: A New Measure and Early Evidence (Mar 5, 2026)
- Anthropic: Introducing the Anthropic Economic Index (Jan 2026)
- Fortune: Anthropic just mapped out which jobs AI could potentially replace. A 'Great Recession for white-collar workers' is absolutely possible (Mar 6, 2026)
- Fortune: Will AI take my job? A new Anthropic study suggests the answer is more complicated than you think (Mar 10, 2026)
- CBS News: Anthropic is tracking which jobs are most exposed to AI. These 10 professions top the list. (Mar 2026)
- The Register: Anthropic bods say AI hasn't had much impact on jobs (Mar 7, 2026)
- The Decoder: Anthropic's new study shows AI is nowhere near its theoretical job disruption potential (Mar 2026)
- Axios: Anthropic launches AI job destruction detector (Mar 5, 2026)
- CNN: The US economy lost 92,000 jobs in February (Mar 6, 2026)
- Fortune: Microsoft AI chief gives it 18 months for all white-collar work to be automated (Feb 13, 2026)
- Fortune: Anthropic CEO warns AI could wipe out 50% of entry-level white-collar jobs (May 28, 2025)
- ResumeTemplates.com: 60% of Gen Zers will pursue skilled trade work in 2026 (Feb 2026)
- LDS: Developers Thought AI Made Them Faster. The Data Said Otherwise. (Feb 20, 2026)
- LDS: Spotify Developers Haven't Written Code Since December (Feb 19, 2026)