The invention of electricity made menial jobs such as lamplighters, elevator operators, and human knockers, the modern equivalent of alarm clocks, unnecessary. The advent of computers eliminated the need for data entry clerks, switchboard operators, and file clerks.
Anthropic, the artificial intelligence (AI) company that emerged in 2026 as an existential threat to Claude Model’s breathtaking new capabilities and an existential threat to its multibillion market value, has returned with a warning about how outdated AI tools can do far more work. Founded by former OpenAI employees who were as obsessed with AI safety as they were with its advancements, the AI giant is a thought leader not only in advances but also in AI risks, and just published a study containing the most detailed map to date of the jobs AI is actively doing and the jobs it’s merely capable of doing. Depending on your line of work, the difference between these two numbers can be both comforting and alarming.
In their report, “AI’s Labor Market Impact: New Indicators and Early Evidence,” authors Maxim Masenkov and Peter McCrory found that actual AI adoption is only a fraction of what AI tools can achieve.
AI could theoretically cover most tasks in business, finance, management, computer science, mathematics, legal, and office management roles. But in most sectors, actual adoption (as measured by researchers using work-related usage data from Anthropic’s AI model Claude) is only a fraction of the theoretical adoption potential.
Business leaders have been heeding warnings about AI’s ability to replace white-collar jobs for months. Anthropic CEO Dario Amodei said last year that the technology could disrupt half of entry-level white-collar jobs. Mustafa Suleiman, head of AI at Microsoft, made a similar prediction, estimating that most professional jobs will be replaced within one year to 18 months.
Researchers believe that the delay is due to existing legal constraints and technical hurdles, including model limitations, the need for additional software tools, and the need for humans to still review AI work. But they predict it will only be temporary.
Who is most at risk?
This study introduces something called “observed exposure.” This is a new metric that compares theoretical AI capabilities with real-world usage data extracted directly from interactions with Claude in professional environments. The discoveries that jumped off the page only scratched the surface of what AI is technically capable of. And even if the gap does narrow, the workers most at risk will be older, more educated, and higher-income workers.
The workers who will bear the brunt of that scenario are not the ones many imagine. The groups most exposed to AI are 16 points more likely to be female, have 47% higher average incomes, and are almost four times more likely to have a graduate degree than the groups least exposed to AI. That’s lawyers, financial analysts, software developers, not warehouse workers. Computer programmers, customer service representatives, and data entry keystrokers are among the most at-risk occupations.
But even the careers most exposed to AI capabilities have yet to fully evaluate them. The researchers cite approving drug refills at pharmacies as an example of a fully exposed task commonly performed by physicians. While AI can certainly automate this task, they point out that they have yet to observe Claude doing it, even though it could theoretically be completed with a large language model.
The results were amazing. For computer and math workers, large language models can theoretically handle 94% of their tasks. However, Claude currently covers only 33% of these tasks in observed professional use. A similar gap exists in office and administrative roles. Theoretical capacity is 90%, and the actual capacity is only a fraction of that.
As the researchers describe it, the “red area” of how AI is actually used appears small compared to the “blue area” of what is possible. As capabilities improve and adoption increases, red will grow to fill the blue, the researchers wrote. Meanwhile, 30% of employees have zero exposure to AI. Cooks, mechanics, bartenders, dishwashers, etc. are jobs that require a physical presence that cannot be replicated in an LLM.
Peter Walker, head of analysis at Carta, extrapolated the blue and red findings into a bar graph. “Universal truth: most radar charts should be just bar graphs,” he wrote to X. “Love what is yours, humanity!”
The paper calls the scenario that everyone involved in the knowledge economy should consider the “Great Recession for white-collar workers” and points out that during the financial crisis of 2007 to 2009, the unemployment rate in the United States doubled from 5% to 10%. The researchers note that their framework clearly detects a doubling of the top quartile of jobs exposed to AI from 3% to 6%. It hasn’t happened yet, but it absolutely could happen.
If you think this is AI companies speaking their books, this is emerging as a distinct possibility in a number of scenarios, far beyond the viral doomsday essays of the likes of Matt Schumer and Citrini Research. In a speech last month, Federal Reserve President Michael S. Barr laid out three possible scenarios for the introduction of AI.
employment slowdown
The U.S. Bureau of Labor Statistics released dire employment figures Friday. Employers cut 92,000 jobs in February, and the unemployment rate rose to 4.4%. Some companies have recently announced large-scale layoffs due to AI. Jack Dorsey’s block cut nearly half its workforce last month, citing AI. “We are already seeing that the intelligence tools we create and use, combined with smaller, flatter teams, are enabling new ways of working that fundamentally change what it means to build and run a company,” Dorsey wrote in a post on (They note that companies may be conducting “cleansing operations” or using this as an excuse to make necessary job cuts.)
However, the study found that, at least for younger workers, the problem is not layoffs but a slowdown in employment in AI-exposed fields, with employment rates in AI-exposed occupations declining by 14% in the post-ChatGPT era compared to 2022. However, the researchers note that these findings are barely statistically significant. And so far, research shows no systematic increase in the unemployment rate. Citadel Securities, not known for publishing market research, was moved by a viral doomsday editorial pointing out that hiring for software engineers has actually increased in recent months.
Still, the Antropic researchers suggest that the small decline may signal a new reality of employment in the age of AI, consistent with other research on the state of the job market for young workers. A similar study found a 16% decline in employment in jobs exposed to AI among workers aged 22 to 25.
For some young workers, that means avoiding the labor market altogether. “Young workers who are not hired may stay in their current job, take another job, or go back to school,” the researchers said.
