Artificial intelligence (AI) has not triggered the wave of mass unemployment that many analysts feared, but it is quietly splitting the global workforce into two tiers: those who can harness AI tools effectively, and those who cannot. That is the central finding from two new reports by Anthropic, the company behind the Claude AI model.
The first paper, titled “Labour Market Impacts of AI: A New Measure and Early Evidence,” was published on March 5, 2026, by Anthropic economists Maxim Massenkoff and Peter McCrory. It introduces a new measurement framework called “observed exposure,” designed to separate what AI could theoretically do from what it is actually doing in professional settings today. The distinction matters because past forecasts of AI-driven job destruction have repeatedly overstated the disruption.
The findings offer a degree of reassurance. McCrory, speaking at the Axios AI Summit in Washington, said the company’s research finds little evidence of widespread job displacement so far. There is no meaningful difference in unemployment rates between workers whose jobs are most exposed to AI automation and those in less exposed roles that require physical interaction with the real world.
The most exposed occupations identified in the research include computer programmers, customer service representatives, data entry workers, and medical records specialists. Yet even among these groups, the feared unemployment spike has not materialised. The researchers found no impact on unemployment rates for workers in the most exposed occupations, though there is tentative evidence that hiring into some of those positions has slowed.
That hiring slowdown is most visible among younger workers. A related study found a 16 percent fall in employment in AI-exposed jobs among workers aged 22 to 25, with researchers noting that young workers who are not hired may be remaining in existing roles, taking different jobs, or returning to school.
The second report, Anthropic’s fifth Economic Index instalment, released this week and focused on learning curves, reveals a more structural concern beneath the stable employment headlines. Workers who have used Claude for six months or more show a 10 percent higher success rate in their interactions with AI, and the longer the usage period, the stronger the effect. Early adopters are not simply automating tasks but are using AI as a thought partner for iteration, feedback, and higher-order problem solving what the researchers describe as augmentation rather than automation.
AI is reshaping how people work, not whether people work. Some 49 percent of jobs now involve AI usage in at least a quarter of their tasks, up from 36 percent in early 2025. But the distribution of that usage is uneven. Claude is used more intensely in high-income countries, in areas with greater concentrations of knowledge workers, and for a relatively narrow set of specialised tasks and occupations.
The implication is that geography and prior skill level are now determinants of who benefits from AI and who risks falling behind. Signs of a two-tier workforce are already emerging, and neither policymakers nor employers have a clear plan for workers on the wrong side of it.
McCrory stressed that the absence of a crisis today does not guarantee stability tomorrow. “Displacement effects could materialise very quickly, so you want to establish a monitoring framework to understand that before it materialises so that we can catch it as it’s happening and ideally identify the appropriate policy response,” he said.
The researchers note that a doubling of unemployment in the most AI-exposed occupations from 3 percent to 6 percent would mirror the scale of job losses seen during the 2007 to 2009 global financial crisis. It has not happened yet, but they say it absolutely could. The message from Anthropic’s economics team is one of cautious vigilance rather than alarm: the job apocalypse has not arrived, but the window for preparing policy responses may be narrowing.
