Anthropic Builds an Early-Warning System for AI Job Displacement — And the Data Is Already Revealing Which Jobs Are Most at Risk

Anthropic, the artificial intelligence safety company behind the Claude family of AI models, has introduced a new research tool designed to track which jobs are most vulnerable to displacement by AI — before the disruption becomes visible in the labour market.

The measure, called "observed exposure," was developed by Anthropic economists Maxim Massenkoff and Peter McCrory and detailed in a new research paper published this week. Unlike previous attempts to quantify AI's impact on employment, the index goes beyond theoretical capability assessments. It combines what large language models (LLMs) can do in principle with what they are actually doing in practice — placing greater weight on tasks that AI is already automating, rather than those where it merely assists human workers.

The result is one of the most grounded and real-world-anchored measures of AI workforce risk to date.

How the Index Works

The methodology examines three interconnected factors: the specific tasks that make up a given occupation, an estimate of how many of those tasks can be performed by a large language model, and crucially, which of those tasks are already being carried out by AI in the real world today.

A job scores higher on the exposure index not simply because AI could theoretically perform its tasks, but because AI is already doing so at scale. This distinction matters enormously. It separates speculative risk from measurable, present-day disruption — and it gives researchers, policymakers, and workers a far more actionable picture of where pressure is building.

The occupations currently showing the highest task coverage by AI systems make for a revealing list. Computer programmers top the rankings at 75% task coverage, followed by customer service representatives at 70.1%, data entry keyers at 67.1%, medical record specialists at 66.7%, and market research analysts and marketing specialists at 64.8%.

By contrast, roughly 30% of all occupations do not meet the minimum threshold to register as exposed at all. These include cooks, motorcycle mechanics, lifeguards, bartenders, dishwashers, and dressing room attendants — roles grounded in physical presence, manual dexterity, and real-world interaction that remain, for now, beyond the practical reach of current AI systems.

The Findings So Far: Calm Before the Storm?

Despite the sophistication of the tool and the clarity of the exposure rankings, Anthropic's researchers are careful to note that evidence of actual, widespread job displacement remains limited.

The paper reports no systematic rise in unemployment among workers in highly exposed occupations since late 2022. However, there are early signals that hiring of younger workers may be slowing in those same fields — a subtle but potentially significant leading indicator worth watching closely.

The researchers are explicit about the purpose of laying this groundwork now. "This approach won't capture every channel through which AI could reshape the labour market," Massenkoff and McCrory wrote, "but by laying this groundwork now, before meaningful effects have emerged, we hope future findings will more reliably identify economic disruption than post-hoc analyses."

In other words, the tool is designed to catch the wave before it breaks — not after.

A Warning From the Top

The release of the exposure index arrives against a backdrop of increasingly frank commentary from Anthropic's own leadership about AI's economic implications.

In January, CEO Dario Amodei published a 19,000-word essay titled The Adolescence of Technology, in which he warned that AI could cause "unusually painful" disruption to the job market. That a company actively building and deploying frontier AI is simultaneously developing tools to monitor its own potential for workforce harm marks a notable — and some would argue overdue — moment of institutional accountability in the industry.

Why This Matters for the Workforce of Tomorrow

For Hamza Baig, founder of the Automation Institute and creator of Hexona Systems, the Anthropic research underscores a point he has been making to business leaders, students, and policymakers for years: the question was never whether AI would reshape work, but whether people would be prepared for it when it did.

"What Anthropic has done here is important — they've moved the conversation from theory to evidence," said Baig. "The occupations showing the highest exposure today are not fringe roles. They are the backbone of millions of careers globally. This data should be a wake-up call, not a reason to panic, but a reason to act. The workers and organisations that invest in automation literacy now will be the ones who lead, not the ones left behind."

Baig, whose Automation Institute has trained over 30,000 students in AI-driven workflow development, argues that tools like Anthropic's exposure index must be paired with accessible, practical education if they are to drive meaningful change rather than simply quantify risk.

"An index tells you where the fire is burning," he added. "Education is the extinguisher. We need both."

The Bigger Picture

The development of the observed exposure index reflects a broader maturation in how the AI industry is beginning to engage with its own societal footprint. For years, discussions of AI and employment were largely polarised — optimists pointed to historical precedents of technology creating more jobs than it destroyed, while pessimists warned of unprecedented structural unemployment.

What Anthropic's research offers is something more useful than either camp: a methodology for watching what is actually happening, updated as AI capabilities and adoption evolve, and capable of identifying the most vulnerable workers before displacement becomes a crisis rather than a risk.

The economists behind the index are clear that its value will grow over time, particularly as AI adoption accelerates and the economic picture becomes harder to read. The measure, they argue, will be most useful when disruption is "ambiguous" — precisely the moment when post-hoc analysis would be too late to be of practical use.

That moment, by most indicators, is not far away.