Anthropic recently published a research paper titled "Labor market impacts of AI: A new measure and early evidence." Its core contribution is something that had not been done before: drawing a clear line between what AI can theoretically do and what it is actually doing in the real world.
This distinction is far more important than you might think.
Why previous predictions always missed the mark
When assessing AI's impact on jobs, the conventional approach has been to list the tasks within an occupation, ask "Can AI do this?", calculate a percentage, and then conclude something like "65% of this profession will be replaced by AI."
The problem is that there is a huge gap between "can do" and "is doing."
In this study, Anthropic economist Peter McCrory introduces the concept of "observed exposure." This overlays theoretical AI capabilities with real-world API usage data from Claude, looking at what users are actually using AI to do.
What the data says
The most representative example is software developers:
- Theoretical exposure: AI could theoretically handle 94% of programming tasks
- Observed exposure: Only 33% of tasks are actually being performed by AI in practice
In other words, while 94% of a programmer's work could theoretically be done by AI, in reality AI has only penetrated about one-third of it. The remaining 60-plus percentage points are not tasks AI cannot do; they are tasks that, for various reasons—inertia, authorization issues, regulation, trust thresholds, integration costs—people are still doing themselves.
| Occupation | Theoretical exposure | Observed exposure |
|---|---|---|
| Programmers (overall) | 94% | 33% |
| Data entry clerks | Relatively high | Relatively high (close) |
| Microbiologists | High | Low (lab work cannot be outsourced to AI) |
| Real estate managers | Medium | Low (negotiation skills are hard to AI-ify) |
For highly repetitive work like data entry, theoretical and observed exposure are relatively close—AI is indeed doing these tasks.
But for microbiologists, theoretical exposure is not low (AI can handle data analysis and literature review), yet observed exposure is very low because the core work is done by hand in a laboratory.
Who is more at risk: not the group you think
Another counterintuitive finding: workers in high-exposure occupations tend to be more educated, higher-paid, and more likely to be women.
- Workers in the top 25% of exposure earn an average of 47% more than those in low-exposure occupations
- They are 4.5 times more likely to hold a graduate degree
- The share of women is 16 percentage points higher than in low-exposure occupations
This upends the conventional wisdom that "AI will eliminate low-skill jobs first." Knowledge work, copywriting, analysis, and programming—these white-collar roles are precisely where AI is penetrating first.
The impact on young people is already quantifiable: 22-to-25-year-olds are 14% less likely to find jobs in high-exposure occupations than before. There has been no mass unemployment, but the entry bar has risen—fewer new positions are opening up in these fields because existing employees have used AI to boost their own productivity.
What McCrory said
"If you want Claude to help you with machine learning, you actually need to understand machine learning yourself to direct it."
The implication is that AI does not replace professional expertise; it requires practitioners to use that expertise to harness AI. The bar has not been lowered, only transformed.
This is reflected in employment data: in high-exposure occupations, wages for experienced workers are rising, but entry-level positions are disappearing. AI helps veterans become more efficient while simultaneously replacing some of the foundational work that used to serve as a stepping stone for younger workers.
Why this research matters
Previous discussions about AI and work have oscillated between doomsday scenarios (AI will eliminate all jobs) and optimism (AI will only create more employment).
The value of Anthropic's study is that it provides a measurable framework that can identify which occupations are being penetrated before employment data collapses. This gives policymakers time to react, rather than starting remediation only after a wave of job losses.
The current observed data shows: we have not yet reached mass unemployment, but the path for young people into high-exposure occupations is already narrowing. The future trend, readers can infer for themselves.
Sources: Anthropic's research shows that AI can already do a huge portion of many jobs (Fortune); CocoLoop; Labor market impacts of AI: A new measure and early evidence (Anthropic Research)