Anthropic and the Real Impact of AI on Employment

A New Approach to Measuring the Impact of AI

Anthropic, a company specializing in artificial intelligence, is proposing a new method to assess the impact of AI on employment. Unlike traditional approaches that rely on the theoretical capabilities of AI models, Anthropic emphasizes the importance of measuring the actual use of these technologies in workplace environments. This distinction is crucial as it allows for differentiation between what is technically possible and what is actually applied in the professional world.

Most research on AI and employment starts with a fairly simple idea: if a model can theoretically perform a task faster, then the occupation containing that task is "exposed." This seems reasonable until reality comes into play. A task may be technically feasible for AI without being utilized in real-world work environments due to various factors such as complex processes, organizational sluggishness, high risks, lack of adequate software infrastructure, or the need for human validation.

That’s why this paper does not claim that "AI is taking jobs now." Instead, it argues that it is time to stop making assumptions based solely on theoretical capabilities and start tracking the actual use of AI in work. Think of it as the difference between having a gym membership and actually going there every day at 6 a.m. The capability exists in both cases, but the impact is real only in one of them. Anthropic is trying to measure this reality.

The 'Observed Exposure' Metric

At the heart of this new approach is the 'Observed Exposure' metric. It assesses not only the theoretical capability of AI to accomplish a task but also its concrete usage. To do this, Anthropic relies on data from three main sources: O*NET tasks for about 800 occupations, prior estimates on the capacity of language models (LLMs) to accelerate these tasks, and real usage data of AI, particularly through their tool Claude.

The core of the document rests on a new metric called Observed Exposure. In simple terms, it measures not only whether AI could assist with a task but whether it actually does. Anthropic evaluates this using three elements:

• O*NET task data for approximately 800 occupations
• Prior estimates on the capacity of LLMs to theoretically accelerate these tasks
• Real usage data from Claude

After considering these three metrics, the concept of Observed Exposure places more weight on professional and automated use than on occasional or purely assistive use.

This is important because not all uses of AI are equal. A marketer using Claude to generate five title options is not the same as a support team integrating AI into a workflow that responds to customer inquiries at scale. One is assistance, while the other is an almost total substitution of human labor.

Key Takeaways from the Report

Anthropic's report reveals several important points. First, the jobs most exposed to AI are those where it is already perceived as useful, particularly in repetitive and language-focused tasks. Computer programmers, for example, show a coverage of 75%, followed by customer service representatives and data entry clerks at 67%.

With its striking counterpoints to some common beliefs, Anthropic's report shares extremely enlightening insights.

1. The most exposed jobs are those where AI is already useful
   Jobs with the highest observed exposure are those where generative AI seems naturally useful: repetitive, language-focused, and screen-based work. The most exposed professions identified by Anthropic include computer programmers with 75% coverage, followed by customer service representatives and data entry clerks with 67% coverage.

2. A large part of the economy remains untouched
   Approximately 30% of workers show zero coverage in Anthropic's framework, as their tasks barely appear in the data. This includes professions such as cooks, motorcycle mechanics, lifeguards, bartenders, dishwashers, and coat check attendants.

3. Higher exposure to AI is linked to lower long-term job growth
   Anthropic compares its observed exposure metric with BLS employment projections for 2024 to 2034 and finds that the most exposed professions are expected to experience lower growth. For every 10 percentage point increase in observed exposure, projected job growth decreases by 0.6 percentage points.

4. The most exposed workers are not who many assume
   Workers in the highest exposure group are more likely to be older, female, more educated, and higher paid. They earn on average 47% more than the non-exposed group.

5. There is still no clear shock to unemployment
   Anthropic finds no systematic increase in unemployment for highly exposed workers since late 2022. The difference between unemployment trends for workers in the highest exposure quartile and those in the non-exposed group is small and statistically insignificant.

6. Young workers may face the earliest pressure
   Anthropic finds evidence suggesting that hiring in highly exposed professions has slowed for workers aged 22 to 25. Job search rates for young workers entering exposed roles have dropped by about 14% compared to 2022.