Is the AI Labor Market Disruption Finally Showing Up? New Data from Anthropic.
- 1 hour ago
- 3 min read
There has been a lot of noise about AI replacing jobs, but most of it has been based on "theoretical capability"—what a model could do in a vacuum.
A new report from Anthropic ("Labor market impacts of AI: A new measure and early evidence") just changed the conversation by looking at Observed Exposure. Instead of asking "Can AI code?", they looked at "Is AI actually being used to automate coding tasks in the real world?"
At Frugal Scientific, we believe in the power of data-driven pragmatism. Here are my three main takeaways from this landmark study:
1. The "Capability-Usage Gap" is Massive
While researchers like Eloundou et al. suggest that 94% of tasks in Computer & Math occupations are theoretically exposed to LLMs, Anthropic’s real-world data shows actual coverage is only at 33%.
The Insight: Theoretical potential is not the same as economic deployment. Legal hurdles, the need for human verification, and complex software integrations act as "frictions" that slow down displacement. We are in the "deployment lag" phase.
2. The "Canaries in the Coal Mine" are Young Workers
While aggregate unemployment hasn't spiked (yet), the data shows a significant "chilling effect" on new entrants.
There is a 14% drop in the job finding rate for workers aged 22-25 in high-exposure occupations (like software dev, customer service, and data entry).
Experienced workers are staying put, but the "entry-level" door is narrowing.
3. High Exposure ≠ Low Income
Interestingly, the workers most exposed to AI are actually higher-paid, more educated, and more likely to have graduate degrees. This isn't the robotics revolution of the 90s that hit manual labour; this is a white-collar transformation.
The Top 10 Most Exposed Occupations
Based on Anthropic’s "Observed Exposure" metric, these are the roles seeing the highest levels of real-world automated usage:
Occupation | Observed Exposure | Leading Automated Task |
Computer Programmers | 74.5% | Writing and maintaining software programs |
Customer Service Reps | 70.1% | Handling complaints and orders |
Data Entry Keyers | 67.1% | Entering data from source documents |
Medical Record Specialists | 66.7% | Coding patient data |
Market Research Analysts | 64.8% | Translating complex findings into reports |
Sales Reps (Wholesale) | 62.8% | Demonstrating products/soliciting orders |
Financial Analysts | 57.2% | Analysing financial info to forecast trends |
Software QA & Testers | 51.9% | Modifying software to correct errors |
Info Security Analysts | 48.6% | Risk assessments and security testing |
Computer Support Specialists | 46.8% | Resolving user hardware/software inquiries |
Theory vs. Reality:Where the Gap is Widest
The difference between what AI can do and what it is actually doing is starkest in technical and administrative fields.
Computer & Math:Theoretically 94% exposed, but only 33% observed.
Office & Admin: Theoretically 90% exposed, but significantly lower in actual automated deployment.
Methodology Spotlight: How Anthropic Measured This How did they move past theory? They utilized the Anthropic Economic Index, which analyzes millions of real-world interactions with Claude to map them against the O*NET database.
The "Observed Exposure" score isn't just a count of mentions; it is weighted based on:
Work-related Context: Distinguishing professional use from casual chat.
Automation vs. Augmentation: Full automated implementations (like API-driven workflows) receive double the weight of simple "human-in-the-loop" augmentation.
Semantic Mapping: Using advanced embedding models to match user prompts to specific occupational tasks.
What does this mean for the "Frugal" approach to science and business? It means the value of "standard" cognitive output is deflating. If a task can be observed in API traffic as "automated," its market value will trend toward the cost of the tokens. The premium is shifting rapidly toward system-level thinking, human-in-the-loop verification, and complex problem-solving that hasn't hit the "Observed Exposure" threshold yet.
The report concludes that while we haven't seen a "Great Recession" for white-collar workers yet, the signals in youth hiring are a warning we shouldn't ignore.
Source: "Labor market impacts of AI: A new measure and early evidence" by Maxim Massenkoff and Peter McCrory (March 2026).
Comments