AI Can Do More Than We Use It For: The Massive Gap Between Capability and Reality

The Research That Changes How We Talk About AI and Jobs

On March 5, 2026, Anthropic published what may be the most important piece of research on AI and employment to date: "Labor market impacts of AI: A new measure and early evidence." Unlike previous studies that relied purely on theoretical assessments, this paper introduces a new metric called observed exposure that combines what AI could do with what it is actually doing in the real world.

The findings are both reassuring and sobering. The reassuring part: AI is far from replacing most jobs today. The sobering part: the theoretical ceiling for AI disruption across white-collar work is enormous, and the gap between capability and adoption is closing every month.

This is Part 1 of our three-part deep dive. In this post, we break down the study's core framework and what the massive gap between AI's theoretical capability and observed real-world usage means for career planning.

Understanding the Radar Chart: Two Very Different Stories

The most striking visual from Anthropic's paper is a radar chart (Figure 2 in the original) that overlays two measurements across 22 major occupational categories:

The contrast is dramatic. The blue area (theoretical capability) sprawls outward like a large, spiky polygon, covering the vast majority of tasks in knowledge-work categories. The red area (actual usage) is a small, tight cluster near the center. In every single category, observed usage is a fraction of what's theoretically possible.

The Numbers That Matter: Category by Category

Let's walk through what the radar chart reveals for the most important occupational categories:

Why the Gap Exists: Five Barriers to AI Adoption

If AI can theoretically handle 94% of Computer & Math tasks, why is actual usage only at 33%? The research identifies several key friction points:

The Closing Gap: Why This Matters for Career Planning

Here's the critical insight: the gap between blue and red is not static. It's closing. As AI models improve, as enterprise software integrates LLMs, as regulations adapt, and as organizations experiment with automation, the red area on that radar chart will expand outward toward the blue.

The research validates this trajectory: 97% of tasks actually observed in Claude usage fall into categories rated as theoretically feasible. This means real-world usage is tracking theoretical capability almost perfectly—it's just not there yet in volume.

Think of it this way: the blue area is a map of where AI will go. The red area is where it already is. The space between is your window of opportunity.

A New Way to Measure AI Risk: "Observed Exposure"

Previous attempts to measure which jobs AI would affect relied on theoretical assessments—experts guessing what AI could do. The problem? Past predictions about technology and jobs have a poor track record. The paper notes that a prominent study on job offshorability identified a quarter of US jobs as vulnerable, but a decade later, most maintained healthy employment growth.

Anthropic's "observed exposure" metric is different because it combines three data sources:

1. O*NET database: The US government's catalog of tasks for ~800 occupations
2. Eloundou et al. (2023) theoretical ratings: Whether each task could theoretically be doubled in speed by an LLM
3. Anthropic Economic Index: Actual Claude usage data showing which tasks are being performed with AI in practice

The measure weights automated use (API integrations, fully automated workflows) more heavily than augmentative use (humans using AI as a helper). This is important: a customer service chatbot handling tickets autonomously signals more disruption potential than a lawyer using Claude to brainstorm case strategy.

The Historical Lesson: Predictions Are Hard

The researchers are refreshingly humble about prediction. They point out that:

• Studies on the employment effects of industrial robots reach opposing conclusions
• The scale of job losses from the China trade shock continues to be debated
• The government's own occupational growth forecasts have added little predictive value beyond simple trend extrapolation

This is precisely why the "observed exposure" approach matters. Instead of guessing the future, it measures the present and tracks how the gap closes over time. It's a thermometer, not a crystal ball.

The Bottom Line

The radar chart tells a story of two realities:

For career planning, this gap is both a warning and a window. The theoretical coverage tells you where disruption will eventually reach. The observed coverage tells you how much time you have to prepare. The smart move is to use that time wisely.

In Part 2, we dive into which specific occupations are most exposed and what the demographics of at-risk workers actually look like. The findings may surprise you.