Anthropic's New Study Reveals the Gap Between AI's Potential and Its Actual Impact on Jobs

In Brief

Anthropic published a landmark study introducing "observed exposure," a metric that compares what AI could do to what it's actually doing. The gap is wide: 94% potential use and 33% actual today.

The most exposed workers are highly educated and earn 47% more than average. Some GenZers trying to get into the workforce are also vulnerable.

Behind all of this, looms a bigger question: What happens to the social contract between workers, employers, and society when AI starts eliminating jobs?

What Happened

Anthropic's "Labor Market Impacts of AI" introduces a fundamentally different way to measure how AI is affecting jobs. Rather than relying on theoretical assessments of what AI models could do, the study creates a metric called "observed exposure" comparing large language models' capabilities against anonymized, real-world Claude usage data.

I examined this study and its ramifications with SmarterX and Marketing AI Institute founder and CEO Paul Roetzer on Episode 201 of The Artificial Intelligence Show.

The gap is enormous. AI can theoretically handle 94% of knowledge worker tasks. But in practice, Claude is only doing about 33% of those tasks today. That 61-point gap is a reflection of the current AI labor story.

The demographics of the most exposed workers challenge common assumptions: That the most educated workers are insulated. That's not the case. Workers in the most AI-exposed occupations, those who earn 47% more on average, are 16 percentage points more likely to be female, and are nearly four times as likely to hold a graduate degree.

The top three most exposed occupations: computer programmers (75% of tasks already covered by AI), customer service representatives, and those in data entry. Meanwhile, roughly 30% of those in hands-on professions, cooks, mechanics, lifeguards, have essentially zero exposure.

No systematic unemployment increase yet. Not since ChatGPT launched in late 2022. But there's a critical early warning sign: The rate at which GenZer ages 22-25, are getting hired in highly exposed AI fields has dropped approximately 14% compared to 2022 levels.

The Key Numbers

• 94% — Share of knowledge work tasks AI can theoretically handle
• 33% — Share of those tasks Claude is actually covering in real-world usage
• 47% — How much more workers most vulnerable to AI earn compared to the average
• 75% — Task coverage rate for computer programmers, the most exposed occupation
• ~14% — Drop in job-finding rate for GenZ (ages 22-25) entering AI-exposed fields 
  since 2022
• ~30% — Share of workers with zero AI exposure today

Why the Gap Won't Protect You

This study matters because the AI labs themselves are shifting how they think about measuring progress. 

"They're realizing that these benchmarks they've previously been looking at that are mainly testing for IQ are saturated, and we're not going to learn much from model to model," Roetzer says.

This real-world impact measurement signals the industry is moving away from academic benchmarks and onto economic impact instead.

The backward-looking data is dangerous. The 61-point gap between theoretical and actual, observed exposure might feel reassuring. But don't be fooled. "I think in some ways it gives people this false sense of security," Roetzer said. The reality is that "looking backwards isn't going to tell us anywhere about where we're going."

The observed exposure data has its own limitations. 

 "The observed exposure right now is coming from the use of their APIs. So it's an imperfect thing because in the fall when we talked about Anthropic's efforts to look at this, it was almost exclusively being used as a coding tool.

"And now that Anthropic is becoming way more popular, even in the last 30 days, you're going to start to get a much better representative data set of how the exposure is spreading."

—Paul Roetzer, founder and CEO of SmarterX

As more industries adopt Claude, the 33% figure will likely climb and fast.

Roetzer has been building toward this kind of task-level analysis for years. He pointed to his own JobsGPT tool at smarterx.ai/jobsgpt, which draws on the O*NET database. "The O*NET database breaks down like 800 professions. You can go in and you can put in a job profession and it breaks it down into 20 or 25 tasks that make up that job."

This task-by-task lens is the only honest way to assess exposure. Not job titles, not industries, but the actual work people do every day.

Then there's the bigger question: What happens when the gap closes?

Roetzer recently listened to Peter Diamandis interview business leader Andrew Yang on his podcast, Moonshots, about the future of work. Particularly, the social contract, the implicit deal between workers, employers, and society about how work should function. When AI starts eliminating jobs at scale, that deal breaks. Yang discussed universal basic services, a model where citizens pay $250 per month and receive healthcare, education, and retirement coverage. Roetzer raised another idea: an automation tax.

"Is there a tax on automation? So if you're claiming you laid off 4,000 people because of AI, not only do you have to pay unemployment, but you have to pay an AI tax."

—Paul Roetzer, SmarterX founder and CEO

These aren't polished policy proposals. They're the early thinking of people trying to get ahead of a problem that hasn't fully materialized but is clearly coming. As Roetzer put it: "We are nowhere near actually having solutions. But we actually do have some people thinking deeply about this, including Andrew Yang."

SmarterX Take

The human side of this conversation doesn't get enough attention. The assumption is that corporate leaders are eager to replace workers with AI. The reality, according to Roetzer, is far more complicated.

"I've done a lot of workshops. I've done a lot of public talks. I've spent a lot of time with executives and entrepreneurs, government leaders.

"No one wants to fire people. When we talk about the inevitability of jobs going away, I have yet to meet a CEO who takes joy in getting to lay off 20% of their workforce. No one wants to do it, but there are going to be CEOs forced to do it."

—Paul Roetzer, founder and CEO of SmarterX

So what happens when that happens? What obligations exist when this day arrives? obligations from employers to displaced workers, from companies to communities, from governments to citizens whose livelihoods evaporate not because of poor performance but because the economics shifted underneath them.

What to Watch

• GenZ hiring rates. That 14% drop in job-finding rates for entry-level workers entering AI-exposed fields is the earliest concrete signal of displacement. Will it accelerate or spread to other age groups?
• Company-level evaluations: Will organizations start building their own observed exposure analyses internally, using tools such as JobsGPT to assess their workforce task-by-task before the market forces their hand?
• Social contract in mainstream policy. Andrew Yang has been ahead of this conversation for years. Watch whether automation taxes, universal basic services, or similar proposals enter serious policy debate, especially heading into the midterm election cycle.
• The observed exposure number itself. As Claude's user base diversifies beyond developers, the actual use figure will shift. Anthropic has committed to updating this data. Each update will tell us how fast the gap is closing.

Resources

• Anthropic: Labor Market Impacts of AI: A New Measure and Early Evidence
• Alex Imas: What Is the Impact of AI on Productivity?
• Bloomberg: Bank of England to Scenario-Plan an AI Shock Amid Job Loss Fears (paywalled)
• Ajeya Cotra: I Underestimated AI Capabilities (Again)
• Fortune: A "Great Recession for white-collar workers" is absolutely possible
• Axios: Anthropic launches AI job destruction detector

Heard on Episode 201 of The Artificial Intelligence Show: This story was covered in depth on the podcast with Paul Roetzer and Mike Kaput. Check out the full discussion for Paul's complete analysis of the Anthropic study, his breakdown of how JobsGPT works, and the Andrew Yang social contract conversation. Listen →