Sponsored by edgeful

Anthropic just put out a labor market study this week and I think most people are going to read the headline wrong. The headline is that AI hasn't increased unemployment for the workers most exposed to it. That sounds like good news. And on the surface it is.

But there's a second finding in the report that I keep coming back to. Hiring of workers aged 22 to 25 into AI-exposed jobs has fallen about 14% since 2022. Not because anyone got fired. Because the openings just aren't there the way they used to be. Companies are keeping their experienced people and not backfilling the junior seats.

That distinction matters a lot. Because it means the disruption doesn't show up in unemployment numbers. It shows up in the career that never starts.

Most research on AI and jobs looks at what AI could theoretically do. Anthropic took a different approach. They combined theoretical task capability with what people are actually doing on their platform in work settings. They call the metric "observed exposure." They gave more weight to cases where AI is fully automating a task versus just helping a human do it faster.

And the gap between what AI can do in theory and what it's doing in practice is still wide. Computer and math occupations, for instance. AI could theoretically handle 94% of those tasks. Right now it covers about 33%.

The jobs with the most coverage right now are computer programmers at 74.5%, customer service reps at 70.1%, data entry at 67.1%, financial analysts at 57.2%. About 30% of all workers have zero measurable exposure. These are cooks, mechanics, bartenders, lifeguards. Work that happens in physical space with physical people.

When you line up Anthropic's exposure numbers against Bureau of Labor Statistics projections through 2034, the pattern is straightforward. More AI exposure, less projected job growth. For every 10 percentage point increase in task coverage, the BLS projects 0.6 points less growth. Small individually. Adds up over a decade.

And the workers most exposed aren't who you'd assume from the popular narrative. They're more likely to be female, more educated, and better paid. The top exposure quartile earns 47% more than the zero-exposure group. Graduate degree holders are four times overrepresented in the most exposed category. This is a credentialed, white-collar story.

Overall unemployment for AI-exposed occupations has been flat. No meaningful difference between high-exposure and low-exposure workers since ChatGPT came out. The chart basically shows two lines that track each other.

But isolate workers aged 22 to 25 and you see something else. Job finding rates in low-exposure occupations held steady at about 2% per month. Entry into high-exposure jobs dropped by about half a percentage point starting in 2024. That's a 14% decline. And it only shows up in young workers. Over 25, the trend is flat.

This is consistent with what other researchers have found. The mechanism isn't layoffs. It's a hiring slowdown at the bottom of the ladder. The 35-year-old analyst keeps their seat. The 23-year-old who was supposed to become one finds fewer doors open.

And this is how technology disruptions tend to work historically. The incumbents hold on. New entrants absorb the adjustment. It doesn't show up in the unemployment data because the person never had the job to begin with. They just end up somewhere else, or go back to school, or take something tangential.

The way to think about it is if you're mid-career and you've spent ten years building expertise in a specific domain, this report says you're not in immediate trouble. If you're graduating this spring into data analysis or financial modeling or customer support, the entry points that existed a few years ago may be tighter than you're expecting. Not closed. Tighter. And you probably won't see it reflected in the macro numbers until well after it's already shaped the decisions available to you.