Anthropic's AI Jobs Report: What 74.5% Really Means

How does what AI can do connect to actual changes in employment

On March 5, 2026, Anthropic published a report on AI’s impact on the labor market. It combined existing research assessing whether AI can be used for specific tasks with actual Claude usage data, then compared that metric against employment changes in the US.

Reading this report, I found the measurement approach genuinely interesting. But some articles introducing this research in Korea took numbers like “74.5% of programmers” and immediately framed them as job-displacement risk. I was also disappointed to see this lead into the familiar refrain that if you’re worried about layoffs, you should go study AI. You can’t read this research properly until you first check what exactly that 74.5% is a percentage of.

Actual Claude usage data, mapped task by task

The report draws on three sources: the U.S. O*NET database’s list of tasks by occupation, the Eloundou research team’s theoretical AI exposure scores, and usage data from the Anthropic Economic Index’s Claude conversations.

The Eloundou study assesses whether AI can cut the time spent on a task by more than half while maintaining a given quality bar. It distinguishes between tasks an LLM alone can handle and tasks that require additional software. This assessment alone can’t tell you whether companies have actually adopted AI for a given task.

Anthropic layered actual observed tasks and usage patterns from real Claude conversations on top of this. Usage classified as automation got full weight; usage where a human collaborates with AI got half weight. Finally, factoring in how much of an occupation’s working hours each task accounts for, they calculated “observed exposure.” This isn’t a metric that assumes zero human involvement just because someone used the API.

Calculated this way, observed exposure for computer and mathematical occupations came to 33%. That’s considerably lower than the same occupational group’s theoretical exposure of 94%. In other words, not every task assessed as technically feasible for AI actually shows up in real usage data. But these two numbers describe the computer and mathematical occupational group specifically. They shouldn’t be read as an AI adoption rate for the economy as a whole.

The 74.5% for programmers isn’t the share who got fired or replaced

The observed exposure score for computer programmers was 74.5%. The earlier figure, 33%, is the value for the broader computer-and-mathematical occupational group, which bundles together multiple jobs. 74.5% is the value for programmer specifically, as an individual occupation within that group.

That 74.5% combines the theoretical automatability of each task, the actual usage patterns observed on Claude, and the share of work time each task occupies. It doesn’t mean 74.5 out of every 100 programmers have been replaced, nor is it a finding that companies have automated 74.5% of programming work. It’s a metric built to compare AI exposure across occupations using real usage data.

There’s also a limitation in that the analysis relies solely on Claude data. What people do with ChatGPT, Gemini, or Copilot isn’t captured here. The characteristics of the users who choose Claude in the first place could also skew the results. And you can’t simply add other services’ market share to this number to estimate total AI usage.

No clear difference showed up in unemployment rates

The researchers used employment data from the U.S. Current Population Survey (CPS) to compare the top 25% of occupations by AI exposure against occupations with no observed exposure. They analyzed how the unemployment gap between the two groups changed before and after ChatGPT’s launch.

The change in the gap came out to roughly +0.20 percentage points, which was not statistically significant. In other words, this analysis found no clear evidence that unemployment rose more sharply in high-AI-exposure occupations. That said, this can’t be stretched into a conclusion that AI has no employment impact across all occupations and countries.

The researchers note that the current data can only reliably detect effects on the order of about 1 percentage point. Smaller changes could be happening in reality without showing up clearly in this analysis — that’s the limitation here.

Fewer 22-to-25-year-olds are starting new jobs

I paid closer attention to the section on young workers toward the back of the report — specifically, the share of 22-to-25-year-olds starting a job they hadn’t held the previous month. This captures job changes as well as first-time employment, so calling it a “new-hire rate” would narrow its meaning too much.

In occupations with no measured AI exposure, the rate of starting new jobs held steady at around 2% per month. In highly exposed occupations, though, the rate started declining around 2024. Using a comparison group and a 2022 baseline, the researchers estimate the decline at roughly 14%. That’s a relative decline, not 14 percentage points, and its statistical significance sits right at the borderline. No comparable decline shows up among workers 26 and older.

This is worth reading alongside Stanford’s Digital Economy Lab’s analysis of ADP payroll data. According to the researchers’ February 2026 explainer, employment for 22-to-25-year-olds in highly AI-exposed occupations fell by about 16% relative to the comparison group through October 2025. That said, the headcount this study measures and the new-job-start rate Anthropic measures are different metrics.

There’s now real reason to look at employment shifts among young workers across multiple datasets. That doesn’t mean the entire decline can be pinned on AI, though. The Stanford researchers themselves factored in interest rates and firm-level variation, and stated plainly that results obtained without an experiment can’t establish causation on their own.

Oswarld's Lens

I was glad to see an attempt to connect actual usage data with employment data. It goes beyond simply listing tasks AI can perform, letting us look at what usage is actually observed and where employment shifts show up across different groups.

At the same time, I felt that unemployment rates alone aren’t enough to judge this. The paths of people who gave up job hunting to pursue further education, or who stopped looking for work altogether, or who moved into different occupations, are hard to capture with a single unemployment figure. We need to look at both the rate at which young people start jobs and what happens to them afterward.

Something I learned while building GTM strategy is that you have to look at existing-customer and new-customer metrics together. Even if existing customers stick around, a drop in new inflow can change the whole picture for a business. So reading this report, my eyes were drawn not just to how many people are currently employed, but to how many are newly starting work.

Companies can scale back hiring plans without laying off existing employees. To determine whether AI is influencing those decisions, or whether other costs or macroeconomic factors matter more, we need to dig further into new hiring and how job seekers are moving between roles. My GTM experience is the background that made me pay attention to this metric — it isn’t evidence that AI has actually reduced hiring.

We also need to check who produced the data and what scope it measures. Anthropic is a company that sells AI, and this report is built on tasks observable within its own service. While I credit them for making the data public, we need to see whether the results replicate in independent employment data from other services.

On this point, Antonio Casilli’s 2025 critique is worth consulting. That piece addresses not this Anthropic report but the earlier Stanford study. It asks whether an occupation can really be described just by a list of tasks, and whether such analyses miss freelance or contract work that falls outside payroll data. Still, pointing out these possibilities isn’t itself proof that employment patterns have actually changed.

The gap between numbers and conclusions needs explaining

This report found no clear difference in unemployment rates, but it did find signs of a decline in the rate at which 22-25 year-olds in high-AI-exposure occupations were starting new jobs. We need to read both results together.

I find it unsatisfying when the case for AI education rests only on anxiety-inducing numbers. To actually help someone preparing for a career, you need to explain which tasks in which jobs are changing, and how companies’ hiring practices are shifting as a result. A phrase like “the end of junior employees” doesn’t capture that distinction.

One plausible explanation is that existing staff are handling more work with AI, so companies are hiring fewer new people. Whether this is actually happening, and to what extent, still needs more verification. When I look at the next batch of employment data, I plan to examine not just layoff numbers but also new job entries and career transitions among young workers.

References

• Anthropic, Labor market impacts of AI: A new measure and early evidence, March 5, 2026. Covers how exposure levels were calculated and the analysis of unemployment and youth employment.
• Stanford, Canaries, Interest Rates, and Timing, February 9, 2026. Explains the study’s original findings using ADP data and the limits of interpreting causation.
• Eloundou et al., GPTs are GPTs, Science, 2024. The theoretical AI-exposure study that Anthropic drew on.
• Antonio Casilli’s critique of the Stanford study, August 29, 2025.
• Anthropic Economic Index data. Task- and occupation-level aggregate data released alongside the report.

The author is Oswarld (Kwangseob Ahn). Current roles: Adjunct Professor at Sejong University, Strategy Consultant at INLEVEL9. Career, research, books, and recent work are kept current on the About page. Latest · July 2026: HEMA-2: A Consolidation-Aware Tri-Memory Architecture with Multi-Channel Scheduling for Lifelong Conversational AI.