AI and Jobs: What the Data Actually Shows

Opening

Dear reader, an interesting report came out two days ago. A company that builds AI decided to measure, in its own words, “how our technology is affecting jobs.”

It’s a report Anthropic published on March 5, 2026, titled “Labor Market Impacts of AI: A New Measure and Early Evidence.” Where prior research assessed job threats based on “what AI can theoretically do,” this report goes a step further. It combines that with actual Claude usage data to measure which tasks are being replaced by AI in the real world, right now.

The media’s reaction was predictable. Headlines like “74.5% of programmers at risk of replacement!” and “The end of entry-level jobs!” poured out. But when you actually dig into this report, the story the numbers tell is far more complicated — and honestly, more unsettling. Today I want to dissect this report with the eye of a data analyst. It was disappointing to see a few influencers here in Korea just parrot the flashiest numbers as-is. Shocking, threatening… that kind of vibe. The “study hard or take my course or you’ll get fired” kind of framing.

Let’s start with what’s actually new about this report

Most existing research on AI and the labor market has asked only one question: “Can AI do this task?” The 2023 study by OpenAI researchers (Eloundou et al.) is the classic example — they analyzed tasks across roughly 800 U.S. occupations and scored each one 0, 0.5, or 1 based on whether an LLM could theoretically at least double the speed of that task.

The problem is that there’s a massive gap between “can do” and “is actually doing.” Take pharmacy prescription information tasks, for example: Eloundou et al. rated AI as fully capable of performing this (β=1), yet in reality, no instances of Claude actually doing this task were observed. Real-world barriers — legal constraints, verification procedures, software integration — get in the way.

Anthropic’s new report tackles this gap head-on. It introduces a new metric called “Observed Exposure,” which combines three data sources.

First, the task lists by occupation from the O*NET database. Second, the theoretical AI-capability scores from Eloundou et al. Third — and this is the key part — actual Claude usage data collected through the Anthropic Economic Index. On top of that, they applied one more weighting: half-weight for “augmentation” use, where a human is assisted by AI, and full weight for cases where the task is fully automated via API without human involvement.

The result is striking. In theory, 94% of tasks in computer and mathematical occupations could be performed by AI — but the actual observed coverage was just 33%. That’s roughly a 3x gap between theory and reality.

Look past the numbers, and the picture changes

This is where actually reading the report carefully starts to matter.

The most widely cited figure is “74.5% exposure for computer programmers.” But to understand what this number actually means, you have to examine the methodology. That 74.5% doesn’t mean “74.5% of a programmer’s tasks are being replaced by AI.” It means “the share of tasks in that occupation that are both theoretically AI-capable and show observed automation patterns in actual Claude usage data.” Both conditions have to be met simultaneously.

Here’s something methodologically interesting, though. The actual usage data behind this report comes only from the Claude platform. Usage patterns from ChatGPT, Gemini, Copilot, and other AI tools aren’t included. That means real-world AI usage is likely broader than what this report measures. Conversely, the characteristics of Claude’s user base (known to skew heavily toward developers) could also be introducing bias into the results.

What matters more is the report’s core finding: the conclusion that “no large-scale unemployment has occurred yet.” Using U.S. Current Population Survey (CPS)¹​ data from 2016 to the present, the researchers compared unemployment rates between the top-25%-exposure occupations and the zero-exposure occupations. Running a difference-in-differences (DID)²​ analysis, they found that the change in the unemployment gap between the two groups after ChatGPT’s launch was just +0.20 percentage points — not statistically significant. And I don’t mean that as wishful thinking; I mean it’s not significant, statistically. So, claims like “74.5% are being replaced” are closer to clickbait than fact. The report itself never actually makes that claim.

The researchers themselves acknowledge a limitation here: the minimum effect size this analytical framework can detect is roughly 1 percentage point. In other words, an unemployment rate change of, say, 0.5 percentage points would simply be invisible to this methodology. The measuring instrument itself has a resolution limit.

The real signal isn’t the unemployment rate — it’s the entry rate

The most noteworthy finding in this report actually comes later on: the analysis of job start rates among 22-to-25-year-olds.

In high-AI-exposure occupations, the monthly rate at which young people start new jobs began dropping noticeably in 2024 — down about 14% versus 2022, though the researchers were careful to note this figure sits right at the edge of statistical significance. By contrast, in low-AI-exposure occupations, the youth job-start rate stayed stable at around 2% per month. And this decline wasn’t observed at all among workers aged 25 and up.

This overlaps with Brynjolfsson et al. (2025) from Stanford’s Digital Economy Lab. Analyzing actual payroll data from ADP, the largest payroll processor in the U.S., they found that employment among 22-to-25-year-olds in high-AI-exposure occupations dropped 6–16% since late 2022. For software developers specifically, the 22-to-25 workforce shrank about 20% versus 2022, while workers 35 and older actually saw a slight increase.

The fact that two studies, using entirely different data sources (CPS survey vs. ADP payroll), arrived at directionally similar results makes it more likely that this signal isn’t just statistical noise. But both studies also admit the same thing: they can’t confirm this decline is purely due to AI. Interest rate hikes, tech-sector layoffs, and the broader business cycle are all in play at the same time.

Oz’s Lens

Honestly, reading this report left me with two conflicting feelings.

One is admiration. This is the first time a company that builds AI has set out to measure its own technology’s labor market impact at this level of rigor. Laying bare the gap between “theoretical possibility” and “actual use” is meaningful in itself. Where past research stopped at theoretical estimates like “AI could affect 80% of U.S. jobs,” this report showed with data that “in reality, we’re still at about a third of that theoretical level.”

The other is caution. Anyone who’s done data analysis knows that the method of measurement determines the conclusion. This report used the unemployment rate as its core metric — but that’s actually one of the least sensitive indicators for catching AI’s early effects. Unemployment only counts “people who want work but can’t find it.” It doesn’t capture people who’ve given up looking altogether, gone to graduate school, or switched fields entirely. The Anthropic researchers were aware of this too, which is exactly why they added youth entry rates as a supplementary metric.

Something I learned building go-to-market strategy: market shifts show up in blocked entry before they show up in exit. Existing customers have inertia and don’t leave easily, but a slowdown in new customer inflow happens much faster. I think the labor market works the same way. Firing existing employees is costly and carries legal risk. But cutting back on new hiring happens quietly, with no friction at all.

So I think the real value of this report isn’t its “everything’s still fine” conclusion, but the framework it offers for what we should be measuring. And what catches my attention is the report’s closing line: the researchers named “the labor market entry path for graduates in high-AI-exposure fields” as the top priority for future research. That’s not just academic curiosity — it means the signal is already showing up in the data.

One more thing worth remembering: this report was published by Anthropic. When an AI company says “our technology isn’t taking jobs yet,” that statement can be true and strategic at the same time. Scholars like Antonio Casilli have pointed out the potential conflict of interest baked into this kind of research. Whenever you read data, you always have to ask “who published this, and why?” alongside it.

Closing

Here’s the summary. Right now, AI isn’t destroying existing jobs on a massive scale. But the door into new jobs is quietly narrowing — especially for people in their early twenties trying to step onto the very first rung of the career ladder. But just lumping all this together and saying “AI is taking human jobs” is something anyone can say. Your neighbor, your grandmother, a job-seeker fresh out of school — anyone can say that.

What’s the point of saying that? What actually matters is identifying the structural problem and thinking through solutions. If someone’s real goal is to stoke fear, they’re either being needlessly cruel, or they’re profiting off that fear by selling courses and the like.

This is a more structural problem than the sensational “end of juniors” framing suggests. If senior employees boost their productivity with AI, companies can hire fewer juniors for the same cost. Restructuring without layoffs — that may be the real nature of what’s happening right now.

If this topic interests you, I’d recommend reading the original report alongside the references below. In particular, the ADP data analysis in the Brynjolfsson study is a great source for cross-checking against the Anthropic report.

References & Further Reading

• Massenkoff, M. & McCrory, P., “Labor market impacts of AI: A new measure and early evidence,” Anthropic, 2026. : The core subject of today’s analysis. I’d especially recommend looking directly at Figure 2 (the theoretical vs. observed coverage radar chart) and Figure 7 (youth entry rates).
• Brynjolfsson, E., Chandar, B. & Chen, R., “Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence,” Digital Economy, Stanford, 2025. : A payroll-data analysis using ADP figures, and a key reference for cross-checking the Anthropic report.
• Eloundou, T. et al., “GPTs are GPTs: An early look at the labor market impact potential of large language models,” Science, 2024. : The original study behind the theoretical exposure (β) measure that the Anthropic report builds on.
• Casilli, A., “Young Workers Haven’t Been Replaced by AI — Economists Are Just Looking for Them in the Wrong Places,” 2025. : A critique of the fundamental limits of the task-based approach that equates “tasks” with “jobs.” Worth reading for a balanced perspective.
• Observed coverage dataset(task/job level): Raw data for anyone who wants to get hands-on with it directly.

The author, Kwangseob Ahn, is a professor of business administration at Sejong University and lead consultant at OBF (Oswarld Boutique Consulting Firm). He teaches statistics and data analysis — business data management and business analytics — while leading GTM and AI strategy consulting in the field, designing the seam between technology and business. He has published academic research on a memory architecture for AI dialogue systems (HEMA) and runs Daily Arxiv, a daily curation of global AI papers. He holds a master’s from Korea University’s Graduate School of Technology Management and a KMBA. He is the author of Homo Brainless: The People Who Outsource Their Thinking.

Footnotes

1. CPS (Current Population Survey): a monthly employment status survey conducted by the U.S. Census Bureau and Bureau of Labor Statistics, covering roughly 60,000 households. It’s the official basis for calculating the U.S. unemployment rate. ↩

2. Difference-in-Differences (DID): a statistical technique that measures the effect of some change (here, AI adoption) by comparing the trends of an “affected group” and an “unaffected group.” By looking at the change in the gap between the two groups, it filters out the effects of other factors (like business cycles) unrelated to AI. ↩