A report citing multiple sources came out on March 14 (local time) saying Meta is internally discussing a plan to cut more than 20% of its total workforce. The company called the report speculative, and when and how much to cut has not been settled inside the company either. Nine days before that, on March 5, Anthropic released a report examining how much AI is used in the work of 800 occupations, and said it had found no evidence that unemployment has risen in occupations highly exposed to AI since the end of 2022. The two pieces of news, which came within a week, point in opposite directions.
Meta employed about 79,000 people as of the end of 2025. The company let go 11,000 workers, about 13% of the total, in November 2022, and in 2023 it announced it would cut 10,000 more jobs. Amazon in January 2026 put forward a plan to cut 16,000 positions, about 10% of its total workforce, and the payments company Block said it would cut more than 4,000 of its 10,000 employees, citing the rapid advance of AI models.
If you line up only the layoff news, the conclusion looks like a single one. But the "observed exposure" measure Anthropic built draws a different picture. Observed exposure is a value that measures how much of the work in a given occupation AI is actually handling. Instead of asking whether an occupation has disappeared, it looks at how far the tasks within that occupation have moved over to the machine side.
Looking at the figures, computer programmers rank highest at 74.5%. Customer service representatives follow at 70.1%, data entry workers at 67.1% and medical records specialists at 66.7%, with market research analysts and marketing specialists at 64.8%. Sales representatives in wholesale and manufacturing, except technical and scientific products, come in at 62.8%, financial and investment analysts at 57.2%, software quality assurance analysts and testers at 51.9%, information security analysts at 48.6% and computer user support specialists at 46.8%.
At the other end is 0%. Cooks, motorcycle mechanics, safety attendants, bartenders, dishwashers and locker room attendants fall here. Work that is finished only when hands and bodies are on site registers no exposure. That stands in exact contrast with white-collar office jobs occupying the top of the list.
What is interesting is the makeup of the people in the highly exposed occupations. Anthropic wrote that workers in these occupations tend to be older, female, more highly educated and more highly paid. It also said that while unemployment has not risen, hiring of young workers in those occupations has slowed. Existing staff keep working while the door for newcomers narrows.
A domestic survey has pointed to a similar spot. In April last year, in its report "AI-Driven Replacement and Change in White-Collar Jobs," the Korea Employment Information Service analyzed AI job replacement rates for some 500 occupations and named patternmakers as the occupation with the highest replacement rate. Broadcast writers, game graphic designers and voice actors also came out high. Anthropic's observed exposure and the Korea Employment Information Service's job replacement rate are calculated in different ways, so the two values are hard to compare directly.
The language on the corporate side is changing too. Meta CEO Mark Zuckerberg said in January this year that there are growing cases in which one outstanding person handles a project that used to be assigned to a large team. Reports also came out that the same company plans to put $600 billion into data centers by 2028, that it acquired Moltbook, an SNS platform for AI agents, and that it is pursuing a plan to buy the Chinese AI startup Manus for $2 billion. The arrangement of cutting people and pouring money into machines is clear.
The fork that separates the two narratives lies in how many of the various tasks that make up an occupation get peeled off, and how fast that peeling proceeds. Whether an entire occupation disappears does not become the fork. A programmer with 74.5% exposure is not in a state where the occupation is gone.
A substantial share of the work is simply being handled differently. That is also why layoff announcements and exposure figures look out of step.
Whether Meta's layoff plan will actually be confirmed, and how far the affected divisions will extend, has not been decided. Counting from November 30, 2022, when ChatGPT appeared, three years and four months have passed, and over that period hiring statistics responded fastest at the entry level. Writing out your own job as roughly ten tasks and marking which of them could pass into someone else's hands is more useful than reading a list of vanishing occupations. If you open the same sheet again next year and count how many more marks there are, you can see how fast this change is for you.
