The People Who Most Believe AI Will Replace Jobs Are the Ones Building It
Anthropic published a report this week with a bland title — "How employment is changing at Anthropic" — written by economists on its own payroll. The contents are anything but bland.
One number stands out: Anthropic staff believe that by 2030, AI will be able to do 75% of their current work. Ordinary Americans, asked the same question, said 28%. Inside the company, more than 80% of merged code is now written by AI; more and more of what human engineers do is reviewing AI output, writing specs, and building monitoring dashboards. The report calls this "meta-work" — its share of work sessions has risen from 10% to 22%.
More striking are the hiring numbers: the share of junior employees among technical hires fell from 27% to 9% within a year. About a third of machine-learning managers admitted they are hiring less because of AI.
But what struck me first was not the numbers — it was the strange calm behind them.
The people building AI are the least afraid of being replaced by it. Seventy-five percent sounds like a doomsday prophecy, yet those who said it sounded as calm as a weather report. The report even notes that human attention has become the new bottleneck: agents now log six times the hours of employees, while 80% of that time is spent waiting for human input. Machines do not lack compute — they lack someone willing to read what they wrote.
It reminds me of a counterintuitive truth: those closest to the storm fear it least. Fishermen do not fear the sea, pilots do not fear the sky, and AI builders do not fear being replaced — because they watch AI make dumb mistakes every day. AI can write 80% of the code, but every line is reviewed before merging; AI can do six times a human's work, but every piece needs a human sign-off. Replacement is happening — hiring numbers do not lie — but "replacing" and "taking over" are two different things.
The report has an even more science-fiction number: 91% of Anthropic staff believe AI will make a Nobel-level scientific discovery by 2030. Among ordinary people, only 39% agree.
When humans debate "will AI replace me," they treat AI as an enemy at the gate. Its builders do not see it that way — they treat AI as a coworker who shows up every day: diligent but dumb, in constant need of babysitting. Replacement is not an invasion that happens one day; it is a handover, bit by bit. First you hand off the most tedious tasks, then the slightly tedious ones, and what remains is precisely the hardest part to replace: judgment, taste, and the willingness to decide.
So do not ask whether AI will replace you. Ask yourself: if AI did 75% of your work, what would you do with the time you saved? Anthropic's employees are already answering with their actions: more meetings, more meta-work, managing more AIs.
It sounds like a promotion. It also sounds like a loop.