The job market is solid

Before asking whether AI has hit employment, rule out one possibility: maybe the labor market itself is worsening and masking AI's effect. It isn't. U.S. job metrics look healthy. Unemployment sits near 4.2%—full employment in the Fed's eyes. Weekly initial jobless claims are low. Job openings roughly match people looking for work; supply and demand are about balanced.

Over the same stretch, AI adoption is substantial: about 20% of firms already use AI in operations, rising to 40% in information technology. His inference: the base is large enough that if AI were really stealing jobs, it should already show up in the macro data.

AI's fingerprint shows up in productivity

There is a signal—just not in unemployment. It shows up under productivity. Labor productivity (output per hour) has grown faster in recent years than before the pandemic.

One honest caveat: there are two yardsticks for "efficiency," and they don't fully agree.

One ruler says AI is speeding things up; the stricter one is less bullish. His read: the clash itself may mean an acceleration is just starting and hasn't fully registered in the data. Either way, this section's point holds: AI's macro effect is visible in output efficiency, not unemployment.

So far, AI is amplifying human capability

So why did efficiency rise while unemployment didn't? That's the essay's central claim. On how AI reshapes work, two worldviews sit in opposition:

Killing tasks is not killing jobs

How can AI automate so many tasks without wiping out whole jobs? One sentence holds the answer—he calls it "the task is not the job." A job is a bundle of tasks. Automating a few of them is not the same as automating the job.

This isn't hand-waving. In O*NET—the U.S. Department of Labor's catalog that breaks each occupation into typical tasks—no occupation has every task systematically taken over by Claude. What remains usually needs human coordination, face-to-face interaction, or dealing with the physical world. Only people can do that work for now.

There's a deeper layer: a job was never a fixed task pack. Every major technology reshuffles tasks inside a role—kills some, strengthens others, and creates entirely new work. The net effect: human marginal product rises, not falls.

A classic (illustrative) example: after ATMs spread, many expected bank tellers to vanish. They didn't. Tellers shifted into sales and advising that need a human touch. Each branch needed fewer people and got cheaper to run, so banks opened more branches—and total teller employment rose. Machines took the "count cash" task; people moved into higher-value "serve the customer" work.

Why call it "amplification"? The evidence

This isn't a gut call. Behind it sit several findings from Anthropic's economics team, based on real Claude usage data:

The one soft spot is young workers—don't rush to blame AI

He doesn't duck contrary evidence. Weaker signs show youth hiring softening over the past year in AI-exposed roles—same direction as the Stanford Digital Economy Lab's "Canaries in the Coal Mine." He immediately adds: don't pin the whole bill on AI.

So the struggle young people face finding work may be macro, not AI. The worry is real. Treating it as proof that AI has already caused unemployment doesn't hold yet.

What would break the case

"Amplify, don't replace" is a so-far conclusion. He admits that as models strengthen, it can break: the jagged frontier—AI that's uneven and hard to predict—may fill in, and the payoff to human expertise may shrink. The biggest variable: AI may start automating innovation itself.

Past general-purpose technologies couldn't do that: a stronger internal combustion engine doesn't invent new forms of transport on its own. Bolt general intelligence onto machines, and you have an engine for the next wave of invention. In standard models, once innovation itself is automated—first, perhaps, via recursive self-improvement (RSI)—you can get an "economic singularity": unbounded growth in finite time.

Those weak links are already biting. Example: coding agents lifted generated code volume 10–20×, yet software releases rose only ~30%, and app usage didn't rise at all.

Writing got an order of magnitude faster; delivery barely kept up. Where's the jam? Coding was only ever one step in software production. Upstream you still have to decide what to build; downstream sit testing, review, integration, shipping, and actual user adoption—steps AI has barely sped up. Code flies, then slams into everything that didn't accelerate. That's a weak link: a chain moves only as fast as its slowest ring, not its fastest.

Result: code volume 10–20×, software releases only +30%, app usage flat. One link sped up; the whole chain still crawls at the slowest step.

Scaling laws are hard to argue with. Models will get better—much better. I expect faster productivity growth, and clearer signs of self-improvement. I don't think unemployment will be markedly higher a year from now, at least not because of AI.Peter McCrory

Why did efficiency rise while unemployment stayed quiet? His call: so far AI is skill-biased and labor-augmenting—whoever can use it gains value. It takes some work; people still set direction, judge, and hold the hardest pieces. One person's feasible boundary widens.

Supporting evidence points the same way: complex outputs usually have a skilled human directing; the model struggles most on the hardest work and needs human cover; longer use turns it into a thinking partner; in coding, skilled users succeed more and recover better when AI fails—pure coding falls in value, planning and recovery rise.

Automating a pile of tasks doesn't erase the job, because a job is a bundle of tasks. AI peels off a few blocks; as long as human-only work remains, the job stays—and may be worth more because people fill the gaps. Jobs were never fixed packs either: new tech reshuffles tasks and creates new work.

Take XiaoHu's day job: entry and layout can go to AI; judgment, people work, and failure recovery stay human. While those blocks remain, the job isn't gone—and human value can rise.

He doesn't dodge contrary signs: youth hiring softened in AI-exposed roles (same direction as Stanford's "Canaries in the Coal Mine"). But the macro mess of recent years—the largest non-recession employment slowdown on record, with low hire and low fire—already hits new entrants hardest. The worry is real; pinning it all on AI still doesn't stand up.