Based on Peter McCrory's (Anthropic's chief economist) piece "Why Hasn't AI Increased Unemployment?" and The AI Daily Brief's (NLW) discussion of it. I'm summarizing and adding my own framing, not reproducing the original — the data and arguments belong to McCrory and NLW.
Everyone is worried AI is coming for their job. McCrory's piece gives a counterintuitive answer: so far, it hasn't raised unemployment at all — because it's an "augment the person" technology, not a "replace the person" one. What's actually happening isn't people getting cut. It's value quietly getting reassigned: the people who can direct and judge AI are becoming more valuable.
The counterintuitive number
McCrory starts with a calm set of facts. The US labor market is stable, close to full employment. Even in occupations "highly exposed to current AI automation," there's been no unusual rise in unemployment in recent years.
The numbers back it up: June unemployment at 4.2% (a level the Fed considers consistent with full employment), the job-vacancy-to-unemployment ratio back above 1, and prime-age employment near decades-long highs.
And this isn't because AI is "too small to show up yet" — it's the opposite. It's already big: 20% of companies use AI in at least one business function, and in the information sector that number is 40%. Quality-adjusted AI output has grown over 2000% a year, two years running. Labor productivity has climbed from a pre-pandemic 1.6% to 2%.
So the real question isn't "when will AI's impact arrive." It's already here. The real question is: everything's being used — so why hasn't it pushed people out?
The answer: augmenting, not replacing
McCrory's core judgment, in one line: AI has all the characteristics of a skill-biased, labor-augmenting technology.
Unpacked, that's three things: it complements human expertise, it relies on humans-in-the-loop to direct and judge the most complex work, and it rewards people who know how to use it well.
The evidence is compelling:
- No occupation, checked against the Department of Labor's O*NET task taxonomy, has had all of its tasks systematically taken over by Claude. And the part that hasn't been automated both caps overall output gains and amplifies the return to the humans still doing it.
- "Complex Claude output" correlates strongly with "high-quality human input" — when Claude builds a sophisticated economic model, there's usually an expert steering it behind the scenes.
- Even after six months of use, people still treat Claude as a thinking partner — and the ones who use it that way get better results. McCrory's point: if AI were good enough to just run solo, we wouldn't see this effect at all.
The keyword: jagged frontier
This is the single most memorable term in the piece. McCrory's framing: model capability is advancing fast, but it stays stubbornly jagged. Filling in those uneven pockets requires expert supervision — people who can guide these very powerful systems and catch them when they fail.
This ties the whole argument together. It's the same idea as the "jagged edge" that Anthropic product managers talk about, and the same thing this newsletter keeps returning to: the capability ceiling keeps rising, but the edges stay rough. Because the frontier is jagged, human judgment becomes the thing that fills the gaps. Whoever can spot where AI is about to stumble — and catch it when it does — is the one who's hard to replace.
Value is being redistributed: what's rising, what's falling
This is the most practically useful part for anyone reading. McCrory doesn't hedge: today's valuable skills won't all stay valuable. Some expertise will depreciate — pure coding implementation, for instance. Some will appreciate — management-adjacent skills like delegation and evaluation.
In plain terms: execution-type expertise is falling in value. Direction-and-judgment-type skill is rising.
There's also a genuinely counterintuitive finding: the more people use Claude, and the better they get at it, the more they believe it can take over more of their work — but at the same time, the less worried they are about losing their job, and the more optimistic they are about pay and job stability. The most commonly cited source of productivity, McCrory notes, is "scope" — doing more, and doing it at a higher level.
This lines up with everything else in this space right now: Replit talks about people moving from doer to director. Netflix says craft is still scarce and they're hiring systems thinkers. Mollick says the skill to practice is "managing agents." Four different sources, one conclusion: value is retreating to the layer machines can't do — and that layer is judgment, delegation, and evaluation.
But don't get too optimistic: three warnings, and one sharper insight
McCrory is careful to add three caveats:
- Hiring for young people in high-AI-exposure roles is genuinely weakening (consistent with Stanford's "canary in the coal mine" research). But it's hard to pin this on AI specifically — these years are also colliding with the largest non-recessionary labor-market slowdown on record, with low hiring and low firing, an environment that hits early-career workers hardest regardless of AI.
- High-exposure occupations aren't losing jobs now, but the Department of Labor projects slower growth for them through 2034.
- Everything could flip. Stronger agents, a flattened jagged frontier, or even AI automating "innovation" itself (recursive self-improvement) — at that point, "augmentation" could turn into "replacement."
NLW adds one point I find sharper: the self-fulfilling nature of the narrative. If the whole world keeps saying "this technology exists to cut your headcount in half," and your investors expect that too, you'll be pushed to actually do it. But if the story instead is "use it to do more, faster, and ship new products and services," that's the pattern we'll see more of.
In other words: what AI ends up doing depends partly on which story we choose to tell. This also explains a trend worth watching — teams getting smaller and more agile, with more concentrated responsibility per person. It's the same curve as "80 people building 2 products" turning into "3 people building 1."
Closing: bet the narrative, and yourself, on "augmentation"
The one-line summary: the data says AI is currently augmenting people, not replacing them — and what it's augmenting is the people who know how to direct and judge it.
Three moves for anyone trying to keep up:
- Stop positioning yourself as an "executor." Move toward "direction + judgment" — knowing how to use it, how to delegate, and how to sign off on the result is the new three-piece skillset.
- Accept that teams will get smaller and sharper. This isn't layoffs — it's everyone's scope expanding and responsibility concentrating.
- Deliberately choose the "augmentation" narrative — not just because it sounds better, but because it's self-fulfilling: however you talk about AI is largely how you'll end up using it.
Ultimately, AI has flattened a lot of "can you do this at all" thresholds — which means what's genuinely scarce has moved back to the most non-AI layer there is: the judgment to fill in the jagged edges, to direct and evaluate the machine. Value hasn't disappeared. It's just retreated into that top 5% — and that 5% happens to be a place you can still choose to stand in.
Source: Peter McCrory (Anthropic's chief economist), "Why Hasn't AI Increased Unemployment?"; The AI Daily Brief (NLW), companion episode discussing the piece (2026-07-24, ~36 min). This article is based on a transcript of that episode plus the original piece; data and views should be attributed to McCrory's original writing and the original podcast — the author has paraphrased and restructured for this summary.
Top comments (0)