No AI job apocalypse, but work is being redistributed

What the essay actually argues

George's essay, published May 7, pushes back on the "AI is going to wipe out jobs" narrative that has dominated business press since GPT-4. He builds the case on four pillars:

• A Federal Reserve Bank of Atlanta working paper covering roughly 6,000 corporate executives across the US, UK, Germany, and Australia. More than 90% of firms reported no AI-related employment impact over the prior three years; expected impact over the next three years lands at about –0.7% of headcount.
• A Yale Budget Lab paper from April 2026 concluding that economy-wide AI labor disruption "remains largely speculative."
• An NBER working paper finding that AI adoption has not "led to meaningful changes" in total employment.
• Research from Stanford, the Dallas Fed, and Census showing that early-career workers (ages 22–25) in the most AI-exposed occupations have seen a 16% relative decline in employment since ChatGPT's late-2022 release — but also that entry-level roles where AI is augmentative are growing.

The intellectual frame is the lump-of-labor fallacy: the assumption that the economy contains a fixed amount of work, so anything a machine does must come out of someone else's job. Economists have been knocking this fallacy down since the 19th century, and George is right to invoke it.

So the headline read is correct: the data does not show an apocalypse.

The number that doesn't go away

Three of the four sources are aggregate. One is not. The Stanford / Dallas Fed / Census strand is doing the heaviest lifting in the essay's "yes, but" section, because aggregate stability and group-level disruption are entirely compatible phenomena.

16% is a large move for a young cohort. It is the kind of move that, if it persists, shapes career trajectories for a decade. And it has a specific shape: it concentrates in the slice of work where an AI model is already a substitute for a junior person — generating boilerplate code, drafting routine documents, answering scripted questions, producing first-pass analysis — not where AI is a force multiplier for someone who already knows what they're doing.

The Atlanta Fed's –0.7% three-year expectation looks gentle on the page. But –0.7% across a workforce of hundreds of millions is a redistribution event, not a quiet one — and the redistribution is not random. It maps onto exactly the tasks that LLMs do well unsupervised.

What aggregate stats are designed to hide

Macro labor statistics are calibrated to detect recessions, not reorganizations. Total employment can be flat while every job inside it has been quietly redefined. That has happened before — the spreadsheet did not reduce the number of accountants in the 1980s, but it did change what "an accountant" meant by 1995. The post-spreadsheet accountant who could not write a model was a different worker than the pre-spreadsheet accountant who could not.

The interesting question for the next three years is not will the headcount change. It is what will the inside of the headcount look like. Three forces are doing the reallocation right now, all of them visible in the same studies George cites.

1. Substitution at the bottom of the skill ladder

The tasks that were entry points into a profession — first-pass research, junior copywriting, template legal drafts, basic data manipulation, level-one support — are exactly where AI is most productive solo. The Stanford 16% is what this looks like in the labor data. It is not catastrophic in aggregate, but it is highly local: a 22-year-old paralegal feels it even if national legal-services employment is up.

2. Augmentation in the middle

For workers with domain context — a doctor reading scans, a senior engineer reviewing PRs, a portfolio manager screening filings — AI extends reach without replacing judgment. The augmentative entry-level roles the same researchers found are the visible edge of this: the new junior job is not "draft the memo," it is "specify, evaluate, and ship the memo the agent drafted." That job did not exist three years ago.

3. Expansion at the top of leverage

A single operator who can compose tools, orchestrate sub-agents, and ship into production now does work that previously required a team. This is the part of the labor market the Atlanta Fed survey cannot see well, because it shows up as new companies, new products, and new internal teams — not as headcount reductions in existing rows of an HR spreadsheet.

The net effect — total employment roughly stable, the content of jobs shifting hard — is exactly what a non-apocalypse looks like up close. It is also why the surface-level debate ("AI will / won't take your job") is the wrong frame.

The lump-of-labor fallacy goes both ways

George is right that "AI does X, therefore X workers lose their jobs" is bad economics. But the symmetric mistake is equally bad: "AI does X, therefore the economy invents new X work for the same people automatically." It does not. Past technology transitions — power looms, tractors, spreadsheets, container shipping — created enormous net employment and produced cohorts whose specific skills depreciated faster than they could retool. Both things happened in the same data set.

The honest read of the four studies is:

• The economy is not shedding jobs in aggregate. ✓
• The mix of work inside the same headcount is shifting, sharply and unevenly. ✓
• The shift favors people who can direct AI. It disfavors people whose work AI can do unsupervised. ✓
• The outcome at the individual level depends on which side of that line you end up on, and that side is mostly a learnable skill. ✓

None of those bullets contradicts the others. All of them describe what a non-apocalyptic technology transition feels like for the people inside it.

Where this read could be wrong

Three honest counter-arguments are worth naming:

• The 16% may revert. Junior roles often dip first in any tech wave and recover once the apprenticeship pipeline reorganizes around the new tool. Three years of data is enough to take seriously and not enough to call a regime change.
• AI's frontier is moving. The studies measure 2022–2025 model capability. If the next two years deliver agents that reliably handle multi-hour tasks, the augmentation category in the middle starts looking more like the substitution category at the bottom — and the same survey would print very differently.
• Aggregate data lags. The Atlanta Fed survey reports what executives have already done, not what they are planning. A three-year lag between adoption and the headcount line is normal in workforce statistics, and the bulk of agentic deployment is still ahead of that survey's window.

So the honest position is: George's read of the past is well-supported. The forward read depends on assumptions about model capability and adoption speed that none of the four studies were designed to make.

What this means for someone building agents

If you stop at George's headline ("no apocalypse"), the takeaway is "relax." If you read the same studies with one more layer of resolution, the takeaway is the opposite of relax: the gradient between people who can direct agents and people who cannot is already visible in the labor data, and it widens with every model release.

That gradient is the part that is actually learnable. The skills it rewards are the same ones AgentWay's curriculum is built around — composing an agent loop end-to-end, wiring real tools and sub-agents through something like MCP, designing systems that survive non-deterministic behavior, evaluating before shipping, and operating in production. To land on the right side of that curve: Agent Basics is the entry point, Design Patterns and MCP & Skills are the skeleton, and Multi-Agent Systems plus Production get the work shipped.

The larger point is not which lesson to start with. It is that for the next several years, headline labor numbers may keep saying nothing is happening while your actual week at work says something is changing quickly. Trust the second signal.

The bigger question

The important question is not whether the apocalypse arrives. George has shown it does not, and the data is on his side. The issue now is what kind of worker the same data is quietly producing on the way to "stable total employment." If it is the worker who composes agents and ships systems, the next decade rewards the people who started learning that craft early. If it is the worker who waits for the apocalypse to confirm itself before reskilling, the same decade quietly leaves them behind.

There is no apocalypse coming. There is a redistribution already in progress. Those are not the same statement, and only one of them is actionable.