The honest answer: it depends on two numbers we're only beginning to measure — and the answer flips over time.
The price of machine thinking just collapsed. Scroll.
The cost of running a frontier AI model has fallen roughly 750× in six years — from tens of dollars per million words to fractions of a cent.
So what happens to the value of human thinking? You'll get two opposite answers — and they're both right, at different times.
Illustrative. Each side caricatures a real strand of the academic debate; both cite real economics.
The Doom story. AI gets cheap. Machines replace cognitive workers. Wages for office jobs collapse the way factory wages did in the 1980s. The new factory worker is the middle manager.
The Utopia story. AI gets cheap. Productivity explodes. Human work that can't be automated — teaching, therapy, judgment, presence — becomes precious. Everyone gets richer, slowly.
Both camps are loud. Both cite real economics. This piece argues they're not actually in conflict — they describe different time periods of the same process. The squeeze is real. So is the rebound. They just don't happen at the same time.
Michael I. Jordan, the Berkeley computer scientist, puts it this way: a market that delivers food into New York every day is itself a kind of intelligence. AI's mistake is trying to rebuild that kind of intelligence inside one model — instead of designing the incentives between many. — paraphrased from Jordan (2025), “A Collectivist, Economic Perspective on AI”; quoted material from Jordan's ICASSP 2023 keynote
Each row is one job. Each figure is 5% of that job's tasks. As AI copies a task, the figure falls — and becomes part of the bar.
Illustrative; synthesized from task-level analyses (Epoch AI, O*NET task structure). See methodology.
Before we resolve the doom-vs-utopia debate, look at where the wave actually lands. Some jobs are mostly routine thinking — the kind a careful intern with a textbook could do. Those are the jobs AI copies most easily.
Data entry, paralegal research, and routine financial analysis sit at the top of that list.
Hands-on, in-person, relationship-based work — nurses, therapists, teachers, tradespeople — barely moves.
Keep this split in mind. The work AI can't copy is about to become the whole story.
Indexed approximations following the polarization pattern documented in Autor & Dorn (2013) and successor BLS work. See methodology.
Here is the half the Doom camp gets right. When a machine takes over part of a job, the workers doing that part get pushed out — and pay for that kind of work falls. Economists call this the displacement effect.
High-skill wages — the solid black line at top — have climbed steadily since 1980.
Low-skill, in-person service wages — the dashed gray line — have roughly held.
Middle-skill wages — the red line at the bottom — have been hollowed out for forty years. Routine clerical and factory work was the first wave to be automated. White-collar routine work — entry-level analysis, paralegal work, basic coding — is the second.
This pain is real, and it comes first. Anyone promising a painless transition is selling something. The Doom camp's mistake isn't seeing the squeeze — it's stopping the story there.
Same force, opposite signs — separated by the time it takes cheap machine capacity to become universal. Scroll to watch the slope flip.
BLS CPI selected categories, cumulative percent change 1990–2024. Magnitudes rounded; directions well-documented.
Here is the half the Utopia camp gets right. Once there is plenty of cheap machine capacity, the bottleneck stops being the AI and starts being the work only humans can do. Scarce things get more expensive.
Sectors whose productivity exploded — manufacturing, electronics, software — got dramatically cheaper over the decades.
Sectors that couldn't be sped up by machines — live performance, education, medical care, childcare — got dramatically more expensive. Not despite the productivity boom elsewhere — because of it.
Economists call this Baumol cost disease: when a job can't be sped up by machines — think teaching or nursing — its cost tends to rise over time, even as gadgets get cheaper.
The things only humans can do don't get cheaper. They get pricier. The Utopia camp's mistake isn't seeing the rebound — it's pretending it arrives without the squeeze first.
This is a teaching model showing the direction each dial pushes — not a prediction of exact wages.
Whether that long-run rebound actually happens isn't guaranteed. It depends on two things economists can measure in principle — but mostly haven't yet.
Dial A — the consumer side. When you buy a poem, a piece of code, a song, a medical opinion — do you care if a human made it? If buyers treat human-made and AI-made as fully interchangeable, the cheap AI version wins everywhere. If buyers increasingly prefer human-made, human work gains value.
Dial B — the production side. When a machine arrives in a workplace, does it replace the worker doing a task, or does it make that worker more productive? If it replaces, wages fall. If it complements, wages rise.
Drag either dial toward its human-favoring end. In this model, either one alone is enough to flip the long-run outcome from falling to rising. The only configuration where the pessimists win is both dials sitting at the unfavorable end.
What little evidence we have hints that machines often complement skilled workers in real workplaces — Dial B leans good. Whether buyers will keep paying for “human-made” once AI output is indistinguishable is genuinely unknown — Dial A is the open question.
The optimistic case is plausible. But it hangs on a number nobody has pinned down.
After Jordan (2019, 2025): the computational, statistical, and economic foundations of AI systems.
Even a great tool makes things worse if the rules around it reward bad behavior.
Jordan's argument: real AI systems sit at the intersection of three traditions. Miss any one and you misread the system.
Computing. The algorithms and the raw scale that make modern AI possible. The part that gets all the attention.
Inference. The statistics. The part that tells you when the model is uncertain — or confidently wrong.
Economics. The incentives. What people will actually do with the tool, once it is released into a world full of self-interested players.
Release a powerful AI into a world of people acting in their own interest, and the system settles at whatever those incentives reward — not at whatever's best. Whoever designs the rules decides who actually gains. — paraphrased from Jordan
Consider a drug-approval AI. The model is honest. The submitters are not.
A smarter model doesn't fix bad incentives.
The two dials in the previous act look like fixed, unknowable parameters — things we just haven't measured yet. Jordan's framework reveals something more unsettling: they are being actively moved right now, by actors with financial incentives to move them in one direction.
Platforms have strong reasons to make AI output indistinguishable from human output — eroding the consumer preference that makes Dial A favorable. Firms face cost pressures that favor replacing workers over complementing them — pushing Dial B the wrong way. The dials are not passively unknown. They are contested.
Illustrative positioning based on task structure across O*NET categories.
Put the work on a single axis. On the left, tasks where a model can do the whole job. On the right, tasks where the job is the human presence.
Data entry and routine legal research cluster on the left. Their price is collapsing.
Code generation and financial modeling sit just to the right of that — assisted, not yet replaced.
Diagnosis and complex analysis are hybrid: the model proposes, the human is liable.
Teaching, therapy, contextual judgment, trust, presence — the right edge of the spectrum — gets more valuable, not less, as the left edge gets cheaper.
Machine thinking is becoming as cheap as electricity.
That won't make human minds worthless — but whether it makes them priceless depends on two dials we're only beginning to measure, and on who gets to set them.
The future isn't a forecast. It's a choice — and some actors have already started choosing.
This piece argues a conditional claim: as AI commoditizes generic thinking, the human work that cannot be automated becomes relatively scarcer — but only under specific conditions on two substitution elasticities. The visualizations illustrate that argument; they do not constitute a forecast. The argument does not rest on any single number; it rests on the equilibrium logic.
Built as vanilla HTML + Scrollama.js
+ D3.js — no framework, no build step, CDN imports only. All datasets are hard-coded in
js/data.js with inline source notes. The full source is included with this submission for reproducibility.