Automation explains 50 to 70 percent of the rise in U.S. wage inequality between 1980 and 2016. Wages of men without a high school degree fell 8.8 percent in real terms over that span. Generative AI is not a break from that history. It is the next chapter of it, and the chapter has a name.
Daron Acemoglu calls it "so-so automation": technology just good enough to replace a worker but not productive enough to grow the economic pie. Self-checkout kiosks were the textbook case before 2023. AI agents handling junior associate work are the textbook case now.
The mechanism, stripped to its skeleton
So-so automation works in three steps, each individually rational, each collectively corrosive.
First, a technology becomes capable enough to substitute for a specific labor task. Not better. Not transformative. Just substitutable. Self-checkout did not improve the customer experience; it shifted the labor of bagging from a paid cashier to an unpaid customer.
Second, the savings flow to capital. The displaced worker is not retrained into a higher-productivity role, because no such role was created by the automation itself. The owners of the technology and the firms deploying it capture the surplus.
Third, the displaced worker re-enters the labor market and competes for whatever roles remain. Wages in adjacent occupations get bid down, not up. The "augmentation" gain economists hoped for — a senior worker becoming twice as productive — does not arrive at scale, because the tool was good enough to replace the junior, not amplify the senior.
This is the inequality machine Acemoglu's Simple Macroeconomics of AI traces in formal terms. It is not a market failure. It is exactly what efficient markets do when the technological frontier shifts toward replacement instead of augmentation.
Generative AI fits this template too cleanly to ignore
Look at what gets automated first under current deployments. Customer service triage. Paralegal summarization. First-pass marketing copy. Junior coding tickets. Audit support work. Insurance underwriting on standardized claims.
Each of these is a substitution play. None of them is paired with a credible "the senior gets ten times more productive" story. A senior litigator with an AI research assistant is somewhat more productive. She is not ten times more productive, because the bottleneck on her output was never the speed of her research assistant. It was court calendars, opposing counsel, client decisions, regulatory cycles, and the cognitive limits of a human who has to be the one in the courtroom.
So the savings show up as headcount cuts at the bottom of the org chart, where the substitutability is real, and not as a productivity surge from those who remain. That is the textbook so-so automation pattern. Acemoglu predicted it before ChatGPT shipped. The 2025 to 2026 deployment data confirms it.
Where the savings actually go
AI deployment is, on average, a transfer from labor income to capital income. The owners of the model, the cloud, the chips, and the firm deploying the model split the surplus. The displaced worker absorbs the loss. The senior worker who keeps her job sees small productivity gains that do not translate into proportional wage growth, because her bargaining position has not changed.
Labor's share of U.S. GDP has fallen from 64 percent in 1974 to 56 percent in 2024. Capital's share has risen by an exactly compensating amount. Generative AI accelerates that trend rather than reversing it, because the deployments that pencil out for firms today are overwhelmingly the substitution kind, not the augmentation kind.
This is not a moral observation. It is an accounting one. If the technology displaces a $60,000-a-year employee and saves the firm $50,000 net of API fees, that $50,000 does not appear as a wage increase elsewhere. It appears as margin, which flows to shareholders concentrated in the top decile.
The escape hatch nobody is funding
There is a version of the story where AI augments rather than replaces — where teachers use it to give every student a tutor, where small businesses use it to compete with chain incumbents, where solo lawyers can run private practice at the productivity of a mid-sized firm. That version requires AI deployments aimed at amplification, not substitution.
Almost no one is funding that direction at scale, because amplification does not show up as a clean cost line on a CFO's deck. The next political question is whether anyone — government, philanthropy, cooperatives — sees fit to fund the augmentation case before the substitution case finishes hollowing out the middle.