Fair Use Just Made AI's Data Free: Own the Scarcity

On September 2, 2026, the U.S. government picked a side in the biggest unresolved question in the AI economy: whether training a model on the world's copyrighted words is theft or fair use. In a 20-page brief filed in New York federal court, the Justice Department told the judge hearing The New York Times' copyright suit against OpenAI and Microsoft that training AI on copyrighted material is fair use — the first time the federal government has weighed in on the question, and the clearest signal yet on which way the law may break.

The stakes are not literary. They are about where the profit in artificial intelligence ends up. If the government's position wins, the single most expensive input to building a frontier model — hordes of human-written text — becomes, for licensing purposes, effectively free and abundant.

That is the kind of policy change that quietly rewrites who captures the value on the other side of the AI buildout. The economist's test applies: when something becomes abundant, value migrates to whatever is still scarce.

Training data is about to be declared abundant. What stays scarce is compute, and the capital needed to buy it. The five largest U.S. cloud and AI infrastructure providers — Microsoft, Alphabet, Amazon, Meta, and Oracle — have committed between $660 billion and $690 billion in capital expenditure for 2026, roughly double last year's level. Microsoft alone expects to spend about $175 billion on AI infrastructure this calendar year, and more than $50 billion in the latest quarter. A rule that makes data free does not distribute that spending across the industry; it concentrates all the cost weight on the machines the small players cannot afford.

Here is the violation hiding inside the government's reasoning. The DOJ frames fair use as a pro-competition move: force AI companies to pay for licensing and you hand the field to big tech, entrench legacy media, and lock out independent publishers and authors who might use AI tools themselves. But look at the mechanics. The teams that actually train frontier models are not the ones the brief claims to protect. They are the ones whose advantage was always capital. Making the input free does not democratize the outcome; it makes capital rarer relative to everything else, and therefore more valuable. The "pro-competition" brief is, in effect, a subsidy for the capex giants.

The government is not shy about the motive. Its lawyers argue that restricting training would compromise national security by ceding a "competitive advantage to foreign adversaries who are not so encumbered." The Times answered in kind, saying the administration is "siding with a handful of trillion-dollar AI companies at the expense of the countless American creators whose work they stole."

A fair reader should weigh the case against the government before trusting the direction, because two of its limits are real.

First, a brief is a recommendation, not a ruling, and it runs against prior ground. A federal judge in a separate authors' case against Meta found that AI training required payment, on an "indirect substitution" theory the DOJ now urges this court to reject. And in the Times case itself, a judge in late 2025 refused to dismiss claims based on what ChatGPT outputs, even as the case's focus narrowed. The government's own brief leans on the distinction: training may be fair use, but reproducing a copyrighted passage in an output could still be infringement. It pushed the judge to test whether AI creates "significant substitutive competition" with the original work rather than whether it costs the publisher readership — a standard that could bind even a friendly ruling.

Second, the AI companies are already hedging with their actual dollars. OpenAI and Microsoft have spent the past two years signing voluntary content deals with publishers, and Microsoft launched a marketplace for licensing AI training data. They are paying some creators even while their lawyers argue they should not have to. That is a tell about where the lasting value sits: if content owners cannot charge for bulk training data, they may still command premium prices for the few corpora whose absence would measurably degrade a frontier model's output. Meta's and the publishers' arguments have not disappeared; they have been rerouted to the outputs and to the highest-quality data.

None of this settles copyright law. A single amicus brief from the executive branch does not bind the next judge, and the same government could flip its position in four years. But it does tell you which side the state's weight is on: data stays cheap, so the economic surplus of the buildout concentrates in the one thing that cannot be made abundant — the machines and the capital behind them. When a critical input is ruled abundant, the disciplined move is not to buy the input. It is to buy the scarcity.