A Federal Court Just Ruled That Training AI on Someone Else's Work Isn't Fair Use, Five Years After the Company It Sank Went Out of Business

A federal appeals court ruled on September 29 that training an AI tool on someone else’s copyrighted work is not automatically fair use, the first time any U.S. circuit court has decided that question. The company on the losing end, ROSS Intelligence, had already been out of business for five years by the time the ruling came down.

A legal first that arrived after the defendant was gone

The case was Thomson Reuters against ROSS Intelligence, and it had been working its way through the courts since 2020. Thomson Reuters accused ROSS, a legal research startup, of copying more than 2,000 headnotes, the short editorial summaries that Westlaw’s attorney-editors write to sum up a point of law in a court opinion, and feeding them into an AI system built to compete directly with Westlaw.

Judge Stephanos Bibas had already ruled for Thomson Reuters at the district court level, and a three-judge Third Circuit panel affirmed that result on appeal. The panel held that Westlaw’s headnotes are original enough to be copyrighted, and that training a rival product on them wasn’t transformative enough to count as fair use. Judges split the four fair use factors down the middle: ROSS won on the nature of the work and the amount copied, Thomson Reuters won on commercial purpose and market harm, and the court said the market harm outweighed everything else. ROSS had built a product meant to take Westlaw’s customers directly, not to transform the headnotes into something new, which is what made the commercial purpose factor cut so clearly against it.

The irony is that ROSS never got to see how this ends. The company shut down in January 2021, saying the cost of fighting Thomson Reuters in court, not the merits of the case, was what sank it. Thomson Reuters kept pressing anyway, winning summary judgment in 2025 and now an appellate affirmation five years after its opponent closed its doors. Lawyers who track AI copyright cases say the ruling matters well beyond one shuttered startup, because it is the first time a federal appeals court has weighed in on whether feeding copyrighted material into a machine learning system is inherently transformative. The court said no, at least not when the result competes head to head with the original. Every AI company currently defending a similar fair use argument, several of which are in cases far bigger than this one, now has to contend with an appellate precedent that was not on the books two weeks ago.

SoftBank finishes paying for its slice of OpenAI

SoftBank announced on October 1 that it had wired the third and final $10 billion installment of a $30 billion commitment to OpenAI, bringing its total investment in the company to $64.6 billion for a stake of roughly 13 percent.

The payment, worth about ¥1,579.6 billion at the exchange rate SoftBank used, closes out a financing plan structured in February as three equal $10 billion tranches due April 1, July 1 and October 1. SoftBank funded this last one by drawing on a foreign-currency bond sale it had priced on September 24, meaning the company borrowed money on the public debt markets specifically to make its last payment into a private AI lab. The $30 billion commitment was itself SoftBank’s share of OpenAI’s $122 billion funding round, a round large enough that OpenAI’s other backers put in tens of billions more between them.

SoftBank’s relationship with OpenAI is now one of the largest single corporate bets on any AI company anywhere, and it is a bet made almost entirely on borrowed conviction: SoftBank has leaned on debt issuance rather than cash on hand to get through all three tranches. Masayoshi Son has staked a sizable share of his company’s balance sheet on OpenAI reaching a valuation that justifies $64.6 billion for 13 percent of it, and for now that bet is locked in, paid in full, with no more installments left to make.

The jobs report says AI has become the default excuse

Challenger, Gray & Christmas released its October 1 report on September hiring and firing, and the headline number looked almost reassuring: 43,281 job cuts, down 18 percent from August and the lowest total for any September since 2022.

AI wasn’t even the top reason cited that month. It ranked fifth, tied to 3,961 of September’s cuts. But zoom out to the full year and the picture flips. Employers have cited AI in 120,136 job cuts through September, roughly 21 percent of the 573,195 total announced so far in 2026, making it the single most-cited reason for layoffs this year, ahead of cost-cutting, restructuring and everything else on Challenger’s list. That total is already more than double all of 2025’s AI-attributed cuts combined. Technology companies alone announced 10,799 cuts in September, a 77 percent jump from August, even as the overall numbers improved.

Andy Challenger, the firm’s chief revenue officer, has described the pattern all year as less about AI directly replacing specific workers and more about where companies choose to spend afterward. “AI is now the leading reason companies give for cutting jobs,” he said, but even when a role isn’t literally handed to a model, “the money for those roles is,” redirected into AI infrastructure and tooling instead of being used to keep headcount steady. Through September, overall job cuts in 2026 are still down 39 percent from the same period last year, so there’s no mass layoff wave by the raw numbers. What’s changed is the language employers reach for first when they explain the cuts they do make, and for nine months running that language has been AI.