DOJ tells a judge that training AI on the news is fair use
The US government has picked a side in the defining copyright fight of the AI era, Tesla finally put its steering-wheel-less Cybercab on the road — quietly — and Oura's IPO filing revealed just how much physiological data one ring company has banked.
The Justice Department filed a 20-page statement of interest in the New York Times' copyright case against OpenAI, and it argues model training is not just defensible but "extraordinarily transformative." Filed September 1 in the Southern District of New York and signed by Associate Attorney General Stanley Woodward, it is the first time the federal government has weighed in on the wave of AI training suits brought by authors, publishers, labels, and news outlets. The brief is advisory — the government has no jurisdiction over the dispute — but it hands every lab fighting these claims a citation to the executive branch's own view of the Copyright Act, and it lands in a case that dozens of others are watching.
The reasoning is worth reading because it goes further than the tech industry usually argues in public. The brief says a training copy makes nothing available to the public, so it cannot act as a market substitute — and it spends five pages dismantling the "indirect substitution" theory that Judge Vince Chhabria floated in Kadrey v. Meta, the one ruling that went against the labs. Under that theory, the government argues, Joan Didion should have owed Hemingway royalties for typing out his stories as a teenager to learn how sentences work. Then it makes a competition argument that sounds odd coming from the state: mandatory licensing is something only the largest tech companies could afford, and the fees would function mainly as a subsidy for old mainstream media. The Times responded that the administration is "siding with a handful of trillion-dollar AI companies at the expense of the countless American creators whose work they stole."
This doesn't end anything — a statement of interest binds no one, and the government explicitly takes no position on whether a licensing regime could be made to work. But it changes the weather. When the executive branch tells a court that losing this case would hand an advantage to foreign adversaries, it reframes a copyright dispute as industrial policy, and judges do not read that in a vacuum.
Tesla's long-promised Cybercab started giving rides in Austin on Thursday, with a launch so subdued that by evening there was no livestream, no stage remarks, and no clear answer on who could actually book one. Tesla updated its robotaxi site to say Cybercab rides are available in "limited areas of Austin." State records show 45 Cybercabs registered in Texas out of 420 autonomous vehicles, against Waymo's 988 — this after Musk promised in July 2025 that the network would reach half the US population by the end of that year.
What makes the quiet rollout rational is a regulatory wall, not a technical one. Federal rules cap how many vehicles without steering wheels or pedals a manufacturer can sell; testing deployments are unlimited, but collecting fares is not. Tesla also lacks permits to operate a paid robotaxi service in California. So the company that spent two years staging Cybercab as the vehicle that justifies a $1.4 trillion market value is launching it in the geopolitical equivalent of a test track.
Oura filed for a US IPO, and the S-1 shows a hardware company that is quietly becoming a health-data company. Fiscal 2025 revenue more than doubled to $907.9 million, and the first nine months brought $1.4 billion — with membership revenue up 121% to $240.5 million at an 89% gross margin, and paid members doubling to 5 million. The filing also describes 42 billion hours of physiological readings across more than 50 metrics, which Oura says is among the largest longitudinal biometric datasets in consumer health and the thing that trains its models.
The AI story here is the asset, not the ring. Predictive health models get better as member histories deepen, which means the moat is a decade of bodies, not a sensor. The risk factors are equally data-shaped: a proposed class action alleges Oura overstated the accuracy of its sleep tracking, and warranty costs hit $84.4 million on battery defects in some Ring 4 units. Accuracy claims are the whole product when the product is a prediction about your body.
What to watch: whether other federal agencies pile onto the Times case now that DOJ has broken cover — and whether Tesla ever says how many paying passengers the Cybercab has carried.
If the government thinks AI training is fair use, should Congress be the one to write the licensing rule that courts keep being asked to invent? Tell us in the comments.
