DOJ's 20-page brief backs OpenAI, argues AI training on copyrighted work is fair use

A licence is worth the cost of not having one, which is why a document carrying no binding legal authority ^[4] can still move a price. The Justice Department did not change copyright law on Tuesday. In twenty pages it told a Manhattan federal court that the United States has a strong interest in the court rejecting any argument that training large language models on copyrighted texts violates copyright law ^[3][26], and it called that training extraordinarily transformative, with benefits that far outweigh any competitive harm ^[6].

The scale gap between the two sides is stark. Goldman Sachs Research projects more than a trillion dollars of AI-related investment globally in 2026, of which $581bn lands in the United States ^[12], so the domestic share the brief invokes is at most about 58 cents of every global dollar ^[24]. The New York Times, which sued OpenAI and Microsoft in 2023 ^[13], has spent more than $28m on AI-related litigation according to its regulatory filings ^[17], which is one dollar of legal spend for every 20,750 dollars of American AI capex this year ^[25]. That disparity does not settle who is right; it settles who can afford to keep going, and it is the same point the government makes about small developers facing licence fees ^[7], read from the other end of the table.

The two concentration stories are the interesting part. Washington says mandatory licensing would raise entry hurdles, bolster incumbent tech firms and transfer much of the profit to publishers with the largest archives ^[7]; the Times says the administration is siding with a handful of trillion-dollar AI companies at the expense of creators whose work they stole ^[14]. The OECD's reading, that access to computing power, data and expertise is already consolidated in favour of established firms ^[11], sits underneath both. The choice on offer is between a concentration priced in licence fees and one priced in GPUs, and the filing picks the second without saying so.

What the brief declines to defend is where the residual value sits. It separates training from outputs that reproduce copyrighted material, and from how the data was acquired ^[8], which is exactly the seam the California cases split on, since a Copyright Alliance reading of Bartz v. Anthropic and Kadrey v. Meta has both courts calling training transformative while reaching different conclusions on piracy and market harm ^[9]. The US Copyright Office holds that some training uses qualify and some do not, depending on source, purpose and market effect ^[10]. What a publisher still sells is clean provenance and indemnity, the parts of the deal the brief left standing.

This is probably wrong, but the deals signed over the next two quarters look to me like data-supply and indemnity contracts at lower headline numbers, because the permission component has been publicly marked down. The counter-thesis is simple: if the court follows the Copyright Office's factor-by-factor line and treats output substitution as dispositive, damages exposure becomes real and licence prices rise. A third path is that judges treat an executive-branch statement of interest in a private suit as the non-binding paper it is ^[4], the sanctions fight over allegedly destroyed evidence prices on its own ^[18], and the only durable effect is diplomatic, since Commerce Secretary Howard Lutnick spent Wednesday urging G20 officials in North Carolina to adopt fair use while devising ways to protect artists ^[19] rather than building a licensing framework, and the EU already runs the other model, with text-and-data-mining exceptions that rights holders may reserve against ^[20].

I would drop the read if a licence signed after Friday clears at a premium to comparable pre-filing deals, or if the opinion adopts the transformative framing and negotiations continue at unchanged numbers, which would mean they were never about litigation risk.