US government files in the OpenAI copyright case, September 2 — 'training is fair use'
The United States government filed a statement of interest in Manhattan federal court on September 2, 2026, in the copyright case brought by the New York Times against OpenAI and Microsoft, arguing that training large language models on copyrighted text is generally fair use. It is the first formal intervention by Washington in the three-year wave of copyright litigation brought by authors, publishers, record labels and news organisations over AI training data. The brief said the United States has a strong interest in the court rejecting any argument that training LLMs on copyrighted texts violates copyright law, citing scientific advancement and national security and warning that constraining training would put the country at a disadvantage in global AI competition. A statement of interest is not binding on the court and decides nothing by itself. The Times sued in 2023, alleging that millions of its articles were ingested without permission to build ChatGPT
The three lines
- First — Washington's first formal filing in any AI training-data copyright case
- Argument — LLM training is fair use, grounded in scientific progress and national security
- Limit — a statement of interest does not bind the court; the judge may weigh or ignore it
Key questions
- Did the US government take OpenAI's side?
- **On one issue in the case, yes — and that position matches OpenAI's.** The vehicle is a **statement of interest**, a filing through which the government, not a party to the suit, tells a court that a national interest is at stake. The reported core sentence: **'The United States has a strong interest in this court rejecting any argument that training LLMs on copyrighted texts violates copyright law.'** The grounds given were scientific advancement, **national security**, and the argument that restricting training would disadvantage the United States in global AI competition. What it is not: a ruling, or anything binding. The judge may weigh it or set it aside. What it is: a document from the executive branch landing on the pile a federal judge reads before deciding a question with no settled answer.
- Why is this a big deal?
- **Because the government had said nothing until now.** For three years, authors, publishers, record labels and news organisations have filed copyright suits against AI companies while Washington stayed out of all of them. This filing ends that silence, and it lands **on the AI companies' side**. The contrast with the same week is sharp: five days earlier, on August 28, **35 music companies including Sony and Warner sued Anthropic**, seeking up to 150,000 dollars per work. Private litigation is expanding while the government moves to defend training. For readers outside the United States this is not remote. Korean copyright law has its own fair-use provision, but it is applied differently from US fair use, and whether to write a text-and-data-mining exception for AI training remains unsettled — while most commercial models used by Korean companies were trained in the United States.
- What happens to the New York Times case now?
- **One issue gained weight; the case did not end.** The Times sued OpenAI and Microsoft in 2023, alleging millions of its articles were ingested without permission to build ChatGPT. The dispute has two layers. **(1) Is training itself fair use?** — the layer this filing addresses. **(2) What about reproduction**, when a model emits source text more or less verbatim? That is fought separately, and permission to train does not settle it. Earlier rulings also turned on a third distinction that this filing does not touch: **how the material was obtained**. Training on lawfully purchased copies has been treated differently from assembling a library of pirated ones. That separation is covered in "What fair use means for AI training data — why bought books pass and downloaded ones do not."
Washington has taken a position on AI training data for the first time. It took the AI companies' side.
1. The filing
| Item | Detail |
|---|---|
| Document | Statement of interest |
| Court | Manhattan federal court (SDNY) |
| Case | New York Times v. OpenAI and Microsoft (filed 2023) |
| Position | Training LLMs on copyrighted text is generally fair use |
| Grounds | Scientific progress · national security · global AI competition |
The reported core sentence: "The United States has a strong interest in this court rejecting any argument that training LLMs on copyrighted texts violates copyright law."
A statement of interest is not a ruling. It is how the government tells a court that a case touches a national interest, and it binds nobody. Its weight is entirely in the fact that a judge with an unsettled question now has an executive-branch brief in the stack.
2. Why the timing is the story
| Date | Event |
|---|---|
| 2023 | New York Times sues OpenAI and Microsoft |
| 2023–2026 | Suits multiply from authors, publishers, labels, newsrooms. Government silent |
| August 28, 2026 | 35 music companies sue Anthropic, up to 150,000 dollars per work |
| September 2, 2026 | US government files for the training-is-fair-use position |
Private litigation is expanding and the government has moved the other way. That divergence, not the legal argument itself, is what changed this week.
3. The case has two layers
| Layer | Question | This filing |
|---|---|---|
| ① Input (training) | Is training on copyrighted works fair use? | Addresses this |
| ② Output | What if the model reproduces the source text? | Does not address it |
Layer ② survives regardless. A model that emits article text close to verbatim is judged on different ground than the act of training.
Earlier rulings turned on a further distinction the filing does not reach: how the corpus was obtained. Training on lawfully purchased or licensed material has been treated differently from accumulating pirated copies. That split is covered in "What fair use means for AI training data."
4. Why this reaches Korea
Korean copyright law contains its own fair-use provision, but it is applied through a different test than US fair use, and the question of a text-and-data-mining exception for AI training has not been settled.
The practical channel is simpler than the doctrinal one. Most commercial models Korean companies build on were trained in the United States, and the price of data licensing deals moves with these rulings. A US decision does not bind a Korean court, but it does reprice Korean contracts.
5. What is still open
- Who filed — the department and signatories were not specified in reporting.
- The date — outlets place the filing on September 1 or September 2.
- The text — the full brief was not reviewed here; quotations come from reports.
- The court's response — how the judge treats the filing, and the next hearing date, are not public.
- Other cases — the effect on suits against Anthropic and others is unknown.
Sources
- Quartz — Trump administration backs OpenAI in NYT copyright lawsuit
- Gizmodo — Trump Admin Teams Up With OpenAI in Its Fight Against the New York Times
- TheWrap — Trump Administration Backs OpenAI in New York Times Copyright Fight
- GV Wire — US Government Backs OpenAI in New York Times Copyright Case
- Crypto Briefing — Trump administration supports OpenAI in NYT copyright lawsuit, citing national interest