A stack of folded newspapers piled on a newsstand shelf.

The Justice Department backed OpenAI on training data, but your copyright risk sits on the output side

The DOJ OpenAI copyright brief backs training as fair use, but it is not binding and treats outputs separately. Marcus Laporte on what buyers should check.

Nineteen pages, filed September 1. In them the Justice Department tells a New York judge that training an AI model on copyrighted writing is fair use. That’s good news for model makers. For a firm that buys AI tools, the DOJ OpenAI copyright brief settles very little, because your copyright risk sits in what you publish and where you sell.

The filing landed in the consolidated New York case where the New York Times and other publishers are suing OpenAI and Microsoft, according to Bloomberg Law. The government isn’t a party, the publishers have pushed back, and Judge Sidney Stein need not agree with anyone. Our earlier look at AI-generated video and copyright risk drew the same line between how a model learns and what a business puts out.

Key points

  • The brief argues training is fair use, asks that output be judged separately, and concedes some output uses may lack a new purpose.
  • A statement of interest binds no one. It still shapes settlements and licensing talks.
  • Your exposure is outputs, contracts and markets. Check your plan’s indemnity, your editor, your EU sales and your “original work” promises.

Per a legal summary of the filing, the government argues training changes the works’ purpose so thoroughly that it’s fair use, and warns that heavy licensing requirements could entrench a few large technology firms. The same summary says the brief rejects a theory that AI output can harm a publisher’s market without copying its text. Gizmodo’s report adds that the brief wants output considered apart from training.

Advertisement

That concession is the line a buyer should read twice. The brief itself allows that reconstructed or redistributed material may not count as a new purpose. Training and output are separate questions, and you only ever touch the second.

A caveat on our method: we haven’t read the filing itself. This column rests on four secondary reports, and they differ on whether the brief makes a national security argument, so we leave that point aside.

Your exposure Does a training win help?
Output you publish No. If a marketing lead pastes three paragraphs that closely track a publisher’s article onto your blog, a training ruling does nothing for that page.
Sales into Europe No. EU law has no general fair use defence.
Client contracts promising original work No. That promise is yours.

On Europe, The Next Web notes that EU law offers a fixed set of named exceptions, and that the AI Act’s Article 53 asks general-purpose model providers to respect rightsholders’ opt-outs. It also reports a Munich court ruling against OpenAI over memorized song lyrics, now under appeal. German courts are already testing AI liability, as our piece on the Munich ruling over AI disclaimers explains, and AI disclaimer liability now has a court behind it. Authorship is just as unsettled, where the Claude watermark leaves the rule to you.

Indemnity: coverage, carve-outs, control

OpenAI’s Copyright Shield, announced in November 2023, reportedly covers ChatGPT Enterprise and API customers for output claims. Free and Plus users are left out. Proskauer’s summary doesn’t spell out the carve-outs, so read your current terms. If your team writes client work on a personal plan, you may have no indemnity at all.

Three points decide what any indemnity is worth: what it covers (outputs, inputs or both, on which plans), the carve-outs, and who controls the defence. Carve-outs usually remove cover if you ignored a filter, fed in material you didn’t own, or edited the output in ways the vendor calls material. An indemnity you can’t invoke without the vendor’s consent is weaker than it reads.

A worked example, our own invention. A 60-person agency posts an AI-drafted client article from a personal-tier account. A publisher alleges the piece reproduces its reporting. The agency promised the client original work. The vendor’s indemnity, if one exists on that plan, is a contract the client never signed. The agency holds the risk, and each link can be checked against its own paperwork.

Who owns the work is a separate matter, which our piece on AI-generated work and copyright takes up. To compare vendor terms, start with our ChatGPT, Claude and Gemini business comparison. Our look at the Musk v. OpenAI verdict and vendor contracts makes the same case from another dispute, where the Musk OpenAI verdict turned on a calendar.

The edit trail, and the date to watch

Keep the prompt, the raw draft and the edited version where a colleague can find them, with a name and date on the edit. Long passages from a tool don’t go out untouched; a paragraph is a reasonable line. If someone alleges copying, you can show a human read and changed the page. If your indemnity excludes material edits, you’ll know which edits you made. Give the job to whoever signs off client deliverables. The prompt writer is rarely the person who gets the letter.

The counter-argument: if government and probably the courts treat training as fair use, AI gets cheaper, vendors defend customers, and small firms should get on with it. The brief does tilt training toward model makers. “Training is probably fine” and “we’re covered” are still different claims, and the second depends on a contract most owners haven’t opened since they clicked accept. We’d change our view if Judge Stein sided with vendors on outputs as well as training, or if the big vendors extended indemnities to every paid tier without carve-outs. This is general information; we give no legal advice and can’t predict the ruling.

Reply briefs on the summary judgment motions are due November 6, 2026, per the legal summary. Open your plan’s terms before then.

Frequently asked questions

No. A statement of interest gives the court the government’s view and doesn’t bind it. Judge Sidney Stein will decide the fair use question, and the ruling could be appealed.

Am I covered if my team publishes AI-written content?

It depends on your plan and your vendor’s terms. OpenAI’s Copyright Shield reportedly covers ChatGPT Enterprise and API customers for output claims, and leaves out free and Plus users. Read the carve-outs in your current agreement.

Does a US fair use ruling apply in Europe?

No. The EU has no fair use doctrine and relies on a closed list of exceptions, so a US win on training would not carry over to sales or publishing there.

Archive entry dated September 7, 2026, written and fact-checked on October 8, 2026. Sources are linked on the claims they support.

Total
0
Shares
Prev
GPT-6 Astra costs 2.5x the GPT-5.6 Sol promo and wants a real test
Shelves of jars with price labels in a shop.

GPT-6 Astra costs 2.5x the GPT-5.6 Sol promo and wants a real test

GPT-6 Astra pricing is $10 input and $50 output per million tokens, against a $4

Next
Why OpenAI’s $600-a-day agent habit is the wrong benchmark for your team
Rows of server racks with red and blue cables in a data centre.

Why OpenAI’s $600-a-day agent habit is the wrong benchmark for your team

Set your AI agent token budget by the finished task

You May Also Like