A government brief is not a court ruling, so nothing about how you build changed this week.
That is the whole verdict, and I want to spend the rest of this backing it up, because the headlines are doing something unhelpful. The U.S. government has sided with OpenAI in the New York Times lawsuit, filing a 20-page brief defending the unlicensed use of copyrighted material to train large language models. The language is direct: “The United States has a strong interest in this court rejecting any argument that training LLMs on copyrighted texts violates copyright law.” The reasoning leans on scientific advancement and national security, and argues that making it harder for models to train on copyrighted content would be “inconsistent with basic copyright law principles” and would thwart creative and scientific progress.
If you build bots for a living, that is a meaningful signal about which way the political wind is blowing. It is not permission.
What a brief actually is
The government filed an argument. A judge will decide whether to agree with it, and judges are not obligated to. Fair use in the U.S. is decided case by case, weighed against a set of factors, and a filing from the executive branch is input to that process rather than an outcome of it. Until there is a decision, the legal question sitting under every pretrained model is still open.
I bring this up because I have already seen the takeaway mutate in developer channels into “training on scraped data is legal now.” It is not what happened. What happened is that one party with a lot of weight showed up on one side of a live dispute.
The part that matters for people shipping bots
Most of us are not pretraining foundation models. We are wiring APIs together, building retrieval pipelines, scraping documentation, caching content, and shipping assistants that answer questions on top of somebody else’s text. That is a different legal posture than the one this brief addresses, and the distinction is worth keeping straight in your head:
- Pretraining is the activity under argument here: ingesting large volumes of text to learn general patterns.
- Fine-tuning on a narrow corpus of somebody’s copyrighted work looks a lot less like the broad, transformative use being defended, because the output is closer to the source.
- Retrieval-augmented generation puts copyrighted text directly into a prompt and often directly into your response. That is a copying-and-distribution question, not a training question.
- Scraping brings in contract and terms-of-service issues that sit entirely outside copyright. A site’s terms can bind you even where fair use would protect you.
None of those four get resolved by a brief about training. If your RAG bot returns three paragraphs of a paywalled article verbatim, no amicus filing helps you.
Build like the question is still open
The engineering response to legal uncertainty is not to pick a side. It is to make your system explainable and reversible. A few habits that have saved me:
- Log provenance for everything you ingest. Source URL, fetch timestamp, license if stated, robots directives at time of fetch. Store it next to the embedding, not in a separate spreadsheet nobody updates. If you ever have to answer where a chunk came from, you want a query, not an archaeology project.
- Keep your corpus swappable. If a source becomes a problem, you should be able to delete its chunks and rebuild the index without touching application code. Tag every chunk with a source ID and make deletion a one-liner.
- Constrain output length on retrieved text. Summarize and cite rather than reproduce. Cap how much source text can appear in a response. This is a prompt and post-processing concern, and it is cheap to add.
- Attribute by default. Link back to the source in every answer. Good for users, good for trust, good if anyone ever asks questions.
- Prefer licensed and permissive sources where quality allows. A slightly weaker corpus you fully control beats a better one you cannot defend.
Why the national security framing should get your attention
The brief’s reasoning is the interesting part for anyone planning more than a quarter ahead. Tying model training to scientific progress and national security moves this out of a narrow publisher-versus-platform fight and into policy territory. That framing tends to produce durable positions. It also tends to produce carve-outs and conditions, and those conditions usually land on the people building downstream, not on the labs.
So the practical read is straightforward. The political environment is tilting toward permissiveness for large-scale training. The legal question is unresolved. And your obligations as someone shipping a product to real users have not moved at all. Build the audit trail now, while it is a design choice rather than a scramble.
🕒 Published: