Somewhere in Austin or Amsterdam, a four-person team is shipping a support assistant that answers product questions by retrieving pages scraped off the web. The demo works. Users like it. And as of yesterday, that product is sitting on a RAG copyright risk that just got much sharper.
On Tuesday, October 13, a federal judge in the Northern District of Illinois denied OpenAI’s motion to dismiss the copyright lawsuit Britannica filed last year. Britannica sued in September 2025, claiming OpenAI trained ChatGPT on its encyclopedia content and now hands that content back to users. The company puts potential damages at up to $500 million. With the motion gone, the case moves into discovery.
You could read this as an OpenAI problem. Don’t.
The claim that should matter to RAG builders is the one about outputs. Training on copyrighted text is unsettled law, and courts are still working out how fair use applies to model weights. Retrieval is different. When your system fetches an encyclopedia passage and returns it to a user, you copied the work and you distributed it. Courts have understood that kind of copying for two centuries. Discovery in the Britannica case will now dig into how content was collected, stored, indexed, and served. Those are the same four verbs that describe your pipeline, and that overlap is the heart of RAG copyright risk.
Most small teams have treated these lawsuits as background noise. The early ones were about training runs, and training runs belong to companies with nine-figure compute budgets. But Britannica’s complaint leans hard on verbatim output, and output is exactly what a retrieval augmented generation system produces every time it quotes a source.
Buried in the ruling is something useful. Nothing here makes RAG illegal or even unusual; it makes provenance a feature you can’t skip. Below: what the court decided, why retrieval changes the copyright math, and five fixes that fit a ten-person team’s budget.
What the Court Actually Decided
OpenAI’s dismissal motion followed the playbook from the New York Times case. The model learns statistical patterns, it doesn’t store copies, and outputs that track the training data are rare accidents rather than infringement. The judge let Britannica’s core claims through anyway.
A denial like this isn’t a verdict. It assumes Britannica’s facts are true and asks only whether those facts could amount to a legal claim. The practical effect is bigger than the legal one. OpenAI now has to open its ingestion records: what was collected, from where, and when. Every internal document about dataset assembly is fair game.
Two details in the complaint matter for anyone sizing up their own RAG copyright risk.
First, Britannica claims OpenAI stripped copyright management information from its articles. If that holds, DMCA claims come into play, and statutory damages for willful violations run up to $150,000 per work, with no need to prove actual harm. Second, the complaint catalogs examples of ChatGPT producing passages that track Britannica text closely. That’s the retrieval-shaped allegation. It frames the product as a distribution channel for the encyclopedia’s expression rather than a student of it.
One more piece of context. Britannica is not anti-AI. CEO Jorge Cauz laid out the position in a Time essay last year: the company wants to license its content to AI developers, and litigation is what happens when developers take it without asking. Publishers increasingly run both plays at once.
Why RAG Copyright Risk Is Different From Training Risk
Training makes one kind of copy. A RAG pipeline makes three.
At ingestion, you store the document. At indexing, you chunk it and keep those chunks in a vector store, which is another copy of protected expression. At generation, you send text drawn from those chunks to an end user. Each step is reproduction or distribution, the two activities copyright law has regulated since long before GPUs existed. If your corpus holds unlicensed work, your pipeline touches it repeatedly and keeps the receipts.
Recent rulings show where courts are drawing lines. In February 2025, in Thomson Reuters v. Ross Intelligence, Judge Stephanos Bibas ruled on summary judgment that using Westlaw headnotes to build a competing legal research tool was not fair use. That system never generated prose. It just used copyrighted text to build a product. That September, Anthropic settled with authors for $1.5 billion over pirated books used in training. The New York Times case against OpenAI and Microsoft grinds on.
The pattern is consistent. Courts forgive studying more readily than serving. Studying text to learn patterns is arguable. Packaging that text into a product response is not, and RAG systems do the second thing by design. That’s the core of RAG copyright risk.
Market harm is the factor that usually decides fair use questions. And Britannica sells exactly what your pipeline consumes: licensed knowledge for AI systems. “We only retrieved it” is not much of a defense when the plaintiff sells the license you skipped.
5 Fixes Before the Rules Get Written
The law will take years to settle, and an appeal will probably follow any verdict. Your next release can’t wait. Here are five changes that cut your RAG copyright risk, roughly in order of effort.
1. Log where every document came from
Store four fields with every chunk you ingest: source URL, acquisition date, the license you acquired it under, and the robots.txt status at crawl time. In any dispute, the first question is where the text came from. Teams that can answer in one query resolve problems cheaply. Teams that can’t either pull features or write checks. The log also doubles as a debugging tool, since tracing sources exposes duplicates and stale pages you’d want gone anyway.
2. Treat public as unlicensed until proven otherwise
A page loading in a browser grants you nothing. Copyright attaches automatically, and plenty of sites state in their terms of service that AI training and retrieval are prohibited. Others sell access. Britannica licenses its corpus to AI developers. Many news outlets run API programs. If a source sells a feed or dataset, that price is the license, and it’s usually cheaper than the alternative.
3. Cap what retrieval can return
The Britannica complaint is about long verbatim passages. So put the limits in the retrieval layer itself, not only in the prompt. Cap chunk size and cap the quoted text an answer may contain. Prefer paraphrase plus citation over dump. Some teams now run an overlap check on generated answers and flag any response whose similarity to a source chunk crosses a threshold. A response that quotes two sentences and links the page is a different product from one that returns three paragraphs, legally and in every other way. Of the five fixes here, this one cuts RAG copyright risk most directly.
4. Read your vendor contracts
OpenAI’s Copyright Shield covers outputs of its own models. It doesn’t cover the corpus you built and fed into the pipeline, and most model providers draw the same line. The contracts to actually read are the ones for your dataset or corpus vendors. When someone sells you a scraped web dump with no rights chain, that liability lands on you, not them. Get the provenance question answered in writing during procurement. The answer becomes part of your defense later. Vendor dumps are an underrated source of RAG copyright risk, and the contract is the only place to catch it.
5. Keep a corpus you can defend in an afternoon
Once a quarter, sample a hundred documents at random and trace each to its origin. Build an allowlist around sources with clean licenses. US federal government works are public domain. Wikipedia is CC BY-SA, which permits reuse if you honor attribution and share-alike terms. Licensed news APIs and publisher partnerships fill the rest. Cut everything you can’t trace. Document why the remainder is safe. A team of ten can run this audit in a weekend.
The Realistic Risk for a Ten-Person Team
Nobody is serving a small startup with a $500 million claim. The realistic RAG copyright risk is smaller and closer: a cease-and-desist from a publisher, a DMCA takedown against a hosted index, or an enterprise deal dying in legal review because the security questionnaire asked where your retrieval data comes from and the honest answer was “the internet.”
That last one is already routine. Procurement teams that once asked about uptime and SOC 2 now ask about data rights, and a RAG product with a documented, licensed corpus answers those questions in an afternoon. One without them stalls for a quarter or more, which makes procurement review the place where RAG copyright risk actually bites.
Time is on your side, at least a little. Yesterday’s ruling starts discovery, not a verdict. The case will take years, and an appeal will likely follow. Use the runway. Teams that spend the next year cleaning up ingestion will be fine. Teams that wait for a bright-line rule may find it arrives in a demand letter.
Where This Leaves You
The Britannica case reads like a training data story. The part that will shape RAG is about serving. A retrieval pipeline copies content three times, and the copies courts forgive are the ones nobody can see.
Five things to take from it: log provenance at ingestion, assume public means unlicensed, cap what your system returns, read your vendor contracts, and audit the corpus quarterly. None of it needs a legal budget. Most of it needs a weekend and a spreadsheet. Do those five and your RAG copyright risk shrinks to something you can explain in one email.
The four-person team from the opening has the easier job here. Their corpus is small enough to audit with a script. A weekend spent logging sources and capping quote length buys them something OpenAI can’t buy retroactively: a pipeline they can defend.
If you’re rebuilding ingestion around this, two earlier guides pair well with it. RAG Data Quality: 5 Defenses Against Machine-Made Junk covers keeping junk out of your index in the first place, and RAG Access Control: 5 Fixes to Stop Your Chatbot Leaking Private Docs covers the other way private content escapes into answers. For the next turn in Britannica v. OpenAI as discovery unfolds, the newsletter at ragaboutit.com is where this kind of analysis lands first. Sign up, and you’ll catch it while it’s still news.
