The Q3 2026 AI copyright map: Sony and Warner, Seattle Times, the DOJ brief, and who gets paid from $1.5B

Between August 31 and September 7 the training-data copyright fight produced four events that point in different directions, and anyone running RAG or fine-tuning on third-party content should understand all four before reading any one of them as the trend. Here is the map.

The four events

August 31: Sony Music Publishing and Warner Chappell sue Anthropic. Filed in California federal court, the complaint alleges unauthorised use of copyrighted lyrics and sheet music to train Claude: scraped from websites, taken from datasets containing protected works, and scanned from physical copies. Three acquisition channels, named separately.

September 3: the DOJ files a statement of interest for OpenAI and Microsoft. In the dispute with the New York Times and a group of authors, the Department of Justice argued that training on copyrighted text is transformative fair use, because the model learns statistical patterns rather than copying works, and framed American AI leadership as a national-security interest. It is the first time the federal government has formally entered the training-data lawsuits. The brief is advisory; it binds no court.

September 5: the Seattle Times and Newsday sue OpenAI and Microsoft. Same theory as the Times case, applied to two more newsrooms. The detail everyone noticed: Microsoft and OpenAI are among the Seattle Times' sponsors.

September 4 onward: Bartz v. Anthropic pays out. The $1.5 billion settlement, roughly $3,000 per book across about 500,000 titles, sent claim notices on September 4. Authors then discovered that publishers and literary agencies were claiming shares. The split that emerged: in-print books with a traditional publisher divide 50/50 between author and publisher; self-published titles or books with reverted rights pay the author in full; literary agents are not rightsholders and get nothing directly.

Reading them together

Two things are true at once. The executive branch has, for the first time, put its weight behind the fair-use argument for training. And the number of plaintiffs is growing, across music, news and books, with a settlement on the record that puts a per-work price on the exposure. Fair use may win in court; it has not yet, and $1.5 billion is what settling looked like for one defendant.

The DOJ brief tells you which way the government leans; the settlement tells you what a wrong guess costs.

What it means for your stack

The lawsuits are about pretraining foundation models. Your risk profile is different, and mostly smaller, if you are precise about what you do with third-party content.

• RAG over licensed or owned content: lowest risk. Retrieval at query time and citation of a source you have rights to is not the fact pattern in any of these cases. Keep the licence record next to the index, and keep it current as the pipeline changes.
• RAG over scraped web content: the exposure is in the scraping and storage, not the model. The Sony complaint names scraping as a channel. Know what is in your corpus and where it came from.
• Fine-tuning on third-party data: the closest to the litigated fact pattern. Prefer data you own or license, document provenance at the row level, and keep the choice between fine-tuning, RAG and prompting partly a legal one.
• Vendor models: ask your provider what indemnity they offer for outputs and for training data, in writing. The settlement shows the vendor can be the one paying, but only if the contract says so.
• Outputs that reproduce protected works: a separate risk from training, and the one you can actually test for. Add reproduction checks to your output validation.

The DOJ brief is the one to track: if a court adopts its reasoning, the pretraining exposure drops for everyone downstream. The Sony complaint is the one to read: its three acquisition channels are a checklist for how a plaintiff will describe your corpus.

The honest gap

Nothing here is a ruling. The DOJ brief is advisory, the Sony and Seattle Times complaints are allegations, and the settlement is a negotiated outcome that sets no precedent. I am an engineer, not a lawyer, and the right reading of these events for your jurisdiction and your contracts needs one. What an engineer can do is make provenance a first-class field in every dataset, so that when the lawyer asks where the data came from, the answer is a query and not a guess.