Two newspapers sued OpenAI and Microsoft on Friday. That part got covered everywhere. What did not get covered is that the complaint contains a second theory, sitting underneath the training-data argument everyone reported, and that second theory is the one that reaches the rest of us.
The Seattle Times Company and Newsday LLC filed a 38 page complaint in the Southern District of New York on September 4, 2026, case number 1:26-cv-07644. Nine OpenAI entities are named, plus Microsoft Corporation. Their lawyers are Klaris Law PLLC. They want a jury.
Nine separate OpenAI entities are named as defendants alongside Microsoft. The corporate-structure sprawl is itself a tell about how hard discovery is going to be.
I read the filing because I build retrieval systems for a living, and because I wanted to know whether this was another training-data case I could file away as somebody else's problem. It is not. Paragraph 64 makes a claim about retrieval-augmented generation that has nothing to do with training, and if it survives a motion to dismiss, it lands on every team running a crawler and a vector index.
What exactly did the Seattle Times and Newsday file?
Seven counts. Direct copyright infringement under 17 U.S.C. 501, vicarious copyright infringement, two DMCA counts under 1202(b)(1) and 1202(b)(3) one for removing copyright management information and one for distributing works knowing it had been removed, and three separate trademark dilution counts: federal under 15 U.S.C. 1125(c), Washington state under RCW 19.77.160, and New York under General Business Law 360-L.
The factual core is that OpenAI and Microsoft scraped hundreds of thousands of articles from both papers, went around their paywalls, and ignored their terms of service, to build products that now compete with them. The opening paragraphs do not hedge. "AI products like ChatGPT and CoPilot are touted as producers of content, but in fact they are rapacious consumers, devouring human-authored content and delivering back to the world copies and derivative imitations of that same original content they consumed to achieve their commercial objectives."
Then the line every outlet quoted: "Like a snake eating its own tail, GenAI that is trained on painstakingly researched, expensive-to-produce content threatens to destroy the very news organizations by competing directly with them through AI-generated substitutive content."
Some context on who is suing. The Seattle Times has been publishing since 1886, Newsday since 1940, and the complaint counts 30 Pulitzer Prizes between them. Newsday's site draws roughly 51 million monthly page views and about 2.1 million monthly unique visitors, with around 80% of that traffic on mobile. These are not opportunistic plaintiffs. They are two regional papers that have already survived every previous thing that was supposed to kill regional papers.
Why does the RAG claim matter more than the training claim?
Because training-time copying is a frontier lab problem and retrieval-time copying is everyone's problem. The complaint draws that line explicitly. Paragraph 64 describes RAG as "a technique that operates independently of the training process," and then spells out why that independence is legally interesting: the copying happens after the model is finished, often in real time, and the output "can extensively reproduce or closely track the language of a specific work even when that work was never part of the dataset used to train the underlying model in the first place."
Read that twice if you ship RAG. The claim is that you can infringe a document your model never saw during training, because your retriever fetched it at query time, dropped it into the context window, and the model paraphrased it back to the user. Paragraph 79 finishes the thought: RAG output "merely repackages the original reporting into a competing form."
Paragraph 87 splits the ongoing conduct into two mechanisms. First, broad web crawls that build fresh indices for retrieval, which store the papers' content. Second, direct live scrapes of seattletimes.com and newsday.com in response to user queries about current events, "without preserving CMI."
| Question | Training-time copying | Retrieval-time copying |
|---|---|---|
| When does it happen | Once, before release | Every query, forever |
| Who is exposed | Whoever trained the model | Whoever operates the retriever |
| Can you fix it after the fact | Not without retraining | Yes, by changing the pipeline |
| Is the source document in the weights | Yes, per the complaint | Not necessarily |
| Does an opt-out signal help | Only before the training run | Yes, at fetch time |
That last row is why I think this theory has legs. A court that is nervous about ordering a model destroyed has an easier remedy available on the retrieval side: enjoin the fetching. It costs nobody a training run. If you were a judge looking for a proportionate order, that is the one sitting right there.
I wrote a longer piece on how these pipelines actually get built in the agentic RAG production guide, and a plain-language version in what RAG is and what it is for. Neither of them treats the retrieval step as a copyright surface. That was an omission, and this filing is what changed my mind about it.
What does the complaint actually show ChatGPT reproducing?
Eighty-eight consecutive words, verbatim, from The Seattle Times' Pulitzer-winning series on the 2019 Boeing 737 MAX crashes. Paragraph 66 says the prompt carried the headline and the URL. No pasted text. Paragraph 65 makes the broader claim, that the models reproduced substantial passages in many instances when supplied with nothing more than a headline, publication date, and URL.
Here is the detail that I have not seen a single outlet mention, and it is the sharpest thing in the filing. Paragraph 60 describes Newsday's own editorial ethics policy, which treats "unattributed use of as few as seven to ten consecutive words from an outside source as plagiarism." Seven words gets a Newsday reporter fired. The complaint alleges the model produced eighty-eight.
That is a framing argument rather than a legal one, and it is a good one. It puts a number on both sides of the same standard.
Paragraph 67 adds side-by-side tables of Newsday text against model output. FEMA funding for Puerto Rico after Hurricane Maria. Long Island employment figures with the Labor Department's seasonal-adjustment caveat carried over almost word for word. A Kate Spade profile carried over almost intact. The overlap in those examples is shorter than 88 words, but it is structural, and the seasonal-adjustment sentence shows up twice from two different articles, which is the kind of thing that is hard to explain as coincidence.
The subscribe link in the corner is the whole business model. The complaint's theory is that a good enough answer elsewhere means that link never gets clicked.
Does robots.txt help, and did OpenAI honor it?
Newsday's robots.txt blocks both OpenAI and Common Crawl, per paragraph 59. Its terms of service, effective October 12, 2023, ban using its content "for the development of any software program, including, but not limited to, training a machine learning or artificial intelligence (AI) system." The Seattle Times' terms separately require compliance with exclusionary protocols and name robots.txt and ACAP directly. Both papers did the technical thing you are supposed to do.
The complaint's answer on timing is paragraph 78, and it is the strongest paragraph in the whole crawler section: OpenAI "did not disclose any bots used to perform earlier scrapes, and did not even begin to attempt to support REP disallows until August 2023." GPT-1, GPT-2, and GPT-3 were already trained by then. You cannot opt out of a crawl that already happened and was never announced.
OpenAI's own docs split the bots by purpose: GPTBot for training, OAI-SearchBot for search, ChatGPT-User for live fetches. The complaint uses that split against them.
OpenAI publishes this split itself. GPTBot gathers content that "may be used in training our generative AI foundation models." OAI-SearchBot surfaces sites in ChatGPT search. ChatGPT-User fetches pages live when a user asks something that needs current information. The docs say each setting is independent of the others, which is a reasonable thing to build. Read the same page closely and the split is narrower than it looks. There are four agents listed on it, not three, because OAI-AdsBot checks landing pages for ChatGPT ads. Only two of them have robots.txt tags at all, GPTBot and OAI-SearchBot. And on the live fetch, the one the complaint cares about most, OpenAI writes that because the action is initiated by a user, robots exclusion rules may not apply. The complaint reframes it as an admission: OpenAI itself distinguishes training access from retrieval access, so it cannot argue that a training-data license covers what the retriever does.
If you want to see which of these bots your own site currently lets through, I built a crawler check that reads your robots.txt and tells you. Most sites I run it against are inconsistent, usually because someone blocked GPTBot years ago and never revisited the file when OAI-SearchBot appeared.
What is the DMCA count, and why should smaller builders care?
Counts III and IV allege violations of 17 U.S.C. 1202(b), which covers removing copyright management information and distributing work knowing it is gone. CMI here means the article title, the author byline, the copyright notice at the foot of every page, the terms of service language. The claim is that the defendants stripped it while copying, and that outputs reach users with the attribution gone.
This is the count that should worry a two-person team more than the copyright count does, and the reason is mechanical. Section 1202 carries its own statutory damages per violation, independent of whether the underlying copying turns out to be fair use. And stripping CMI is the default behavior of almost every ingestion pipeline I have ever reviewed. You fetch a page, you run it through a readability extractor, you chunk it, you embed it. The byline and the copyright line are boilerplate, so the extractor throws them away, because throwing away boilerplate is what it was written to do.
Nobody decides to remove the attribution. The library removes it, quietly, because that is its job. Then the model answers a question from a chunk that no longer knows who wrote it.
The fix is not hard. Carry source metadata on the chunk, not just in a sidecar table, and put it in the prompt so the model can cite it. I made the same argument from the other direction when the music publishers sued Anthropic and the interesting part turned out to be the data pipeline rather than the model. Two cases, two plaintiff groups, same underlying finding: the liability lives in the plumbing.
What are the plaintiffs really asking for?
The prayer for relief asks for statutory damages or actual damages plus profits, separate statutory damages for each DMCA violation, a permanent injunction, and treble damages plus attorneys' fees on the dilution counts given what it calls bad faith and willful conduct.
Then subsection (d), which is the one the headlines picked up. It asks the court to order "the impoundment and/or destruction, pursuant to 17 U.S.C. Section 503, of all copies of Plaintiffs' works, and all LLMs and training datasets incorporating Plaintiffs' works or derivatives thereof."
Model deletion is a real remedy under 503 and it is also, in practice, a negotiating position. No court has ordered a frontier model destroyed. What that clause does is set a ceiling high enough that a licensing conversation looks cheap by comparison, and the complaint helpfully establishes what that conversation costs. Paragraph 89 lists OpenAI's existing deals with the Associated Press, News Corp, Axios, Axel Springer, The Atlantic, the Financial Times, Dotdash Meredith, and Vox Media, and notes that the disclosed terms of just three of them total more than $300 million.
The argument the plaintiffs build from that is neat. A market for this license demonstrably exists, OpenAI participates in it, so OpenAI already concedes a license is required. It just never sought one from these two papers.
Does the traffic damage argument hold up?
The complaint's number is that search referral traffic to mid-sized regional and metro daily publishers fell roughly 47% between December 2024 and December 2025, against roughly 22% for larger national publishers, citing Search Engine Land data. The stated reason is that mid-sized outlets depended more heavily on incidental search traffic, which AI-generated answers now absorb.
I can corroborate the shape of this from my own analytics, at a much smaller scale. Around 65% of the Google impressions this site earns are for questions phrased the way people talk to an assistant, and those impressions convert to clicks at effectively zero. The content is being read. The visit never happens. I wrote about what that does to a publishing strategy when Google shipped the preferred sources button, and the honest summary is that being cited and being visited have come apart as measurements.
Where I would push back on the complaint is causation. Referral traffic fell during a period when Google also shipped AI Overviews, changed its core ranking several times, and expanded zero-click features that predate generative AI entirely. Attributing a 47% decline to two defendants is going to require expert work that the filing does not attempt, which is normal at the complaint stage but is exactly where a motion to dismiss will aim.
Paragraph 91 makes a subscription argument that I find more durable, because it needs less causal machinery: a reader who gets a satisfying answer from ChatGPT has less reason to ever subscribe. That harm does not require you to identify which specific article was copied. It just requires the substitution to work.
Why is a paper suing its own funder?
Because the money and the injury are unrelated, and the filing is what happens when a newsroom decides that stops being a reason to hold off. Microsoft Philanthropies underwrites some Seattle Times journalism projects. In 2024, Microsoft and OpenAI jointly funded a $10 million Lenfest Institute AI fellowship whose inaugural newsrooms included both the Seattle Times and Newsday.
GeekWire, based in Seattle, had the story within hours of the filing and surfaced the funding relationship the national coverage mostly left out.
Seattle Times Co. president and CEO Alan Fisco told staff in a memo that "This was not an easy decision," and said the company felt strongly it had to defend content that costs millions of dollars a year to produce from being used without consent or compensation. The paper says it keeps editorial independence from its funders.
Microsoft's response was mild. A spokesperson said the company was surprised by the lawsuit, said it appreciates the importance of the Seattle Times to the region, and offered to sit down and talk about this kind of dispute. Nobody has said whether licensing talks happened before the filing.
This is not the first of these, either. The complaint's own footnotes cite The New York Times against Microsoft and OpenAI from December 2023, the Center for Investigative Reporting from June 2024, Ziff Davis from May 2025, and U.S. News & World Report from November 2025, and it cites them for a specific purpose: to argue willfulness. The defendants have been on notice since 2023 and kept going.
What should you change if you ship retrieval this quarter?
Four things, and none of them require a lawyer to start.
Keep the attribution on the chunk. Whatever your extractor throws away as boilerplate probably includes the byline and the copyright line, and those are the exact fields Section 1202 protects. Store them alongside the embedding and put them in the prompt.
Log what your retriever fetched, not just what your model answered. If somebody asks you in eighteen months which of their pages you read and when, an answer of "we do not keep that" is a much worse answer than a log file. This is the same discipline as zero data retention on the monitoring side, applied at the other end of the pipe.
Honor robots.txt at fetch time, per request, not as a one-off check you ran when you built the crawler. Sites change their files. Newsday tightened from a metered paywall in January 2019 to a hard gate in August 2022 and updated its terms in October 2023, and a crawler that cached a permission decision from 2021 would have missed all three.
Decide now whether you are licensing or gambling. If your product's value depends on a specific publisher's content, the market rate is knowable, because the complaint just published a floor for it. There is a version of this where you find out what a license costs and it is less than you feared. I made a similar argument about consent when Twitch changed its AI training opt-out, and the pattern holds: the teams that ask early get better terms than the teams that get asked late.
None of this is legal advice and I am not a lawyer. It is what I would want in place before a letter arrives, based on which parts of this complaint would be hardest to answer.
Where I think this lands
The training-data counts will get litigated the way the New York Times case has been litigated, slowly, with fair use doing most of the work, and probably settled. The trademark dilution counts strike me as the weakest part of the filing, because dilution requires the mark itself to be tarnished and "ChatGPT summarized our article" is a hard fit for a doctrine built around famous brands on unrelated products.
The RAG counts and the CMI counts are the ones I would watch. They do not require a court to decide anything philosophical about whether training is reading. They require a court to decide whether fetching a paywalled page in real time, stripping its byline, and paraphrasing it back to a user is copying. That is a much more ordinary question, and ordinary questions get answered faster.
If the answer is yes, the compliance work does not stay at the frontier labs. It arrives at every team with a crawler, which by now is most of them.
If you are trying to work out how much of your own stack this touches, the AI readiness assessment walks through where your data actually comes from and what you would be able to prove about it. Same questions I ask at the start of an engagement.
Frequently asked questions
Who filed the lawsuit against OpenAI and Microsoft?
The Seattle Times Company and Newsday LLC filed jointly in the Southern District of New York on September 4, 2026, as case 1:26-cv-07644. They are represented by Klaris Law PLLC and have demanded a jury trial. Nine OpenAI corporate entities are named alongside Microsoft Corporation.
What is RAG copyright infringement?
It is the claim that copying happens at retrieval time rather than during training. A system fetches a copyrighted page in response to a user query, puts the text into the model's context, and returns an answer that reproduces or closely tracks the original. The complaint argues this can infringe even when the work was never in the training data, because the copying is performed by the retriever.
How much of an article did ChatGPT reproduce?
The complaint alleges 88 consecutive words verbatim from The Seattle Times' Pulitzer-winning coverage of the Boeing 737 MAX crashes, produced from a prompt that paragraph 66 says carried only the headline and the URL. For scale, Newsday's own ethics policy treats seven to ten unattributed consecutive words as plagiarism.
Does blocking GPTBot in robots.txt protect a publisher?
Only going forward, and only for the bot you named. Only partly, and less than most publishers assume. Per the complaint, OpenAI did not begin supporting robots exclusion until August 2023, after GPT-1, GPT-2, and GPT-3 were trained. Going forward, OpenAI's docs offer robots.txt tags for exactly two agents, GPTBot and OAI-SearchBot, so blocking GPTBot leaves search crawling untouched. The live fetch is worse: OpenAI states that because a ChatGPT-User request is initiated by a person, robots.txt rules may not apply to it, which means there is no robots.txt line that reliably stops it.
What does the DMCA claim add to a copyright claim?
Section 1202(b) covers removal of copyright management information such as bylines and copyright notices, and carries its own statutory damages per violation. It runs independently of the infringement claim, so it can matter even where fair use is arguable. Standard content extraction pipelines strip this metadata by default.
Can a court really order an AI model destroyed?
Impoundment and destruction of infringing copies is an available remedy under 17 U.S.C. 503, and the plaintiffs request it explicitly for the models and training datasets. No court has yet ordered a frontier model destroyed, and in practice the request functions as pressure toward a licensing settlement.
Sources: Complaint, The Seattle Times Company and Newsday LLC v. OpenAI et al., No. 1:26-cv-07644 (S.D.N.Y. filed Sept. 4, 2026), full filing via CourtListener · GeekWire (Sept. 4, 2026) · TechCrunch (Sept. 5, 2026) · The Verge (Sept. 6, 2026) · OpenAI, Overview of OpenAI Crawlers.
Top comments (0)