Microsoft OpenAI Copyright Fight Expands as Seattle Times and Newsday Sue

Microsoft and OpenAI now face another publisher lawsuit, despite years of funding experiments intended to help local newsrooms adopt artificial intelligence. The Seattle Times and Newsday filed the case on September 4, 2026. They allege that the companies copied their journalism without permission to build and operate commercial AI products.

The complaint targets more than model training. It says ChatGPT, Microsoft Copilot, and Bing features can reproduce or closely paraphrase protected reporting. Those answers allegedly compete with the publishers for readers, advertising, subscriptions, and licensing revenue.

That makes this Microsoft OpenAI dispute more complicated than a familiar fight between technology companies and copyright owners. Both publishers participated in an AI fellowship funded by the defendants. They are not rejecting AI outright. They are arguing that collaboration does not authorize uncompensated copying or products that substitute for their reporting.

The case also arrives during a decisive phase of the broader copyright conflict. The New York Times and other publishers are already litigating related claims in Manhattan. Meanwhile, the U.S. Justice Department has backed a broad fair-use defense for AI training. The new plaintiffs must therefore separate the alleged copying used to teach models from outputs that can expose protected expression to users.

The Lawsuit Targets Training, Retrieval, and AI Answers

The publishers are attacking the entire content pipeline, from website scraping to the answers shown to users.

The Seattle Times Company and Newsday LLC filed their 38-page complaint in the U.S. District Court for the Southern District of New York. The named defendants include Microsoft and several entities within OpenAI’s corporate structure.

The publishers allege that automated systems copied hundreds of thousands of their articles. According to the complaint, some material sat behind paywalls that were designed to restrict access to subscribers. The defendants have not admitted these allegations, and the court has not determined that the copying occurred as described.

The complaint identifies several alleged uses for the material. It says copies entered datasets used to train and fine-tune large language models. It also alleges that content was stored for retrieval-augmented generation, commonly called RAG. That process retrieves external material when a user asks a question and supplies it to a model when producing an answer.

This distinction matters because training and retrieval present different factual questions. Training changes a model’s parameters by exposing it to patterns across large collections of data. Retrieval can bring a specific document into the generation process when a user submits a relevant query.

The publishers also accuse the companies of removing or changing copyright management information. That information can include titles, author names, and copyright notices. Their claims extend beyond direct copyright infringement to related federal and state-law theories, including trademark dilution.

The most tangible allegations concern model outputs. The complaint presents examples in which ChatGPT allegedly returned passages resembling the publishers’ work. One test reportedly produced an 88-word stretch from The Seattle Times’ Pulitzer-winning Boeing 737 MAX reporting after receiving an article headline and web address.

That example does not establish how often similar outputs occur. It does show why the case cannot be reduced to an abstract debate about machines learning statistical patterns. The publishers are placing specific output behavior beside their claims about upstream copying.

The requested remedies are correspondingly broad. The newspapers seek damages and an injunction against continued infringement. They also ask for the destruction of unauthorized copies, affected training datasets, and models that incorporate their work, according to the filed copyright complaint.

A destruction order would create substantial technical and legal questions. A modern model reflects training across a large dataset rather than maintaining a simple library of readable articles. Courts would need evidence connecting particular works, datasets, model versions, and allegedly infringing uses before considering such a remedy.

For now, every central assertion remains an allegation. The defendants have not filed their detailed answers, and no judge has ruled that the publishers’ works entered the products unlawfully. The filing changes the dispute by establishing a new case, not by resolving the contested facts.

The Microsoft OpenAI Defense Faces a Local-News Test

The central conflict is whether AI training creates a new analytical tool or a substitute built from the same journalism it competes against.

OpenAI says its models are trained on publicly available data and that the practice is grounded in fair use. Fair use is a legal doctrine that permits some unlicensed uses after considering purpose, transformation, market effects, and other factors.

Microsoft did not offer a detailed legal defense when the lawsuit became public. A company spokesperson said Microsoft was surprised by the filing and remained willing to discuss possible solutions. That response leaves the complaint’s technical allegations unanswered at this early stage.

The publishers frame the same activity differently. They say the defendants copied expressive works to create services that answer the same questions those works were written to answer. In their account, ChatGPT and Copilot do not merely analyze journalism. They can deliver information that removes a reader’s reason to visit the originating publication.

That alleged substitution is important because market harm is part of fair-use analysis. A court could treat model training, model outputs, and retrieval systems as separate uses. An activity considered transformative at one stage does not automatically protect every downstream output.

The Justice Department has argued for that separation in related litigation. In a September 2026 filing, the government described LLM training as highly transformative. It compared the process to using existing material to build a system that recognizes relationships and responds to new information.

However, the government’s fair-use brief also distinguished training from outputs that reconstruct and distribute protected works. It argued that potentially infringing outputs should receive a use-by-use assessment rather than determine the legality of training as a whole.

That position gives both sides something to work with. OpenAI and Microsoft can argue that training serves a different purpose from publishing news. The newspapers can focus on retrieval, memorization, repeated passages, and answers that allegedly replace article visits.

The burden will fall heavily on evidence. The plaintiffs need to connect registered works with particular acts of copying or actionable outputs. The defendants can challenge whether prompted reproductions represent normal product behavior, exceptional testing, or attempts designed to elicit protected text.

The new case also raises a local-news dimension. National publishers have diversified audiences and subscription businesses, although they face the same referral pressures. Regional organizations depend more heavily on readers seeking information about local government, transportation, schools, weather, and public safety.

Those queries are well suited to summarized answers. A user may ask an assistant about a transit closure and receive the essential details without opening the reported article. When that answer derives from original local reporting, the publisher bears the reporting cost while another interface controls the audience relationship.

The complaint cites industry data claiming that search referrals to midsized regional and metropolitan publishers fell 47 percent between December 2024 and December 2025. It compares that figure with a 22 percent decline for larger national publishers.

Those figures provide context, but they do not prove causation in this case. Referral traffic changes for many reasons, including search ranking updates, consumer behavior, platform design, and competition among publishers. The plaintiffs will need more specific evidence to connect their losses with the defendants’ products.

The legal question is therefore narrower than whether AI affects journalism. It is whether the defendants made legally protected uses, produced infringing outputs, or harmed markets copyright law recognizes. The distinction will determine whether this becomes a precedent for training or a case focused on particular content pipelines.

AI Funding Did Not Settle the Permission Question

The sharpest reversal is that both newspapers accepted AI-development support while preserving their claim that adoption is not consent.

In 2024, Microsoft and OpenAI supported an AI collaborative organized by the Lenfest Institute for Journalism. Each company committed direct funding and software credits. The combined program was valued at up to $10 million and included eight independent metropolitan news organizations.

The Seattle Times and Newsday were both participants. The initiative aimed to help local publishers test AI-supported products and develop more sustainable business models. It treated AI as something news organizations could use, not simply a threat they should block.

The Seattle Times planned to explore advertising, sales training, and analytics. Newsday planned public-data summarization tools for its newsroom, readers, and business customers. The published fellowship plans placed both organizations inside the industry’s experimentation with generative AI.

The Seattle Times has also said its newsroom permits limited uses after risk review. Those uses can include transcription and data analysis. The organization says it does not use AI to generate its published stories and labels significant uses when appropriate.

This history weakens any simplistic description of the lawsuit as resistance to new technology. The plaintiffs are drawing a boundary between using licensed tools and authorizing their journalism for training, retrieval, or competing outputs.

It also creates an uncomfortable question for Microsoft and OpenAI. Their local-news investments promoted AI as a route toward sustainability. The lawsuit argues that the same companies undermined that goal by using reporting without compensation.

Seattle Times President and CEO Alan Fisco told employees that filing the case was difficult. He said the organization spends millions of dollars annually producing its content and must defend it against unapproved, uncompensated use.

Fisco also said the suit was not intended to impede AI innovation. His stated concern was that innovation should not come at the expense of the newspaper’s business model. That position turns the conflict toward licensing and market design rather than a total ban on AI.

A labor response adds another layer. The Seattle Times Union, representing more than 160 employees, supported the copyright case. It also criticized the publisher for declining to guarantee that AI would not replace non-reporting newsroom positions during contract negotiations.

The union’s position exposes a second boundary. A publisher can defend its intellectual property while still facing scrutiny over how it deploys AI against its own workers. The lawsuit does not settle whether newsroom adoption protects employment, improves productivity, or shifts work away from staff.

For knowledge workers, the dispute illustrates why provenance matters. Provenance means knowing where information originated and what permissions govern its use. An AI answer can appear useful while concealing the reporting, documents, and human labor that made the answer possible.

Organizations building an AI knowledge base face a related operational question. They need rules for what enters the system, which users can access it, and whether generated answers preserve source context.

The conflict is not collaboration versus litigation. It is licensed collaboration versus alleged uncompensated extraction. A grant, product license, or innovation program does not necessarily convey copyright permission for unrelated uses.

That distinction will matter beyond journalism. Enterprises increasingly combine vendor models with internal documents, licensed databases, and public web material. Buyers need clear contractual answers about training rights, retrieval sources, attribution, data retention, and responsibility for generated outputs.

Licensing Deals Make the Market-Harm Argument Harder to Ignore

Existing publisher agreements show that news content has an AI licensing market, but they do not automatically prove every unlicensed use is illegal.

OpenAI has agreements with publishers that chose negotiation instead of litigation. Those arrangements help establish that archival and current news material has commercial value for AI developers.

The Associated Press announced an agreement allowing OpenAI to license part of its text archive. The companies also planned to explore potential uses of generative AI in news products. The archive agreement did not disclose complete commercial terms.

News Corp later announced a multi-year partnership covering current and archived content from publications across several markets. OpenAI received permission to display material within its products, while News Corp gained a formal commercial relationship around that use.

The publisher partnership demonstrates a workable permission path. It also strengthens the plaintiffs’ argument that unauthorized copying can bypass a market in which other publishers receive compensation.

Still, the existence of licenses does not decide fair use. Copyright defendants often argue that rights holders cannot convert every transformative use into a controlled market merely by offering licenses. Otherwise, the decision to sell permission could predetermine whether permission was legally required.

The newspapers must show more than a hypothetical lost fee. They must explain how the challenged uses affect established or reasonably expected markets for their works. Evidence about licensing negotiations, product substitution, referral losses, and subscriber behavior will become important.

OpenAI and Microsoft will likely emphasize the difference between a voluntary commercial arrangement and a legal obligation. They can argue that some companies license content to improve products, obtain reliable feeds, or display attributed material, even when training itself remains fair use.

The output allegations may again prove more consequential than the existence of training copies. A model that uses patterns learned from journalism presents one market question. A service that returns recognizable passages, summarizes a paywalled investigation, or falsely attributes invented material presents another.

False attribution also moves the dispute beyond lost clicks. The complaint says generated content has been associated with The Seattle Times or Newsday even when the publishers did not create it. That alleged behavior threatens brand trust, which local news organizations build through accuracy and accountability.

Microsoft’s role will receive particular attention. The company is not merely an outside investor. It has integrated OpenAI models into Copilot and Bing products, supplied infrastructure, and maintained a substantial economic interest in the partnership.

The complaint states that Microsoft committed approximately $13 billion to OpenAI beginning in 2019. It also cites the companies’ October 2025 announcement that Microsoft held an investment valued at roughly $135 billion, representing about 27 percent on a diluted basis.

Those figures come from the plaintiffs’ filing and cited corporate disclosures. They explain why Microsoft appears as a defendant, but investment alone does not establish liability. The publishers must connect Microsoft to the alleged copying, control, product operation, or infringing outputs under each legal theory.

The plaintiffs seek to portray Microsoft OpenAI products as a connected commercial system. Microsoft can contest that characterization by distinguishing corporate entities, technical responsibilities, and product-specific behavior. Discovery should clarify who assembled datasets, who controlled crawling, and which systems generated the cited responses.

That evidence will also shape any remedy. Destroying a stored article is simpler than removing its influence from a trained model. If multiple datasets, model generations, and product layers are involved, the plaintiffs must identify which assets contain actionable copies and why narrower controls would be inadequate.

The licensing landscape therefore cuts both ways. It supports the argument that journalistic content has measurable commercial value. It also highlights the need to analyze each use rather than treating all access to published material as legally identical.

Three Signals Will Show Whether the Case Changes AI Publishing

The next phase will turn on consolidation, technical evidence, and the court’s separation of training from user-facing outputs.

The first signal is whether this lawsuit joins the existing multidistrict proceedings involving The New York Times and other publishers. Related cases have been coordinated before U.S. District Judge Sidney Stein in Manhattan.

Consolidation can reduce duplicated discovery and inconsistent rulings. It can also place the new plaintiffs inside a mature dispute where both sides have already developed arguments about datasets, outputs, fair use, and market harm.

A coordinated proceeding would strengthen the view that this is part of a sector-wide legal contest. A separate track would preserve more room for facts specific to regional publishers, paywalls, local-news economics, and Microsoft’s relationship with its hometown newspaper.

The second signal is what discovery reveals about content provenance. The complaint alleges that Seattle Times material appeared among the prominent domains used for an earlier OpenAI dataset. It also asserts that later models incorporated both publishers’ work through web datasets, search indexes, and other sources.

OpenAI disclosed more information about early model datasets than it has about recent systems. That opacity makes provenance difficult to establish from public materials alone. Internal records could show whether the publishers’ articles entered specific datasets, how copies were processed, and whether later systems retrieved them directly.

Discovery could also test the significance of the 88-word example. The parties will examine whether ordinary prompts produce similar passages, whether safeguards prevent repetition, and whether the output came from memorized training data or retrieved content.

Frequent, easily produced reproductions would strengthen the publishers’ substitution argument. Rare outputs created through unusually targeted prompts would support the defense claim that exceptional behavior should not determine the legality of an entire technology.

The third signal is how the court separates input copying from public output. The Justice Department’s position favors strong fair-use protection for training while leaving room for claims involving reconstructed works. A ruling that adopts that division would narrow the publishers’ path without eliminating it.

Such a result would shift attention toward output filters, retrieval permissions, source attribution, and compensation for displayed content. It could encourage developers to maintain stronger boundaries between model training and systems that fetch or quote live documents.

A ruling that treats training and competitive outputs as one connected commercial process would create greater risk for AI companies. It would also strengthen publishers seeking licenses for both archival data and continuing access.

The destruction demand remains the highest-stakes remedy. Courts generally tailor injunctions to proven violations. Requiring the removal of complete models would demand evidence that narrower measures cannot provide adequate relief.

That makes an immediate shutdown unlikely at the opening stage. More realistic near-term outcomes include contested discovery, product-level safeguards, licensing discussions, or orders addressing identifiable copies and outputs. The ultimate remedy depends on facts that are not yet public.

Developers should watch whether the case produces clearer standards for source tracking. Enterprise buyers should ask vendors how protected content reaches generated answers. Publishers should separate negotiations over training, retrieval, quotation, attribution, and referral traffic instead of treating them as one permission.

Knowledge workers have a simpler test. When an AI answer replaces a source, can the user still identify and evaluate the reporting behind it? If not, convenience has removed information needed to judge accuracy, ownership, and context.

The Microsoft OpenAI lawsuit will matter most if it forces courts to draw enforceable boundaries across those stages. Until then, neither broad claims of theft nor broad claims of fair use answer every question.

Watch the court docket, the discovery record, and any licensing talks during the next three months. Together, those signals will show whether this case changes AI training rules or concentrates liability on specific outputs. Readers should also keep opening and evaluating original reporting when an AI system summarizes consequential local news.