Advance Local Media v. Cohere is a copyright and trademark action in the Southern District of New York by fourteen news and magazine publishers against Cohere Inc., the developer of the Command family of language models. The publishers allege that Cohere copied their articles to train its models and that the models, when retrieving content, reproduce full copies, substantial excerpts or substitutive summaries of articles, including articles behind paywalls.[1][2] The complaint was filed on 13 February 2025, and the court denied Cohere’s partial motion to dismiss on 13 November 2025.[6][2] The case is pending in the cited records.
Background
The court described the case as one of more than fifty lawsuits then before the courts challenging the use of copyrighted works by artificial intelligence companies to train large language models.[2] It recorded that Cohere is a Canadian company that develops, operates and licenses AI models, and that its primary product is the Command family, which customers can use through a Chat interface priced on the length of prompts and outputs, or run as their own instances under negotiated agreements.[2] Those two commercial routes are the same ones described in Cohere licensing: pay-as-you-go use and dedicated or private deployments.
According to the complaint as summarised by the court, Cohere collects text from the internet with crawlers, uses a Common Crawl dataset, and crawls particular websites including the publishers’. Command uses retrieval-augmented generation (RAG) to reach external sources when answering, and the publishers allege that with RAG turned on it returns verbatim copies, excerpts or substitutive summaries, and that with RAG turned off it sometimes invents articles that carry the publishers’ trademarks.[2]
The dispute
The complaint pleads four counts: direct copyright infringement, secondary copyright infringement, trademark infringement under 15 U.S.C. 1114(1), and false designation of origin under 15 U.S.C. 1125(a)(1)(A).[2] The publishers ask for a finding of willfulness, a permanent injunction, an order precluding use of their works to train or fine-tune models, an order requiring Cohere to filter its RAG connector, destruction of infringing copies, and statutory damages of up to USD 150,000 per work or actual damages and profits.[1]
Cohere moved to dismiss the secondary infringement and Lanham Act counts, and the direct infringement count only to the extent it rested on a theory of substitutive summaries.[2] It argued, among other points, that the publishers’ 75 examples came from investigators and did not show typical use, that investigators had breached Cohere’s terms of service to obtain them, that plaintiffs had not alleged actual knowledge of specific infringement, and that its models are bounded by private business deployments.[2]
Decision or outcome
On 13 November 2025 Judge Colleen McMahon denied the motion in full.[2] On the substitutive-summaries theory, she held that the publishers had plausibly alleged outputs similar in quantity and quality, pointing to examples said to copy paragraphs nearly verbatim, and noted that the Second Circuit authority Cohere relied on expressly declined to set a rule that copying about twenty percent can never be infringement.[2]
On secondary liability, the court found that the investigators’ examples were enough at the pleading stage to raise an expectation that discovery would show third-party infringement, and that a possible breach of Cohere’s terms of service by an investigator did not defeat the claim. It applied the Second Circuit’s objective “know or have reason to know” standard rather than an actual-knowledge standard, and held that advertising Command as a tool for the latest news, together with a demo that pre-populated a request to summarise technology news, plausibly pleaded inducement.[2] On the Lanham Act counts, the court held that the publishers had adequately alleged use in commerce through the free trial and the display of marks on invented articles.[2]
Later developments
Discovery followed. By a joint status letter of 27 March 2026 the parties reported document productions, including Cohere’s productions of training data and log data.[5] On 2 July 2026 Magistrate Judge Ona Wang considered the parties’ dispute over Cohere’s proposed sampling protocol for log records and noted Cohere’s proposal of 45,000 records from the Chat endpoint alone.[4] By a stipulation entered on 9 July 2026, and citing the Supreme Court’s 2026 decision in Cox Communications v. Sony Music Entertainment, the publishers withdrew with prejudice their theory of contributory infringement based on knowledge and material contribution, while the contributory infringement claim based on inducement remains.[3] The cited records do not show a trial date or a decision on the merits.
Significance for software licensing and SAM practice
The case is about copyright in the content used to train and feed a model, not about the licensing of Cohere’s software, but it bears on how customers should read Cohere’s terms. Cohere’s SaaS Agreement contains a Copyright Assurance under which Cohere defends and indemnifies customers against third-party claims that an output infringes copyright, subject to exclusions such as unpaid fees, breach of the agreement, intentional generation of infringing outputs, finetuning by the customer and continued use after notice.[7] The suit alleges infringement in the outputs of the same product family, so a customer deploying Command in a news or content-summarisation workflow should read those exclusions closely and keep a record of fees paid and prompts used.
The court’s discussion of logs also links to data terms. Cohere states that on its SaaS platform logged prompts and generations are deleted after 30 days unless a legal requirement or customer contract requires otherwise, and that zero data retention is available to approved enterprise customers.[8] Whether litigation preservation affects a particular customer’s logs is not addressed in the cited records, and a customer relying on a retention commitment should confirm it in its Order Form.
Lessons learned
- Retrieval is a licensing issue as well as training. The court allowed claims based on what the system retrieves and outputs, so a licence that covers only training or only the model weights does not address the risk of outputs that reproduce third-party content.[2]
- Marketing statements travel. The court treated advertising a model for a specific use, here accessing the latest news, as possible evidence of encouraging infringement at the pleading stage, so an organization’s own description of how it uses a model can matter.[2]
- Indemnities have exclusions. Cohere’s Copyright Assurance depends on payment of fees, compliance with the agreement and non-intentional use, so an unpaid invoice or an out-of-scope use can remove the protection.[7]
- Logs can become evidence. Sampling of customer prompt and output logs was a live discovery dispute, which makes retention periods and zero-data-retention terms part of legal risk, not only privacy.[4][8]
- Legal theories shift during a case. After a 2026 Supreme Court decision the plaintiffs dropped one contributory theory and kept another, so a customer should check the current status of the case before relying on any summary of it.[3]