TORONTO — Cohere is asking a U.S. court to toss out a copyright lawsuit brought against it by major media companies, claiming the publishers manipulated its AI tools to generate misleading examples.

Toronto Star Newspapers, Vox Media and Condé Nast are among the media companies that sued the Toronto-based firm in the Southern District of New York in February. Their filing alleged that Cohere was engaging in “massive, systematic copyright infringement and trademark infringement.” The filing cited examples of a tool on Cohere’s website regurgitating word for word or near-verbatim sections of news stories when prompted to provide the articles in question.

In a motion to dismiss filed Thursday, Cohere claimed the tool the publishers used was a demo to show developers how its technology works, not a product it sells to customers. The publishers “deliberately misused” the system to “try to manufacture a case,” the filing said. The company also claimed that prompting its tool to generate text that infringes on intellectual property violates its terms of service. 

Firms like Cohere use large amounts of data scraped from the internet to train their large language models (LLMs). Those systems are then used to power tools like chatbots or autonomous agents that can do tasks on their own. Several news organizations and music publishers are suing AI developers like OpenAI and Anthropic, alleging that the tech firms violated their copyright by training models on articles or lyrics.

Cohere says it shouldn’t be grouped with those firms in terms of copyright litigation. The Canadian company sells its tools to businesses in highly regulated industries, but does not offer a ChatGPT-style bot or other consumer products. While other firms’ AI tools are used for everyday tasks like homework assistance or planning trips, Cohere’s clients use its tools in the workplace to search through their data or translate documents, the filing claimed. 

The publishers have not shown that “any real customer of Cohere ever has—or even would—use Cohere’s solutions to infringe [their] copyrights,” the motion filed by Cohere’s lawyers states.

Some of the media companies that brought the suit have licensed their content to other LLM makers. Those deals are focused on providing material for consumer-facing products, Cohere co-founder Nick Frosst told The Logic earlier this year. He said the firm isn’t pursuing similar agreements because its business clients don’t need access to that kind of content.

In its filing, Cohere said that when it trains LLMs on text, it’s not trying to copy that specific material, but rather to teach the system patterns in language and information. “Cohere does not sell, and customers do not purchase, particular outputs,” it said. Cohere invoked the fair use provisions of copyright law, which it said are currently being tested in generative AI lawsuits in the U.S.