The federal government has entered one of the most consequential copyright disputes involving generative AI, supporting OpenAI’s position that using copyrighted written works to train large language models generally constitutes fair use.
The intervention remains highly relevant. Since the Department of Justice filed its Statement of Interest on September 1 in the Southern District of New York, OpenAI, Microsoft, The New York Times and a group of authors have filed competing motions for summary judgment. The parties are now asking the court to decide whether the use of books and news articles to train generative AI systems is protected by fair use.
The eventual ruling could influence technology companies, publishers, authors and nearly every business developing, licensing or deploying generative AI.
The DOJ’s Statement
The DOJ described the training of large language models as “exceedingly transformative.” Its position is that an AI model uses written works to learn statistical relationships involving language and information, rather than to provide the public with copies of those works.
The government also argued that training, outputs and the methods used to obtain training materials must be evaluated separately. An AI-generated output that reconstructs protected material could raise different legal questions from the earlier use of that material during training.
According to the filing, restricting AI training based on an overly narrow interpretation of fair use could hinder scientific progress, competition, national security and American economic growth. The government also expressed concern that mandatory licensing could disproportionately benefit the largest technology companies and established publishers, which are better positioned to negotiate and pay for broad content libraries.
The DOJ’s September 1 filing applies its arguments across the consolidated litigation involving news organizations, authors and publishers.
Why the intervention matters
The government’s position is not binding on the court. The court must independently evaluate the statutory fair-use factors and the factual record developed by the parties.
The filing is nevertheless significant. It appears to be the federal government’s first formal intervention in the growing wave of copyright cases challenging the use of protected material for AI training. It also gives the court a policy argument for separating model training from potentially infringing outputs.
Recent cases have already reached different conclusions when examining related issues. Judges in cases involving Anthropic and Meta have treated certain training uses as transformative, while also recognizing that the origin of training materials and the possibility of market substitution may affect the analysis.
The New York case may provide one of the clearest answers yet because both sides are now seeking a ruling on the fair-use defense. Reuters reported on the competing summary judgment motions on September 8.
What remains unresolved
The DOJ did not argue that every AI-related use of copyrighted content is lawful.
Its filing focused specifically on the training-stage use of written works. It did not eliminate potential liability involving outputs that reproduce protected expression, nor did it provide a blanket defense for unlawfully obtained datasets, circumvention of access controls or violations of contractual restrictions.
The government also acknowledged that fair use requires a fact-specific, use-by-use analysis. The source of the material, the purpose of the copying, the operation of the model, the content of its outputs and the effect on existing markets may all remain relevant.
The Times and the author plaintiffs argue that generative AI products compete with their work, reduce demand for original content and threaten established licensing markets. OpenAI and Microsoft contend that their systems learn unprotectable patterns and do not provide substitutes for the underlying books and articles.
Those competing market arguments remain central to the court’s analysis.
What businesses should do now
AI developers should continue documenting where training materials originated, how they were obtained and what permissions or restrictions apply. Businesses should distinguish between lawfully accessed public content, licensed databases, paywalled material and content obtained through third parties.
Contracts with data providers and AI vendors should clearly address training rights, permitted uses, output controls, ownership, representations, indemnification and responsibility for infringement claims.
Companies deploying third-party AI tools should also determine whether their own materials or customer content may be used for model training. Confidential documents, proprietary data and client materials should not be uploaded without reviewing the platform’s terms and the company’s contractual obligations.
Content owners should maintain clear records of ownership, evaluate licensing opportunities and monitor whether AI outputs reproduce meaningful portions of protected work. Technical restrictions, contractual terms and enforcement procedures should be reviewed together rather than treated as separate issues.
Fair use is not a company policy
The DOJ’s filing strengthens the argument that AI training can be transformative. It does not make fair use a substitute for responsible data governance.
A court may ultimately agree that training an AI model is protected while finding that the acquisition of certain materials, particular outputs or separate commercial uses create liability. A different model, dataset or business practice could also produce a different result.
For now, companies should avoid treating the government’s position as universal permission to train on any available content. The better approach is to understand each use, document how the material entered the system and build contractual protections around both training and outputs.
