Microsoft tells court Copilot rarely copies books while selling OpenAI’s most capable model

Microsoft moved for summary judgment in the New York AI copyright litigation on 4 September, arguing Copilot reproduced book passages 24 times in 8.2 million conversations. Two days earlier it began selling GPT-6 Astra through Foundry, a model OpenAI rates Critical for cybersecurity, and European law measures neither the output rate nor the sales pitch.

Microsoft spent this week making two arguments at once. It told a Manhattan federal court that training large language models on books is fair use, in a summary judgment memorandum saying an expert found 24 matching responses across 8.2 million Copilot conversations.

That is 0.00029 percent. For 202 of the 212 books at issue, the filing says, the expert found nothing at all.

The case sits in the multidistrict litigation before Judge Sidney Stein, which gathers claims from authors and publishers including 400 local newspapers. Microsoft wants the whole question decided as a matter of law.

Two days earlier it started selling the most capable model it has. Microsoft announced GPT-6 Astra in the Foundry Limited Access Program on 3 September, at $10 per million input tokens.

Astra does computer use. The Azure post says it can read on-screen information and act inside approved interfaces, updating records, testing software and navigating development tools.

Microsoft pairs that with scoped credentials, human checkpoints for consequential actions and activity records. It says prompts and outputs are not used to train the models.

OpenAI rates Astra as meeting the Critical cybersecurity threshold under its Preparedness Framework, the first model it has designated at that level. That means finding unknown flaws and building exploits without step-by-step human guidance.

So one filing argues the model barely retains anything, while the other sells what it can do. In Europe, neither claim is the test.

The Munich Regional Court ruled on 31 July that works memorised in a model’s stored parameters infringe the reproduction right at that point. The case was GEMA against Suno, and it is under appeal.

If the copy exists in the weights, how rarely it surfaces does not change what happened. Fair use has no European equivalent to fall back on.

The product side runs on the same logic. Article 55 of the AI Act attaches duties to general-purpose models carrying systemic risk, and those duties describe capability rather than output.

Providers must run documented adversarial testing, report serious incidents to the AI Office without undue delay, and secure the model and its physical infrastructure. None of that turns on how often a model repeats a book.

One gap sits in the pricing table. Astra launched with Global and US data zone deployments, though Azure has offered an EU data zone since November 2024, in a region where cloud dependency is already a political argument.