Microsoft informs the court that Copilot seldom replicates books while marketing OpenAI's most advanced model.

      On 4 September, Microsoft filed for summary judgment in the New York AI copyright litigation, contending that Copilot reproduced passages from books 24 times across 8.2 million conversations. Just two days prior, the company started selling GPT-6 Astra through Foundry, a model that OpenAI categorizes as Critical for cybersecurity, yet European law does not assess the output frequency or the sales claims.

      This past week, Microsoft presented two arguments simultaneously. In a memorandum for the summary judgment, it stated to a Manhattan federal court that training large language models on books constitutes fair use, citing an expert who identified 24 matching responses in 8.2 million Copilot dialogues. This corresponds to 0.00029 percent. The filing indicates that for 202 of the 212 books involved, the expert found no matches at all.

      The case is being heard in the multidistrict litigation led by Judge Sidney Stein, which consolidates claims from authors and publishers, including 400 local newspapers. Microsoft is seeking a legal resolution to the entire matter.

      On 3 September, just two days before, Microsoft introduced its most advanced model, GPT-6 Astra, through the Foundry Limited Access Program, priced at $10 per million input tokens. Astra is designed for computer use, with an Azure post stating that it can read on-screen information and operate within approved interfaces, updating records, testing software, and navigating development tools.

      Microsoft complements this with scoped credentials, human checkpoints for significant actions, and activity logging. It asserts that prompts and outputs are not utilized to train the models. OpenAI has classified Astra as meeting the Critical cybersecurity criteria under its Preparedness Framework, marking it as the first model to achieve this designation. This indicates it can identify unknown vulnerabilities and develop exploits without detailed human oversight.

      Thus, one filing claims that the model retains minimal information, while the other emphasizes its capabilities. However, in Europe, neither assertion acts as a legal standard.

      On 31 July, the Munich Regional Court ruled that works memorized in a model's stored parameters infringe upon the reproduction right at that moment. This ruling arose from the GEMA v. Suno case, which is currently under appeal. The occurrence of a copy in the model's weights, regardless of how infrequently it appears, does not alter what has transpired. There is no European equivalent to fair use to rely upon.

      The product implications follow the same reasoning. Article 55 of the AI Act assigns responsibilities to general-purpose models that present systemic risks, specifying obligations based on capability rather than output. Providers must conduct documented adversarial testing, promptly report serious incidents to the AI Office, and secure the model along with its physical infrastructure. None of these responsibilities depend on how often a model reproduces a book.

      There is a notable gap in the pricing details. Astra was launched with Global and US data zone deployments, whereas Azure has offered an EU data zone since November 2024, in a region where reliance on cloud computing is already a contentious political issue.