Microsoft Tells Court Copilot Rarely Copies Books While Selling OpenAI’s Most Capable Model
Two Filings, One Week, and a European Framework
September 4, 2026 – 6:45 pm
Microsoft argues in a New York AI copyright litigation that its Copilot model reproduces book passages only 0.00029 percent of the time, with no matches for 202 out of 212 books analyzed. They submitted this argument while selling GPT-6 Astra, their most capable model, through Foundry.
Microsoft made two key points:
- Training large language models on books is considered fair use. This was outlined in a summary judgment memorandum citing an expert’s findings.
- GPT-6 Astra has limited book retention capabilities. The model, rated Critical for cybersecurity by OpenAI, does not retain or reproduce content from memorized works stored in its parameters, according to Microsoft.
In Europe, however, neither the output rate nor sales pitch are considered under the AI Act’s framework. A recent Munich Regional Court ruling determined that works stored in a model’s parameters infringe reproduction rights, regardless of their rarity.
Microsoft further emphasized that while Astra can perform tasks like reading on-screen information and updating records, fair use has no direct equivalent in European law.
Pricing is another notable aspect, with Astra launching globally but Azure offering an EU data zone since November 2024, amid growing political debates over cloud dependency.