Microsoft defends Copilot's book usage while selling advanced GPT-6 Astra
Tech giant argues minimal copyright impact as it launches OpenAI's most secure AI model.
1Microsoft's copyright defense
Microsoft filed for summary judgment on September 4, revealing Copilot reproduced book passages only 24 times across 8.2 million conversations (0.00029% rate). Their expert found no matches for 202 of 212 disputed books. The case consolidates claims from authors and 400 newspapers before Judge Sidney Stein, with Microsoft seeking a legal resolution on whether AI training constitutes fair use under US law.
2GPT-6 Astra launch
Two days before its legal filing, Microsoft began selling GPT-6 Astra through its Foundry program at $10 per million tokens. OpenAI classifies Astra as meeting its Critical cybersecurity threshold - the first model to achieve this rating. The AI can autonomously read screens, update records, test software, and navigate development tools while maintaining activity logs and human oversight checkpoints.
3European legal landscape
The Munich Regional Court ruled on July 31 that AI models memorizing works in their parameters infringe reproduction rights, regardless of output frequency. This contrasts with US fair use arguments. Meanwhile, Article 55 of the EU AI Act imposes systemic risk duties on general-purpose models like Astra, requiring adversarial testing, incident reporting, and infrastructure security - none tied to content replication rates.
4Astra's cybersecurity credentials
OpenAI's Preparedness Framework rates Astra Critical for its ability to find unknown vulnerabilities and build exploits without human guidance. Microsoft implements scoped credentials and human checkpoints, claiming prompts/outputs aren't used for training. The model launched in Global/US data zones despite Azure offering EU zones since November 2024, where cloud dependency remains politically sensitive.
5Legal vs product narratives
Microsoft simultaneously argues its AI retains minimal copyrighted material while marketing Astra's advanced capabilities. This dual approach highlights the disconnect between legal defenses (focusing on output rarity) and product value propositions (emphasizing capability). The disparity is particularly stark in Europe where memorization itself may constitute infringement regardless of output frequency.
6Pricing and deployment
Astra's $10/million token pricing positions it as a premium offering. While available globally, its EU deployment lags despite Azure's existing infrastructure. This timing gap occurs as European regulators intensify scrutiny on AI systems, with the Munich ruling potentially setting precedent for how courts view AI training data retention across the continent.
7Industry implications
The case could set important precedents for AI developers worldwide. A US fair use determination would contrast sharply with Europe's reproduction rights approach, potentially creating compliance challenges for global companies. Meanwhile, Astra's Critical rating may become a benchmark for enterprise AI security standards, especially in regulated industries.
8Regulatory evolution
The litigation highlights how AI regulation is developing asymmetrically across jurisdictions. While the US focuses on output-based assessments, Europe examines training processes and systemic risks. Microsoft's concurrent legal defense and product launch demonstrate how tech giants must navigate these diverging frameworks while advancing AI capabilities and market positions.