Microsoft Copilot Logs Show Near-Zero Copyrighted Text Overlap

Overview

As major generative AI developers face legal action over web-scale training data, Microsoft has taken an empirical stance in court to refute claims of copyright infringement. In recent legal filings, the company presented extensive telemetry data from its AI assistant, Copilot, aiming to prove that the system rarely reproduces copyrighted source material in its responses.

By analyzing millions of actual user conversations targeted with publisher-specific keywords, Microsoft contends that output overlap with copyrighted works is statistically negligible. This empirical defense is designed to bolster the argument that foundation models process content transformatively rather than serving as alternative content delivery mechanisms.

Quantifying Content Overlap in the Wild

To test the claims of publisher plaintiffs—which include major news organizations and book authors—a dataset of 8.2 million Copilot interaction logs was subjected to forensic analysis. These logs were specifically filtered using key terms closely associated with the plaintiffs' published content to maximize the likelihood of finding matching passages.

The resulting data demonstrated an extremely low frequency of verbatim duplication:

• Minor Overlaps: Out of the 8.2 million keyword-matched interactions, approximately 59,545 logs contained at least 16 matching words in sequence with publisher content.
• Substantial Matches: When evaluating longer passages of 30 consecutive matching words or more, the total dropped dramatically to just 24 individual responses across the entire multi-million log dataset.
• Book Content Duplication: Across 212 literary works evaluated by opposing experts, only 10 books produced any observable textual overlap within the prompt logs.

This data indicates that instances where an LLM outputs extended, verbatim passages of copyrighted material are statistically rare anomalies rather than systemic operational behavior.

The Transformative Use Defense

The central pillar of Microsoft’s legal argument rests on the doctrine of fair use under U.S. copyright law. Fair use evaluates whether a new technology alters the original work with new expression, meaning, or message, and whether it disrupts the primary market for the original content.

Microsoft argues that because Copilot synthesizes information across vast web corpora to generate novel context-aware text, its function is fundamentally distinct from the original journalism or literature upon which it was trained. The statistical absence of significant verbatim regurgitation supports the position that the AI does not act as an economic substitute for reading original articles or purchasing books.

Strategic Push for Summary Judgment

By filing this evidence, Microsoft is seeking a summary judgment to halt the litigation prior to a full trial. A favorable ruling for Microsoft and its partner OpenAI would establish a critical legal precedent for the broader artificial intelligence industry, affirming that scraping public web data for model pre-training falls squarely within fair use protections if the resulting runtime system does not duplicate original works at scale.

Conversely, if the court determines that even minimal, incidental overlaps constitute non-transformative copying, AI developers may face heightened obligations regarding pre-training data licensing, safety alignment filtering, and output auditing.