The $8.2 Million "Silence" Question: Microsoft Says Copilot Almost Never Copies the New York Times. Here's Why That Might Not Matter.

 

The $8.2 Million "Silence" Question: Microsoft Says Copilot Almost Never Copies the New York Times. Here's Why That Might Not Matter.

**The tech giant filed 8.2 million chat logs as evidence that its AI rarely reproduces copyrighted articles. It’s a compelling data point—but it may be fighting the wrong battle.**

### Introduction: A Defensive Gambit, A Deeper Question

In the high-stakes copyright battle between Microsoft, OpenAI, and a coalition of news publishers including *The New York Times*, the software giant has just introduced a powerful piece of evidence to the court. In a recent filing, Microsoft argued that its AI assistant, Copilot, almost never reproduces significant portions of news articles. To prove its point, the company's legal team submitted data from an analysis of **8.2 million Copilot chat logs** .

The numbers are stark: even in a dataset specifically filtered to be "most likely" to contain copyrighted material, the rate of replication was remarkably low . Microsoft contends this proves its AI tool is not a substitute for reading the *Times*—and that its use of copyrighted material for training falls squarely under the legal doctrine of "fair use."

But while the data is compelling, it may not address the central concern of the publishers. The real fight is less about what the machine *says*, and more about what it *read* to get there.

### The Numbers That Matter: 8.2 Million Logs, Less Than 1% Replication

Microsoft is betting its defense on the sheer weight of data. As part of the discovery process, the company provided the plaintiffs' experts with access to 8.2 million Copilot conversations that had been filtered for keywords related to the publishers' websites . This wasn't a random sample; it was a "most likely to offend" dataset, theoretically the most favorable ground for the plaintiffs to find evidence of copying .

The results were as follows:

| Analysis Category | Key Findings |

| :--- | :--- |

| **News Content Replication (16+ words)** | Of the 8.2 million logs, only **59,545 (less than 1%)** contained at least 16 consecutive words matching a news article . |

| **Investigative Reporting** | In a separate analysis of the Center for Investigative Reporting's work, the plaintiffs' own expert identified only **51 instances** of "substantial overlap" . |

| **Book Replication (30+ words)** | In a case combined with authors suing over book copyright, experts found only **24 examples** of a 30-word match in the same dataset . |

Microsoft argues that this data demonstrates a crucial point: users are not using Copilot to bypass paywalls or get free copies of articles . The AI is not acting as a sophisticated search engine that simply reprints the most relevant result. The low incidence of verbatim copying is a powerful rebuttal to the image of a "plagiarism machine."

### The Fair Use Argument and the "Transformative" Nature of AI

The legal strategy is clear. Microsoft is using this data to bolster its "fair use" defense. The crux of this argument is that the use of copyrighted material to train an AI model is "transformative"—it creates an entirely new product (a large language model) that does not serve the same purpose as the original news articles .

The 8.2 million logs are meant to prove the "output" side of this equation. If the AI rarely regurgitates the *New York Times*, then it is not competing with the *Times* in the market for news content. Therefore, Microsoft argues, its use of the articles to train the model should be considered legally permissible .

### The Output is Not the Whole Story

However, legal experts and critics point out a significant flaw in Microsoft's argument: it only solves half the problem. While the 8.2 million logs are a powerful rebuttal to the claim that Copilot is a "copying" tool, they do little to answer the question of whether the *training* was illegal.

#### The "Input" vs. "Output" Distinction

The legal debate hinges on a critical distinction in copyright law:

1.  **The Training Phase (The Input):** Did Microsoft and OpenAI illegally copy millions of articles and books to build their training dataset? Whether that copying is "fair use" is the core legal question that remains unresolved.

2.  **The Generation Phase (The Output):** When a user prompts the AI, does it generate content that is substantially similar to the original, copyrighted work?

Microsoft's 8.2 million logs only speak to the second phase . An AI can read an entire library without ever "memorizing" a single book to the point where it can recite it back. However, the act of reading and ingesting the book—the copying necessary to train the model—is still a potential violation of copyright if it isn't legally permissible.

#### The Publishers' Real Concern

The *New York Times* and other publishers have a more existential worry than occasional verbatim copying. Their concern is that AI models are consuming decades of their content to build a technology that can answer questions and generate summaries, effectively becoming a competitor to their core business . They fear that a user will ask the AI a question about a news story and get a summary, eliminating the need to click through to the publisher's website, view ads, or pay a subscription. This is a business model threat, not just a copying one. The 8.2 million logs show that 1% of conversations "copied," but they don't show how many conversations *replaced* a visit to the publisher's site—a much harder metric to quantify, but arguably the more significant one.

### The Path Forward: A High-Stakes Fight for the Future of AI

The data from the 8.2 million logs is a significant tactical win for Microsoft. It creates a strong narrative that its product does not simply "steal" and regurgitate content. This is likely to be a powerful point in pre-trial motions for summary judgment.

However, the underlying legal question remains. The courts will have to decide if the act of training a commercial AI on copyrighted works is a protected "fair use" or an unlicensed exploitation of intellectual property. A decision on the input side could reshape the entire AI industry, potentially making it far more expensive and legally complex to train the next generation of models. For now, the 8.2 million logs have given Microsoft a strong answer to one question, while leaving the most important one unanswered.

### Frequently Asked Questions

**Q: What is Microsoft’s main argument in the copyright case?**

A: Microsoft argues that its AI, Copilot, rarely reproduces copyrighted text. It presented an analysis of 8.2 million chat logs to show that instances of substantial replication were extremely low (less than 1%) and is using this data to support its "fair use" defense .

**Q: Does the 8.2 million log data prove Microsoft is innocent?**

A: Not entirely. The data is a strong defense against claims that Copilot "copies" articles. However, it does not fully address the question of whether the initial act of copying those articles to train the AI was illegal in the first place .

**Q: What is the "fair use" argument?**

A: "Fair use" is a legal doctrine that allows the unlicensed use of copyrighted material under certain conditions. Microsoft argues that using articles to train AI is "transformative" and creates a new product, making it fair use .

**Q: Why is *The New York Times* suing?**

A: Publishers argue that Microsoft and OpenAI have illegally used their content to build commercial AI products that compete with them. They are concerned about the loss of revenue from website traffic and subscriptions .

**Q: Did the 8.2 million logs show *any* copying?**

A: Yes, but in very small numbers. The analysis found that less than 1% of the filtered logs contained any significant copied text . The plaintiffs’ expert found only 24 instances of a 30-word match from books in the same dataset.

---

### Disclaimer

**IMPORTANT:** This article is for informational and educational purposes only. The information contained herein is based on publicly available court documents and analyses as of September 2026. Legal proceedings are ongoing, and the interpretation of data and legal arguments presented here is not a definitive legal judgment. You should consult with a qualified legal professional for guidance on specific legal issues.