NYT’s examples ensure it.

Timothy Lee and James Grimmelmann, “Why The New York Times might win its copyright lawsuit against OpenAI,” 02-20-2024, https://arstechnica.com/tech-policy/2024/02/why-the-new-york-times-might-win-its-copyright-lawsuit-against-openai/

In its lawsuit against OpenAI, the New York Times provided 100 examples of GPT-4 generating long, near-verbatim excerpts from Times articles:

Many people in the AI community have underestimated the significance of these examples.

Those who advocate a finding of fair use like to split the analysis into two steps, which you can see in OpenAI’s blog post about The New York Times lawsuit. OpenAI first categorically argues that “training AI models using publicly available Internet materials is fair use.” Then in a separate section, OpenAI argues that “‘regurgitation’ is a rare bug that we are working to drive to zero.”

But the courts tend to analyze a question like this holistically; the legality of the initial copying depends on details of how the copied data is ultimately used.

For example, when the Second Circuit considered the fairness of Google’s book scanning, it closely scrutinized how Google’s book search engine works. The ruling noted that users were only ever shown short snippets and could never recover longer passages from a copyrighted book.

OpenAI dismisses regurgitated results as a “rare bug”—and maybe it is. But Google was able to tell the courts that its search engine never reproduces more than a small fraction of any copyrighted book without permission, because it cannot.

Examples of verbatim copying undermine the argument that generative models only ever learn unprotectable facts from their training data. They demonstrate that—at least some of the time—these models learn to reproduce creative expression protected by copyright. The danger for AI defendants is that these examples could color the judges’ thinking about what’s going on during the training process.