The prompts shown literally invite the LLM to complete the copyrighted text by providing unedited selections and asking the machine to finish those. Even if this is problematic in a small number of cases it is not a use case that undermines the business model of the newspaper since it requires the reader to have access to the original text. Nor will it be easy to demonstrate economic harm since this is not how readers consume news and is very far from how users interact with LLMs. Nor are the archival materials used for training remotely reflective of the "time-sensitive" articles that newspapers sell. And archival materials are easily available elsewhere so where is the case for economic harm?

The courts are going to rule that LLM training is a transformative use case that is protected as fair use under copyright law. They may rule that if an LLM-powered service is explicitly designed to enable copyright violation that is illegal, but there is no way any court is going to look at these examples and see it as anything other than the NYT fishing to try and generate a violation by using the LLM in a way that is very different than the service is intended to be used and which -- even if abused -- doesn't hurt the business model under which the text has been produced.

The most likely outcome is that LLM providers will add some sort of filter on output to prevent machines from regurgitating source documents. But this isn't a court case the NYT can win without gutting fair use protections, and that would be a terrible thing.

Meanwhile, open source LLMs are excluded from stringendo regulation in the US, abd with Mistral there is some knowhow that isn't in SV, which is also jicem