This story should terrify any open source project or company that allows LLM-generated code into their project.
Someone created a new app, using Claude. Only it wasn’t a new app, it was clearly plagiarised because it happened that there was an example in the training set that exactly matched the requirements. The similarity was well within the range that courts have previously used to determine a derived work.
When you have this level of similarity, the requirement comes to you to prove that there was no way that the original work could have flowed to your project. Companies that have this concern usually do it by ensuring that no one on the team has been exposed to the original and that the code for the original never goes near their systems. But when one of the systems that you use is a language model trained on, among other things, all of the open-source code that it could scrape (and which does not disclose its training set), being able to prove that there was no path from some other codebase to yours is impossible.
Just because the US copyright office has ruled that you, as the person promoting an LLM, cannot assert copyright on the output, does not mean that someone else can’t. If you take a DVD and transcode it to H.264, there is no creative step and so the new copy is not something subject to independent copyright, but it is a derived work of the DVD (itself a lower-quality derived work of the original masters) and so subject to the same copyright.
Importantly in this story, the person prompting Claude had no idea that the original app existed. To safely use the code, they would need to search everything in the training data and discard outputs that would meet the bar of being substantially similar. And that’s something that requires human judgement.