That is an opinion, but certainly isn’t settled law in any jurisdiction. Indeed, the answer to whether some, all, or none of an LLM’s output is ever copyrightable and under what terms is the billion dollar question.

A project that incorporates code with shaky legal foundation will find it tough to convince others to contribute, if it’s possible one day that their contributions were in vain. The right answer would be to extricate such code upon discovery, like what OpenBSD had to do when the IPFilter license turned out to be incompatible with the project.

What if license and copyright was washed by using an LLM to translate Claude into another language?

The law doesn’t allow you to launder copyright like that. That’s just a derivative work, which can be restricted by the copyright holder in the original. As an example, in fictional writing, distinct characters are copyrighted, and using an LLM to generate new works using those copyrighted characters would still be a derivative work that the original copyright owner would have the right to deny distribution.

So if you have a copyrighted codebase and you try to implement that codebase using some kind of transformation of that code, that’d still be a derivative work and infringe the original copyright.

Now if you have some kind of clean room implementation where you can show that it was written without copying the original code itself, only working to implement its functionality through documentation/reverse engineering how the code worked, you’d be able to escape out of calling it a derivative work and could distribute it without the original copyright holder’s permission (Compaq did this with the IBM BIOS to make unauthorized/unlicensed PC clones, and Google did this with the Java API to make Android without a license from Sun/Oracle and won at the Supreme Court).

Claude can’t be copyrighted because it’s a product of an LLM.

No, because Claude’s code is still created by humans with the assistance of non-human tools. There’s a spectrum from spelling correction and tab completion in IDEs all the way to full vibe coding with a prompt describing the raw functionality (where the prompt is so uncreative that it isn’t itself copyrightable). Anthropic has never claimed that there was no human in the loop, or that the prompts it uses are so uncreative and purely functional so that the outputs aren’t copyrightable.

Claude can’t be copyrighted because it’s a product of an LLM

You claim Claude itself was coded by an LLM (exclusively)?

This has already been done.

https://github.com/Kuberwastaken/claurst

They got one AI agent to read the code and output documentation (a specification) that contained no actual code.

Then they fed that to a second AI agent and asked it to implement that specification in Rust. Technically, this is a cleanroom implementation.

This is interesting from a legal perspective because it leverages the same legal loophole that Anthropic rely on for their own operations. They can’t take down this repo without creating a precedent that will be used against them.

Depending on how you’re taking “inspiration” I’m fairly certain you can get sued if Anthropic feels like it.

Clean room development is there for a reason (although can you use AI in such an effort?)

The original code is reportedly 100% slop. Slop cannot be copyrighted in the US, and probably any other half-sane country. So double slop is ofc also not licensed.