Who owns the code Claude Code wrote?

> What to preserve: Commit messages that describe what you changed and why, not just what the AI generated. “Restructured Claude’s module architecture, rejected initial state management approach, rewrote error handling from scratch” is evidence. “Add rate limiting module” is not.

> The second commit message versus the first is the difference between a defensible authorship claim and a clean “Claude wrote this” record.

That makes no sense to me, as the commit message is probably LLM generated as well. (and even easier to generate as it doesn't have to compile or pass automated tests).

[flagged]

Indeed a developer owns copyright over the source code and on the compiled binaries, because there is no expansion happening here but just a translation from one format into another, the kind of thing that has been ruled copyrightable since copyright exists. The same goes for translations from one human language into another, and anybody with knowledge of more than one language will be happy to acknowledge that translating is hard work. Even so, the translator does not hold copyright on the result, at best they can say they have created a derived work and it is the original author that continues to hold copyright.

Compilation and translation happen in a generic manner and does not rely on a mountain of other IP, it is really just a transformative tool that happens to do something useful, someone constructed it to be a very precise translation to the point that any mistakes in it are called bugs and we fix them to ensure the process stays deterministic. Translators try hard to 'get it right' too: to affect the intentions of the original author as little as possible.

When you use a model loaded up with noise or that you have trained exclusively on code that you actually wrote I think a strong case could be made that you own the copyright on that work product. But when you train that model on other people's work, especially without their consent or use a model that has been trained in that way you lose your right to call the output of that model yours.

You did not write it, and the transformative process requires terabytes of other people's IP and only a little bit by you.

As soon as you can prove that your contribution substantially outweighs the amount of IP contributed in total you would have a much stronger case.

This is a big question that makes my employer nervous about using LLM-generated code, along with the even-more-unresolved question "what happens if the LLM outputs an algorithm that is protected by patent?" (particularly worrying because we know the base training included patent descriptions.) Questionable copyright can often be worked around (particularly since we don't distribute source) but infringing on a patent can destroy a company.

+1

Adding two subtle points:

>> Indeed a developer owns copyright over the source code and on the compiled binaries, because there is no expansion happening here but just a translation from one format into another ... does not rely on a mountain of other IP

... and, the license agreement of the compiler and libraries used / linked to practically always explicitly waive copyrights over the said non-mountain of IP.

>> As soon as you can prove that your contribution substantially outweighs the amount of IP contributed in total you would have a much stronger case.

... a much stronger case that you have a partial copyright over the work, which is now likely a derivative work. You still may not have a case that you own the copyright exclusively (or as the original article says, that your employer does).