On the regurgitation/attribution point specifically: no, I don't think anyone can currently give a categorical guarantee that an LLM's output contains zero verbatim reproduction of training-data code. That's not specific to Claude or to this browser, it's inherent to how these models are trained, and I'd rather say that plainly than assert a certainty I can't back up.
What actually mitigates the risk in practice, for my own contributions: - I read and understand every suggested change before accepting it — I don't merge blocks I haven't reviewed line by line. - Short, idiomatic, boilerplate-level matches (common loop constructs, standard API call patterns) aren't a meaningful risk on their own — they're not original enough to be protectable regardless of whether a human or a model typed them. - Larger verbatim reproduction is the real risk category, and catching that honestly needs the same tooling the industry is already building for this exact problem (clone/license scanners like ScanCode or FOSSology, GitHub's own Copilot duplicate-detection feature) — not a developer's assurance after the fact. I'd rather disclose the practice and its limits than claim a certainty nobody actually has — that's the same reason I disclose AI-tool use up front rather than using it quietly. If a sponsor or the project wants a specific scan run before upload, I'm open to it. (Sending this as a reply as the reply from the top of this email might be incorrect information) Sincerely, Ştefan Andrei Ionuţ În mar., 20 oct. 2026 la 10:09, Andrei Ionuţ Ştefan < [email protected]> a scris: > Yes, but let me tell you something: I use Claude Code for my development > too. Its standard tooling, BUT i do use standard licences, and i disclose > clearly that i used AI tooling. > Still, that is just an assumption by the United States of America, I live > in the EU (specifically Romania), so its still unknown if there is a ruling > about whether AI tooling can be copyrighted or not. > Its still not global consensus yet, so your info might be inaccurate. > > Sincerely, > Ştefan Andrei Ionuţ > > În mar., 20 oct. 2026 la 10:04, Stephan Verbücheln <[email protected]> > a scris: > >> There are also questions regarding the license, since the project >> appears to have a significant number of code generated by an LLM >> (Claude). >> >> * Since LLM-generated output is not copyrightable, can you even >> legally dictate what license the browser has? Can we rely on the >> assertion that it is GPLv3-licensed? >> * Can you rule out that copyrighted code sections from the training >> data are in the browser's code base without the author's permission >> or without correct attribution as required by many software >> licenses? >> >> Please not that many non-copyleft licenses also require correct author >> attribution. >> >> Regards >> Stephan >> >>
