DominicP <[email protected]> writes: > On Mon, 12 Oct 2026 at 08:54, Richard Sandiford <[email protected]> > wrote: > >> Dominic P <[email protected]> writes: >> > Richard Sandiford <[email protected]> writes: >> >> What kind of assistance did it provide? >> > >> > While working on vectorisation, Claude spotted a preexisting bug. I >> > agreed, the patch was created, and Claude ran the tests for me. I have >> > reviewed Claude's work at every stage. >> >> Ok, thanks. The use of the passive tense kind-of leaves open the question: >> was any of the patch produced using Claude, or was it all your own work? >> My understanding of the AI policy is that changes outside of the testsuite >> need to be authored entirely by a human (unless the patch is trivial). >> >> In other words, it's ok (AIUI) for someone to use AI to detect flaws and >> create testcases for those flaws, and for that person to then go off and >> write a fix without the help of AI. But using AI to write the fix is out >> of bounds. >> >> I realise the policy somewhat incentivises people to lie about the use >> of AI, but... >> > > In this particular case Claude suggested the solution while I was working > on something else. I > reviewed the solution and was unable to come up with a better solution. My > understanding is > that small amounts of code are fine if they have been reviewed. If you feel > this isn't within the policy > then we can drop it and leave the bug as is.
I think the "problem" (as for as the policy goes) is that you're submitting a lot of patches in this area (thanks!) and so it's hard to argue that the net effect is not copyrightable. It would now be difficult for me to submit a patch for the bug without being influenced by the patch that was posted. I think in future, if someone is using an AI tool that generates both a problem description and a candidate patch, the best thing would be to file the bug as a PR and leave someone else to write a patch independently. How does this affect: https://gcc.gnu.org/pipermail/gcc-patches/2026-September/733012.html ? In that case, Richard and I suggested a different approach, so it seems like in that case there was an opportunity to write a fix without using an AI tool, even if an AI tool was used to find the bug. Thanks, Richard
