Re: [Hamlib-developer] The use of LLM generated code in Hamlib (long)

[Greg T. <gd...@le...>]

I too fall into the camp that LLM output is a derived work of its
training inputs, and that LLMs trained on code without permission are
basically copyright infringement at scale.

+1 to require Signed-off-by: asserting that the person has the right to
license the code under the project's license -- which means it is not
from an LLM.


Separately from licensing, I am seeing discussion in multiple projects
where people are submitting code that is LLM-generated.  It's buggy and
project maintainers end up spending time dealing with it (review, fix).
I recently saw a "new project" with pretty big claims, and then a human
commented fairly extensively and quite reasonably (technically sensible
and polite).  That resulted in an LLM-generated reply to the comments,
at which point the human called the LLM-sock-puppet on their behavior
and the converation ended.

For unison, a non-ham project I maintain, I decided that LLM code was
not ok for both of these reasons, and CONTRIBUTING.md reads:

  ## LLM usage

  LLMs essentially always use training input without copyright
  permission, and the unison maintainer considers the output from LLMs
  to be a derived work of the training data.  Thus *LLM-generated code
  is simply not acceptable for submission to the project*.  This is
  implied by the previous requirement that the submitter has permission
  to license the code under the project's license, but stated explicitly
  because that seems to be necessary.

  It is ethically unacceptable to ask humans to review and deal with
  code generated by LLMs.  Even if there were no licensing issues,
  LLM-generated code would still be unwelcome.

(I, a human, am the sole author of that text and you are welcome to take
it!)

At least for now, I strongly recommend hamlib take a "no LLM output is
acceptable within the project" stance.

73 de n1dam