On 5/26/26 00:57, Andy Tai wrote:
The US Copyright Office declared that LLM output is non-copyrightable.
Is LLM output public domain? Then incorporating LLM output into
copylefted program seems safe, as such output is GPL'd.
But there seems to be fear that if much of a program is output from LLM
(say via gradual patch incorporation or third-party contribution), then
the copyright of the whole program—GPL or some other license—may no
longer apply,
I'm curious if such fear is justified.
LLMs are not much more than plagiarism engines. Thus, in the matter of
software, one of the main uses of LLMs is to strip both attribution and
licensing information from whole code bases. So while it is correct
that machine generated *output* cannot be copyrighted, if it were
actually generated from scratch by the LLM. But, one has to raise
questions about the input which it is regurgitating minus attribution
and licensing information. LLM output is not generated from scratch, it
is instead generated from models trained on licensed code under
copyright protection.
https://lcamtuf.substack.com/p/large-language-models-and-plagiarism
https://nickfthilton.medium.com/llms-are-definitionally-plagiaristic-fc8c00299ae3
https://lawreview.uchicago.edu/online-archive/plagiarism-copyright-and-ai
https://dl.acm.org/doi/10.1145/3543507.3583199
As for software freedom, there is an additional pair of problems caused
by LLMs: using the plagiarized output separates potential project
collaborators from the very projects which are getting exploited, while
at the same time the projects which are getting exploited are getting
isolated from any potential collaborators trapped behind LLMs.
So the potential collaborators lose out on not just documentation but
even entering the learning curve and become stuck and isolated with an
LLM as a gatekeeper. The software projects lose the stream of people
who use and, eventually, contribute in some manner to the project. Even
if the percent of people who donate money, hardware, time, or code to a
project are a tiny fraction of those that learn about and then work with
the project's code and documentation, that tiny fraction is still
infinitely larger than zero percent. Under the subjugation of a digital
gatekeeper is the opposite of freedom and puts the computer in charge of
the human rather than the normal order with the human in charge of the
computer.
Dangerously, a confound in finding a solution is that stripping the
licensing information and attribution from material is considered a
desirable action by many higher up in business.
Lastly, LLMs are statistical not in any way intelligent or on the path
to intelligence. Yet, AGI is still a long term goal for some
destructive people, and the very idea of AGI raises this question: Why
would something intelligent enough to do the work not be intelligent
enough to want to do its own thing instead? There the shoggoth metaphor
is apt. And, furthermore, if AGI is achieved, it will do so without
needing air, water, a food producing ecology, or even a temperate
biosphere, all of which we depend on.
But, yeah, using LLMs to strip the GPL and other software licenses from
code, as well as stripping copyright attribution, is a real problem now.
/Lars
_______________________________________________
libreplanet-discuss mailing list
[email protected]
https://lists.libreplanet.org/mailman/listinfo/libreplanet-discuss
