Falsities of LLM Negationists

home

2026-07-24

Falsities of LLM Negationists

There is one claim about LLMs that irritates me more than all the discussions about prompting, AGI, and stochastic parrots combined: “We don’t know if it is yours”.

Sometimes the wording changes: “There might be a copyright owner out there”. “It may be public domain”. “We don’t know whether it can be copyrighted”.

A person spends a month building something.

They generate hundreds of outputs, most of them garbage. They rewrite prompts, edit results, combine fragments, throw entire approaches away, start over, and fix the same bug three times because the model confidently reintroduced it. Eventually, after thousands of decisions, they arrive at the thing they were trying to make.

Someone looks at the finished work and says:

“It isn’t yours”

The month disappears. The rejected attempts disappear. The selection, editing, arrangement, and judgment disappear. Nobody has identified another author. Nobody has pointed to an infringed work. Nobody has shown substantial similarity to anything.

The only contribution we can actually observe - the work of the human sitting behind the keyboard - is treated as the least credible fact in the case.

This is usually presented as caution. It is nothing of the sort.

Work which might not qualify for copyright protection is not evidence that the person did not create it. Whether some work components might be in the public domain does not mean that the composition is. Pointing out that unknown copyright owner might exist somewhere does not give that imaginary person a better claim than the actual author standing in front of us.

Yet this is exactly how the argument works. A possibility is raised; Author loses by default.

That default is dehumanizing in a precise sense. It treats the human as a non-significant operator while assigning all meaningful causation to the machine. Their choices count only when someone can prove, decision by decision, that the model did not somehow make them first. Selection becomes “curation”, revision becomes “cleanup”, composition becomes “prompting” and judgment becomes nothing at all.

We normally infer agency from sustained, directed action. A person forms an intention, rejects results that do not satisfy it, changes course, repairs failures, and stops when the work matches what they were trying to make. With an LLM, the same evidence is suddenly treated as presumptively irrelevant. The human must establish their personhood inside the process before their contribution is even considered.

This is more than scepticism about a tool. It reduces someone who acted, judged, and created to the role of a delivery mechanism for software. And it does so by default - not after showing that their contribution was trivial, but before examining it at all.

Imagine applying the same reasoning to photography; A photographer takes a remarkable photograph of the Eiffel Tower.

“It isn’t yours”

Why?

“The Eiffel Tower is in the public domain, and it dominates the photograph. Your picture may also resemble someone else’s photograph”

The tower may be in the public domain. The photograph is still a photograph made by someone. Framing, position, timing, light, exposure, and composition don’t go away just because the subject itself is not protected.

A photograph of the Mona Lisa may add almost nothing and deserve no independent protection. A carefully composed photograph of a public figure walking through a crowd may belong to the photographer. A photograph of another photographer’s print may merely reproduce the earlier work.

It’s not a camera that settles the question. The public-domain status of the subject does not settle it either. You have to examine what the photographer contributed.

Same is with music.

A musician can take fifty classical pieces, all comfortably in the public domain, and arrange them into something new. The notes may come from Bach, Chopin, or composers dead for several centuries. The arrangement can still be the work of the person who selected, transformed, ordered, and combined them.

Another musician can hear a protected pop song once, reproduce it perfectly from memory, and acquire no rights to the song at all.

The first musician used only public-domain material and may own copyright in the resulting arrangement. The second copied without touching a file.

Copyright does not follow a substance through every transformation like radioactive dye. It depends on the work before us: what was contributed, what was copied, and by whom.

This is why “the model contained copyrighted material” does not resolve the authorship of its output. Neither does “the model learned from public-domain works”. Those statements may matter when investigating training or infringement, but they do not tell us who made the final composition.

More...