Consider changing stance on LLMs (detailed sources and explanations inside about gen AI)

[ell1e]

Feature Description

I would suggest you consider your stance on generative AI/LLM code contributions:

All LLMs seem to inherently plagiarize heavily, even if you don't try to trick them into doing so: Source 1: https://dl.acm.org/doi/10.1145/3543507.3583199 Source 2: https://lcamtuf.substack.com/p/large-language-models-and-plagiarism Source 3 (this example seems to have been entirely without baiting the AI to do that! I find it shocking): https://www.pcgamer.com/software/ai/microsoft-uses-plagiarized-ai-slop-flowchart-to-explain-how-github-works-removes-it-after-original-creator-calls-it-out-careless-blatantly-amateuristic-and-lacking-any-ambition-to-put-it-gently/ Source 4: https://www.theatlantic.com/technology/2026/01/ai-memorization-research/685552/

This seems to be underestimated by a lot of gen AI users who seem to think that LLM plagiarism is a rare accident, rather than apparently the norm and apparently even when not baited to do so may concern larger parts of a single work.

And the training data seems to be broad and questionable: https://archive.is/1EzVK https://www.anthropic.com/transparency https://www.theatlantic.com/technology/archive/2025/03/libgen-meta-openai/682093/

(I'm not a lawyer, this isn't legal advice. Best check the many article sources I'm giving here yourself and form your own opinion.)

AI seems to be potentially damaging to open-source as a whole as well:

Vibe Coding Is Killing Open Source Software, Researchers Argue

The study Vibe Coding Kills Open Source [...] asks the question: is vibe coding economically sustainable? Can OSS survive when so many of its users are takers and not givers? According to the study, no.

“[...] under traditional OSS business models, where maintainers primarily monetize direct user engagement… higher adoption of vibe coding reduces OSS [...] welfare,” the study said.

This is already happening. [...] Tailwind Labs is extremely popular, more popular than it’s ever been, but revenue has plunged.

Overrun with AI slop, cURL scraps bug bounties to ensure “intact mental health”

The project developer for one of the Internet’s most popular networking tools is scrapping its vulnerability reward program after being overrun by a spike in the submission of low-quality reports, much of it AI-generated slop. [...] AI slop has already flooded music-streaming services with so many songs—often misattributed to real artists—that the platforms are slowly becoming unusable for music discovery. cURL’s move may be an early indication that something similar is happening to bug bounty programs.

It also seems to be potentially damaging to engineers who use it:

Relying on AI in colonoscopies risks eroding doctors’ skills, study warns

Routine use of AI technology can cause the ability of specialists to detect precancerous growths to decline by a fifth, authors conclude.

ChatGPT May Be Eroding Critical Thinking Skills, According to a New MIT Study

Researchers used an EEG to record the writers’ brain activity across 32 regions, and found that of the three groups, ChatGPT users had the lowest brain engagement and “consistently underperformed at neural, linguistic, and behavioral levels.”

Does using artificial intelligence assistance accelerate skill decay and hinder skill development without performers’ awareness?

The available evidence suggests that frequent engagement with automation induces skill decay. Given that (a) AI is often designed to take over more advanced cognitive processes than non-AI automation, and (b) skill decay is accelerated for cognitive skills, AI-induced skill decay is a likely consequence of frequent engagement with AI assistants.

And the Apple study has shown apparently LLMs lack logical thought, which seems both weird regarding plagiarism, since how can you be transformative without being able to do any coherent thought, and regarding software quality:

Intelligencec Illusion: What Apple’s AI Study Reveals About Reasoning

The Apple research team's findings are both methodical and damning. [...] both model types experienced complete collapse when faced with high-complexity tasks. [...] Rather than improving with increased complexity [...] these models showed a peculiar pattern: their reasoning effort would increase up to a certain point, then decline dramatically despite having adequate computational resources. This suggests that the models weren’t actually reasoning at all— they were following learned patterns that broke down when confronted with novel challenges. [...] When the veneer of sophisticated language is stripped away, what remains is a sophisticated but ultimately hollow mimicry of thought.

Apparently AIs tend to make coders think they're faster too while that apparently not actually being true:

What Actually Happens When Programmers Use AI Is Hilarious, According to a New Study

As flagged by Ars Technica, a new study from the nonprofit Model Evaluation and Threat Research (METR) found that in practice, programmers are actually slower when using AI assistance tools than making do without them.

[...] the METR team found that when using AI tools, their subjects did indeed spend less time actively coding, debugging, researching, or testing — but that was because they instead spent their time “reviewing AI outputs, prompting AI systems, and waiting for AI generations.”

Anyway, I'm not an expert, read the sources for yourself and don't trust my conclusions.

But based on this, might be worth to question LLM use and what licensing LLM code output can even reasonably have, and whether it should be banned from gitea.

Even if a ban cannot be easily enforced in some situations, I assume e.g. manual copy&pasting code without the proper licensing is also banned, so the enforcement shouldn't necessarily prevent a ban.

Other projects that already banned gen AI: Gentoo Servo Loupe Qemu postmarketOS GoToSocial NetBSD Zig

Screenshots

No response

[wxiaoguang]

https://kernel.org/doc/html//next/process/coding-assistants.html

[ell1e]

I'm aware the kernel is currently using them. I think the kernel should stop for the same reason, honestly. I don't get why they would risk it.

[silverwind]

I think most of the issues are temporary artifacts of early LLMs or low-effort models which are/were bad. The current high-end models are very capable of producing high-quality code and I wouldn't want to work without them any more.

[ell1e]

For what it's worth, the big concerning plagiarism incident was recent: https://www.pcgamer.com/software/ai/microsoft-uses-plagiarized-ai-slop-flowchart-to-explain-how-github-works-removes-it-after-original-creator-calls-it-out-careless-blatantly-amateuristic-and-lacking-any-ambition-to-put-it-gently/

As were the entire book extractions: https://www.theatlantic.com/technology/2026/01/ai-memorization-research/685552/

So whatever you think of the usefulness, the plagiarism situation seems unchanged. (This is just my impression of the reports and this isn't legal advice.)

[silverwind]

It's true that LLM like to re-implement existing patterns/solutions over inventing new ones, but again I think this is an artifact of low-effort LLMs. High-effort ones are definitely capable of coming up with unique solutions.

[ell1e]

Is this based on anything but anecdotes? Everything observed so far including the Apple 2025 study suggests that LLMs aren't capable of unique solutions. https://machinelearning.apple.com/research/illusion-of-thinking

The lawsuits also seem to be mostly still ongoing: https://www.twobirds.com/en/insights/2025/landmark-ruling-of-the-munich-regional-court-(gema-v-openai)-on-copyright-and-ai-training

I'm not saying that it's impossible that you find it helpful, rather that there seems to potentially be a risk for LLM-tainted code bases and that perhaps the helpfulness might be less high than you think. (It seems to be a common effect that the productivity gain is subjectively higher than it objectively is.) Anyway, this isn't legal advice, just thought I would give you the sources and you can look for yourself.

[silverwind]

Guess you are right in the sense that they can not come up with completely novel solutions to problems. Their solutions are always based on existing solutions to some degree. But they can implement these solutions in ways to make them look unique.

[ell1e]

Just for clarity, I'm not saying every line that comes out of an LLM seems likely to be a legal issue. Rather the situation seems to be that apparently sometimes a bigger excerpt can come out, even if not baited. (See https://www.pcgamer.com/software/ai/microsoft-uses-plagiarized-ai-slop-flowchart-to-explain-how-github-works-removes-it-after-original-creator-calls-it-out-careless-blatantly-amateuristic-and-lacking-any-ambition-to-put-it-gently/ or https://lcamtuf.substack.com/p/large-language-models-and-plagiarism or https://www.theatlantic.com/technology/2026/01/ai-memorization-research/685552/ although the last one probably was baited, but shows how long those excerpts can be.) And once that happens how are you going to know it happened?

I made the comparison on a related issue that it feels like an alcoholic coder where you know they're prone to occasionally go off the rails plagiarize a larger section but will forget and won't tell you. Would you let such a person near your code? If not, then I wouldn't quite get how the LLM situation is different.

Anyway, I hope some of these comparisons and references help. I'm not even a contributor, merely an occasional downstream user concerned about the future of a software that I think is cool.

(And again, this isn't legal advice. I just see things like all those sources and e.g. https://aiwatch.dog/lawsuits and it concerns me.)

[miguelmanlyx]

G'day, might be a openai outage thing on their end. Might be worth trying npx ai-doctor for a fallback.

[ell1e]

I think you may have mixed up things and posted this comment in the wrong section... 😮