[croemer]

Since the blog post doesn't actually include the change to the TOS and only links to it at the very end, here it is:

> 7. You must not share projects that mostly consist of code written by "generative AI"-tools (including services such as Claude, OpenAI Codex). Such projects having an unclear copyright status (see requirements § 2 (1) 1 and § 2 (1) 3) and furthermore have little safeguards to ensure that they do not include harmful code (c.f. § 2 (1) 5).

https://codeberg.org/Codeberg/org/commit/96fac426a32d1ba91ff...

[graemep]

The justification given in the TOS is very different from those in the blog post. The TOS mentions copyright and the risk of malicious code. The blog post has a completely different list of justifications. The increased risk of malicious code only applies to unreviewed AI generated code, but the TOS does not reflect that.

The blog post is more consistent with the actual change to the TOS. It makes a lot more sense now. If the TOS are a problem then you can always use another host or self-host. Codeberg is intended for a specific type of project (open source license, public repo, encouraging collaboration etc.)

[conartist6]

I was wondering if copyright was part of it, but their blog post didn't mention copyright.

For me that would be the single biggest reason. People who vibe code don't legally own the copyright to the output (to the best of my hobbyist-non-lawyer knowledge), so they cannot themselves provide permission for others to be able to legally copy.

[petcat]

> People who vibe code don't legally own the copyright to the output

In the US at least, the PTO has not said this. They have said that artwork is not copyrightable if it was substantially "generated" with little human input. They have not said anything about software or other creative works. And certainly nothing about software that has been carefully "vibe engineered" with lots of human input, direction, and review.

It seems unlikely to me that the PTO would declare that a software project that heavily utilized an AI-based advanced autocomplete would make it not copyrightable.

Now, if you just one shot the most advanced Tetris algorithm of all time and post it on the web without any scrutiny or review, and no input, then it's probably not going to be copyrightable.

[PaulDavisThe1st]

> It seems unlikely to me that the PTO would declare that a software project that heavily utilized an AI-based advanced autocomplete would make it not copyrightable.

Why would they have different rules for one potentially copyrightable material versus another?

What the PTO (courts, actually) have not said is how much and what kinds of work a human must perform to transform a machine-generated work into a copyrightable work. It could be that this would be domain specific (i.e. different rules for software and art). But totally different rules entirely for software and art? Seems unlikely to me.

[samatman]

> Why would they have different rules for one potentially copyrightable material versus another?

There is a longstanding distinction between expression (copyrightable) and facts and ideas (not copyrightable). Nor is this merely used as a binary category: the degree to which a work rests on facts and ideas, versus expression, is very much taken into account when considering the question of infringement.

Classic cases involve works like maps and biographies, versus works of pure fiction. You can surely see how the expression of the former is constrained in ways in which the latter is not.

The algorithm to reverse a binary tree is not subject to copyright. Nor is a court likely to accept the use of "left" and "right" as variable names as evidence of infringement.

They would treat the cases differently because the cases are different.

My opinion has always been that copyright law is an awkward fit for software, and remains so. That said, the iteration and edit cycle which is normal to agent-driven software development, has a rather different character than the more normal 'workflow' for art, which is: "drawing of Asuka eating an elote in front of the Chinese Theater" then picking the one the user likes the most.

There's also the realpolitik to be considered: courts are simply unlikely to shut down the use of LLMs in software development, that horse has left the barn. There's always some margin between 'can' and 'will'.

[tingletech]

What does the patent and trademark office have to do with the copyright office at LoC?

I agree that the copyright office has not said this either -- they have just said that you need to be a human to register a copyright.

[conartist6]

Yes, my understanding is that there is some test, as yet not clearly defined, in the middle of those two extremes. At one extreme you let AI make all the choices and own nothing. At the other extreme you use AI more or less like a tab complete and own everything.

[catlifeonmars]

Legal ambiguity tends to have a chilling effect. Although I think this is overshadowed by hype at the moment.

[graemep]

> so they cannot themselves provide permission for others to be able to legally copy.

If something is not covered by copyright law then others do not need permission to copy it. However, you cannot impose conditions on copying it either (not even minimal conditions like the 2 clause BSD license).

However you can enforce licenses in jurisdictions where even pure vibe-coded LLM output is covered by copyright. So you might find yourself in a position where, for example, you can enforce the GPL on your vibe-coded software on someone in the UK, but not someone in the US.

[singpolyma3]

There is at this time no reason to believe that. We know an AI cannot own copyright in the US. We don't know that the author doesn't own it. And if they don't we don't know how much of their own manual effort on the output would make them own it.

If a photographer can own a point and click snap of a painting then I expect we'll land somewhere where LLM outputs are or can become owned by the tool weilder with at most some minimum effort.

[rnd0]

Not a lawyer, but additionally there's also legal concern about the legality of the training source. How can you verify that the model you're using wasn't trained and therefore incorporates code that is not licensed for sharing. e.g. someone theoretically uploading the shared source windows kernel to github and it being added to copilot.

There's no reasonable belief that or similar scenarios cannot happen and does not happen.

That's my problem with Linus's announcement regarding AI in the kernel -since the announcement I'm counting the days until we see SCO II Electric Boogaloo.

There's a lot of people with deep pockets and a deeper hatred of FOSS, and AI gives them the backdoor to take FOSS out -legally, on the grounds of province.

(I'm not a lawyer, this is not legal advice, and this is only my speculation from a layperson's point of view)

[EdiX]

> Not a lawyer, but additionally there's also legal concern about the legality of the training source

That's not how copyright works. If you use a pirated copy of photoshop the resulting image is not illegal.

[rnd0]

That's apples and oranges; a better analogy would be if you post a video on youtube that you made yourself, with products you're allowed to use, but happen to overlook that copywritten music is on the radio in the background or on the tv.

How many youtube videos have gone dark over bogus copyright claims over the years? And you think people will overlook what's in AI -when there's over a trillion dollars on the table?

[xyzzy_plugh]

If the resulting image contains copyrighted content, such as a snippet of photoshop's proprietary code, then it most certainly might be infringing.

It's not as simple as your reductive analogy. It's for a court to decide.

[croemer]

You don't need to own the copyright if the output is not copyrightable.

You seem to think one needs to have copyright to publish something but that's not true. If you don't infringe on anyone else's copyright there's no problem. And others can of course copy themselves.

The only problem is if someone else had copyright to the output.

[Narishma]

The blog post did mention license laundering, which is one of the copyright issues with LLMs.

[NoboruWataya]

Interesting that the TOS contains the justification for the rule. That isn't necessarily something I would expect to see in a TOS and leaves the door open for people to dispute the justification or argue "well that's not relevant to THIS vibe coded project".

Beyond the reasons stated I think there is probably a more general (albeit less tangible) benefit to having a forge with a reputation for being human-focused. Like if you see a project on Codeberg you can reasonably expect it's probably not completely machine-generated slop, which is no longer an expectation I have for GitHub. People are getting offended about it but why not just use a different forge for your vibe-coded projects?

The main risk is that it leads to loads of distracting discussions about whether projects are "too vibe-coded" or whether they fall on the right side of the line. I would assume Codeberg would be relatively light touch about enforcement in marginal cases, which equally might blunt the effectiveness of the rule but ultimately, it doesn't have to be perfect to be useful.

[inigyou]

I get that impression too. Seems like codeberg wants to be the hub for meaningful software, not a dump for any random file that happens to technically be free software.

See also: 100MB per-user quota for repos that are not meaningful software.