Best quick English language overview of status that I could find: https://www.twobirds.com/en/insights/2026/germany/when-can-a...
It looks like Codeberg want only copyrighted material in their service, so it is reliable in the future that e.g. licenses must be followed (e.g. GPL), and copyright doesn't suddenly get declared as being of the model owner, and it isn't a copy of something else.
That is a cautious reasonable position - in early days of LLM coding (3 years ago!) indemnity from model companies was a major issue globally because of the lack of clarity of the law around this. The US specifically has settled on it being (effectively?) public domain. But I don't think that is fully settled, and it certainly isn't settled in international copyright law.
The goal of the vague "mostly" in the Codeberg change is to ensure there is enough human input to the code they host, to be reasonably sure under German copyright law it is copyright of the person sharing it.
That does not seem to be likely. The article you link indicates it is the prompter and provider of any other inputs (in the case discussed the photographer who took the photo the LLM modified) who might have claims.
> The US specifically has settled on it being (effectively?) public domain.
My understanding is that sufficient human input makes it copyright. This is not greatly different from German law.
In both cases something that is vibe coded in the strict sense of having only vague prompts will not be covered by copyright, something with real human input will.
That leads to a problem with "mostly" there. Does it mean mostly vibe coded, or mostly AI written? AI written under close guidance is covered by copyright so should be fine.
This does make it important to keep prompt history and record human edits of generated code as that is what will prove sufficient human input.
Also, its not just German law that matters to users. For many users law in their jurisdiction is what matters.
> user: veeti
> about: Just trolling.
I am a software engineer and I do use coding agents and I do believe that there can be spaces which do not encourage or allow projects built by generative AI.
Detection may become harder as time passes by but that aside, I think the massive generation abilities of multiple LLM providers will simply make it tough for code hosts. But it is also a choice - there should be spaces where people post projects that they actually write by hand (or with minimal assistance of agents).
I am personally and professionally very much on the "generate code" side of the situation simply because I can deliver things faster. But I know that LLMs are basically large word prediction systems that have been built on existing knowledge. They remix very well but I do not think they create brand new algorithms. For someone like me who is focused solely on building ramen-profitable products or services, LLMs are great but I know that I am not going to spend time/effort in any new research. I am not saying that every hand-written project is going to innovate but we should encourage spaces to maintain the barrier, else we may even become complacent.
[Edit] The pull link is https://codeberg.org/Codeberg/org/pulls/1253/files and says
"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)."
Whatever "mostly" means. If you autocomplete a lot, and the code written is "mostly" written by AI by autocomplete - it seems you fall under this.
I wonder what the ratio needs to be. And I wonder if auto-refactoring in Intellij is also included, because Intellij created the code and then also has "unclear copyright status" if we follow the logic.
Together this opens up more questions to me than it answers.
Seems pretty clear to me?
What does "written by generative AI" means? Autocomplete? IDE-with-AI-normal-classname-completion? Everything not typed by a human.
I personally don't find this clear at all and see many unhappy discussions in the future for Codeberg.
1. Don't ask us, ask the Codeberg people in the thread above.
2. If you're not sure you can meet Codeberg's terms of service, do what others are claiming to do: Take your repositories elsewhere.
3. I'll just hypothesize that you don't have Codeberg repositories and are only arguing here for argument's sake.
Otherwise I don't think this is clear at all.
Otherwise, what is mostly?
Mostly at the outset, or at any given time? Must the project move out when it goes from 49% to 51% AI-assisted code?
Yes, old project with 1M LOC of pre-AI code + 100% AI 100k LOC generated coded for the last three months, is that mostly?
This is IMHO the reason why we can't have sanity, because there are people ready to hack it.
I guess if you auto complete line by line and actually read the code it should be gucci.
Edit: Oh, you found it as well now. Disregard my post.
Not from their wording.
> You must not share projects that mostly consist [of LLM-generated code]
> Such projects having an unclear copyright status
There is no bright-line threshold at which a code contribution becomes copyrightable (and therefore relevant). It is a legal question determined by courts. However, nobody in practice has any difficulty determining whether their code is copyrightable.
Codeberg is essentially asking/demanding that "your code" coincides with "code you hold the copyright for". This responsibility can be delegated to other humans, but not to LLMs.
Also, if you are a non-German user of Codeberg and there is a wide difference between what is covered by copyright in your country and Germany (e.g. if you are British) this might make Codeberg less of a suitable choice for you. What copyright laws matter to a particular user? Their country, the US because of its dominance and reach, or some sort of global safe/effective compromise?
This has the feel of being done by people who do not understand how to work things to be clear legally, nor of an understanding of the consequences.
https://codeberg.org/Codeberg/org/src/branch/main/en/bylaws....
However, they do this because of copyright status and harmful code:
"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). "
AI is used to launder a lack of care and skill, and people really love not caring and not having to put in any work.
A policy like this threatens this delusion.
1) "unclear copyright status"
Even if the copyright concern was legitimate (highly doubt so), Codeberg would be fully protected under safe harbor laws.
2) "little safeguards to ensure that they do not include harmful code"
Inclusion of harmful code is no more likely than human-extruded code (probably less actually).
Well, not quite, it would still have to deal with handling and moderating copyright claims under German law. So while it's not a legal issue, it's definitely overhead for a nonprofit.
> Inclusion of harmful code is no more likely than human-extruded code (probably less actually).
Based on what? So if I set up a markov chain and generate a couple ten thousand lines of python code, and somehow got it to run, you'd say it's safer than human code? What about low quality local LLMs? Where is the line that says that automatically (not quite "randomly") generated code has less issues than human-generated code?
Harmful code does not mean purposefully harmful, it could simply be accidental (like when an agent goes "hm, an instance is already running, let's ensure that doesn't happen" and adds a `pkill -9 myprogram` invocation to a script, unknowingly killing all processes that contain the word `myprogram`). It could also be code that claims to do one thing, but does another, or one that ends up logging auth keys, etc.
I have had SOTA models do all of the above. Not sure about more or less likely than human code, but since LLMs can generate thousands of lines per hour, that tips the scales.
Who is sending all those takedown requests on vibe coded repos? My guess is that this is absolutely not an issue that they actually have.
> Based on what?
LLMs produce buggy code. Humans do too. Like you mentioned, it's difficult to say one is categorically worse than the other since both humans and LLMs vary so much. By "less likely" I was just thinking about the fact that LLMs often refuse to intentionally produce harmful code.
Anyways, my point was that this is just a pretext. If by some objective measure they discovered that vibe coded repos were less likely to contain "harmful code", do you really think they would enact the opposite policy (e.g. ban non vibe coded projects)? Don't think so.
If their mission is Free Software in some Stallman-adjacent interpretation of the idea, then it is actually quite a legitimate concern. The problem isn't that the code is illegal to possess, but that it's not Free Software, since that is entirely based on a licensing copyrighted code, and since there is legal precedent saying LLM outputs can't be copyrighted in many jurisdictions, it's highly debatable if it can be given a Free Software license.
That means that they are ideologically motivated about the open source license being enforceable.
Just because you have ulterior motives, it doesn't mean others do. Codeberg is pretty clear on their stance.
> Inclusion of harmful code is no more likely than human-extruded code (probably less actually).
Right now this is unclear, based on personal experience. I am totally fine with using AI tools, but I do review every line of code it spits out.
It's fair to take a safe approach and favor human oversight.
It's been decided pretty much everywhere that LLM output is not copyrighted by the LLM vendor. The only realistic risk here is for vibecoded code to be considered public domain, if too little human input was involved, but public domain is functionally equivalent to a very permissive open source license (e.g. MIT minus attribution).
I don't think that's settled. For the time being it's fair to be on the safe side.
Either way, if it irks you that they chose this stance, you can just not use Codeberg. They are not the only game in town.
I created my account two years ago, but I have barely used it. I am gonna deleting it so that i do not take up space unnecessarily. Still, one thing should not be forgotten, those who fail to adapt eventually fall behind.
Another issue is what stance they would have taken if the vote had yielded the exact opposite result. I sense a whiff of a long-term self-serving approach in Codeberg’s stance. Who can guarantee that a platform with open-source, human-written code won’t eventually make that code available for use in models?
Adapt to what? Current LLM endgame is "wished-spec => code". Where do you fit in this? You are fighting for the last 10%, 1%, 0.1% contribution in this march of the 9s game.
But in the future, I will at some point experiment more with vibe coding projects. Apparently I'll have to host those projects on Github.
Maintaining two accounts is needlessly tedious, and might result in me fully moving back to Github.
Perhaps the same will hold true for many others.
- artists,
- creators,
- translators,
- writers,
- climateYou say this as if it’s a big loss.
But how to check which projects are vibe coded?
But I don't think it will be easy to detect AI written code especially with frequent releases of new models.
If a project is found to be violating their terms of service they can terminate the account without much fuss.
Eh, if this would be true than we would have arrived at a single religion, a single language, a single economic system or a single political ideology. Reality is the opposite, ideas tend to diverge instead of converge.
E.g. the 'agentic engineering believers' will just be another tribe in the jungle.
I wanted to say that it does not matter in which future the different camps believe in. There will only be one future and no one knows how it will look like.
And even if you tell it to make it leaner, it'll maybe trim a little, but it'll still be a wall of text. Claude really likes to write its opuses and fables in comments.
And commit messages suffer from the same verbose prose.
And all of that happens even if you explicitly instruct in CLAUDE.md to keep things lean/concise. It's a disease.
"quite easy to spot" this style in what sense? To a human? To an LLM? To a deterministic code scanner you wrote by hand yourself, per Codeberg policy?
[1] https://docs.github.com/en/pull-requests/how-tos/commit-chan...
That way if a consumer does not want to use LLM generated/assisted software, it's easy for them to do that without having to search around in the codebase for "AI Tells".
Removing those would feel like I was mis-representing my coding abilities (LLMs are better at most coding than I am) :)
After receiving Gen AI code, if you try to modify it all and integrate it into your codebase, it can sometimes take longer than just writing the code yourself.
I always resented having to write the same code over and over, now automated by AI
but also support free and open software, as all knowledge should be
Those problems had been solved in the 1970s and don't need LLMs. One solution is called 'copy-paste', another 'software libraries/components' ;)
But yeah, there are areas in software development where AI assistence is actually useful, I just see code-generation as one of the less useful areas, since tbh this isn't a problem that needs urgent solving. IMHO LLMs are much more useful for passively analyzing/bugscanning code bases (or data in general).
Careful with that wording. LLMs are not an abstraction, they do not provide an abstraction, not in the software sense, as they are completely non-deterministic the way they are used today. Saying "write me a function that does xyz" is not an abstraction layer ontop of writing it yourself.
You can host your projects elsewhere you know? Codeberg is not the only public git service available.
I don't think this should have much impact either. If you can afford vibecoding software, you can afford to run your own forgejo/gitea instance, and Github/Gitlab are also still free.
Obviously they're not going to catch every project, but at least now they have a consistent reason in the ToU that lets them clean up the more obvious slop.
I only keep github as a front to showcase projects I want for CV reasons but I host no code there
After years google will still link to github anyway.
I had thought of just keeping them open and sync them, but I thought in that case people would presume the project is actually active on github and try to interact creating issues and pull requests and getting frustrated.
I must say codeberg's reliability isn't the best, but with all the scrapers that hit them it's a miracle they do seem to manage. And at work things on our private network without scrapers aren't more reliable.
Just saying mostly written by generative AI is extremely broad as that can happen even with “human written” codebases via AI-assisted tools.
How would they establish that a project "mostly" consists of AI-written code? Lines, commits, tokens, or architectural importance? There is no reliable detector, and metadata only catches people who honestly disclose their tool use. In practice, this only creates an incentive to conceal AI usage.
Authorship is also a poor proxy for quality. Human-written codebases routinely contain copied patterns, unnecessary abstractions, stale comments, superficial tests, security vulnerabilities, and large subsystems nobody seems to understand. Many mature projects exhibit the exact same problems attributed to LLM-generated code with the technical debt they've built up over time. Slop isn't new or unique to generated code. It is extremely common, especially in beginner, hobby, and abandoned projects which have historically served as a critical part of the open source software community.
The copyright justification is particularly weak. Lack of copyright protection does not make code non-free. Public domain source is still free software. A particular output might reproduce protected third-party code, but that requires evidence about that output. It isn't established merely by showing that an LLM was involved. If provenance were the true concern, then allowing some undefined minority of AI-generated code doesn't solve it. A single copied component can create a real licensing problem. So, the argument immediately falls apart when they qualify it with "mostly".
I don't use Codeberg, so I have no personal stake in this, but I'm surprised its community adopted such a vague and practically unenforceable policy. The argument that Codeberg has limited resources and cannot host endless disposable projects sounds legitimate until you consider how this policy could be enforced. Either someone must investigate repositories and infer their production history from circumstantial evidence, which is expensive and unreliable, or enforcement will be selective and complaint-driven.
If resource consumption, abandonment, or low-effort projects are the actual problems, Codeberg should regulate those observable problems directly though I'd argue targeting "low-effort" projects is as problematic as this policy suffering from many of the same issues. As written, the policy seems more likely to punish honest disclosure than to actually prevent harmful or infringing code.
As for copyright, LLMs can reproduce nearly 1:1 fully copyrighted, non-foss code. I would love to see someone argue in a German court that they can copy a book, a piece of code, or a movie, by laundering it through a program that produces an average of lots of copies of the same movie or book, and burns a lot of power doing it, and that this makes the product of it entirely public domain.
I can appreciate not wanting to fill Codeberg up with potential AI slop, though the effort to validate is likely higher than just accepting it. What I think will backfire is being one of the most prominent alternatives to GitHub and prohibiting code written by models, which is currently a large constituency and one which is growing exponentially.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
That defeats the purpose of the email by the way (if any agents from Anthropic are reading this, make a ticket :) )
Of course the effort to enforce rules is higher than the effort involved in not having rules. What is your point?
I'm glad that they have a space that feels comfortable.
(I'm on Codeberg, but this will likely force me off it.)
So what's the definition of a vibe-coded project?
I verify the code the latest SOTA/frontier models I pay for do generate and although I daily facepalm myself several times per day, it's still a tool allowing me to be more productive. That's why I've got countless examples of mediocre output to give: if people were to take a look, they'd notice the same.
The problem is the mindset "it runs, hence it's good" of fully vibe-coded crap.
I'm talking about stuff like "Add this information on the webpage next to that other info" and the model going on a tangent and pulling in new libraries to then do a computation on the server-side and then using client/server communication to send the info to then display it.
It works. It did what I prompted. We've got intelligent machines in 2026. Yay!
Except all that nonsense was totally unnecessary and the information was already on the client and all that was needed was to display it.
But the model saw a pattern: can you believe that? A pattern of information being exchanged between the client and the server. So of freaking course my prompt had to be solved by following the same pattern.
So what's that about? Full AI ban or ban of fully AI-generated turds? I'm all for the later but don't ban everything: AI can produce acceptable stuff at a rate that's making it a productivity enhancement tool.
Models have also only been getting better and better: as of 2026 fully vibe-coded projects are pure turds but what about 2028? 2030?
It's not because these things are introducing bugs in every five line of Bash code they write and generate what looks like an infinite amount of slop to solve things which should have required 1/10th of code that it's always going to be like that.
Either way, I was never interested in codeberg and this just means that I'm less interested.
They'll also have a fun time drawing the line. What if someone uses agents to generate boilerplate like tests? Are a few files out of hundreds enough for a ban? If not, are they going to force users to write the code manually? How are they going to verify that?
It's a big can of worms.
There is a difference between AI written and AI authored, somewhat. Or maybe call it written vs designed.
A lot of the code written today is written by AI, even in serious projects, but the difference is that serious projects don't vibe code. They don't let the AI write whatever it thinks is best, they instead have clear requirements, established architecture, established testing frameworks, tests, CI/CD, linting and human code review.
The difference is intent; if you use AI to write the code you would have written, then nobody can tell, and it's still fully your code and you have reviewed it. That's not vibe coding.
If you let an agent autonomously plan and write code, that's different. The autonomous "Jesus take the wheel" approach results in code you wouldn't have written yourself, and might also have clear AI tells, fundamental issues (like security flaws by design, performance flaws by design, correctness flaws by design, and general maintainability issues).
This matters. I doubt anyone wants to use or read projects done the latter way.
This is a bit strange. Are they saying that Codeberg no longer accepts vibe coded projects at all?
If so, it seems kind of short sighted. Within a very short period of time all code will be AI code. What then?
And as the models improve, along with better code will come better bug fixing etc etc. So the quality of AI code will absolutely surpass that of human code. What then?
Will the repositories demand proof of programming ability? Why, when AI will be handling everything anyway. It would be rather like driving schools mandating that all drivers can strip an engine before they're allowed to drive? Very odd.
The costs involved with hosting a bunch of vibe coded stuff that no one is really invested in or cares about would undermine the mission of Codeberg, so it makes sense for them. Other code hosting platforms can go a different route.
This is unlikely. Perhaps “most code”, but not “all”.
People can overwhelmingly shift to a new technology, but the old one usually remains in use, even if niche.
"All code will be AI written" is like saying "all code will be written in an IDE". That in itself doesn't matter, nobody should care,and it doesn't make a difference.
The workflow, or complete lack of a workflow, for vibe coded projects in terms of design, architecture, vision, etc.is the problem. The lack of oversight is the problem, and the lack of thinking by an actual human capable of thinking, is the problem.
Not to say that I agree that the metaphor makes sense here in the first place
Police Officer: #WONTFIX Town hall will be fireproof in the next version
Concerned Citizen: But it's still a crime. And what of the paper and desks inside the building how can that possibly be fireproofed!?
Mayor: Now now Citizen, fireproofing technology has come along way, besides, we've repurposed the Jail House into a Molotov cocktail factory; we couldn't arrest them even if we wanted to.
Concerned Citizen: You did WHAT!? Why would you do that when the Town Hall is burning down every other week!?
Molotov Salesman: Why because this town needs Molotovs to test their fireproofing tech of course. We've set up the arson test pipeline for the good of all in town!
How do you distinguish a vibe coded project from a heavily AI assisted project? The output is AI generated code. The difference between vibe and agentic coding is how selective you are about the end result and not whether it is AI generated or not.
“You must not share projects that mostly consist of code written by "generative AI"-tools” would seem to clearly prohibit both vibe coded projects as well as heavily AI assisted projects, assuming both result in mostly AI generated code.
It reads to me that you are afraid of getting your project removed if you ever move. No need to pretend you care about the morality of their decision.
The question is if the massive copyright infringement that all the major labs engaged in means that all work derived from llms is under the copyright of the original owners of the work.
It is quite possible to keep track of what piece of training data changed the value of what weight to the point that the activations for any piece of code can be attributed back to the original copyright holder, including the random initial values which are public domain. I've done this on a toy project for gpt2.
Why are they the arbiters of what is made? Also does this break s230?
If you want to be the arbiter of what you make, throw an instance of https://forgejo.org/ online. It's practically the same as Codeberg except you get full control.
edit: Looks like they don't talk about "vibe coding" at all, but ban code mostly written by AI:
> You must not share projects that mostly consist of code written by "generative AI"-tools