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.
> copyright doesn't suddenly get declared as being of the model owner
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.
Given a German court talked about not having the prompt logs making it hard in a specific case to prove it had sufficient human input, maybe someone could do an open source hosting service where the Claude Code logs were fully uploaded with each commit, so that could be later proven? It would be cool anyway to have such a service, so people could learn how to do LLM coding better from the examples.
Once you look at how German anti-nuclear lobbying set Europe back decades on its energy independence and made us dependent on authoritarian hellholes of the world, it all makes sense. They just want to do the same with AI.
It’s probably pointless to argue, but there was never a lot of nuclear power in Germany to begin with. This is just turning energy policy into culture war.
Not a lot is inaccurate given the carbon intensity of coal and its importance in Germany's energy mix. Anyway, I think the nuclear debate is quite off-topic here.
Open source code that is AI authored is worthless. If I need an LLM authored, idk, terminal emulator, I will simply ask my AI or use one of the tens of thousands out there.
I use Linux every day, where AI coauthored analysis and code has already contributed to harden its security. Like it or not, any project that categorically rejects AI contributions or engages in vague activist type purity spiral testing over "vibe coding" seems like a security liability at best.
It depends what you mean by "AI authored". If you mean one-shotted vide code, then yes. If you mean someone providing detailed instructions for architecture and functionality and reviewing the code, then it does have a lot of value.
"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.
It means if the guy running Codeberg doesn't like it. This is his fiefdom, not a public square. Make your own fiefdom for AI coding. You could call it Vibeberg.
> I personally don't find this clear at all and see many unhappy discussions in the future for Codeberg.
You don't understand their goal in doing this?
Frequently (much more than we'd like to admit) actual contracts leave loopholes in, said loopholes which go against the spirit of the contract. It is not unusual to have a concrete contract that allows more (or less) than the spirit the contract was signed in.
Rather than nitpicking the contract (the TOU), why do you think they need this updated contract in the first place?
I agree that they have different connotations. However, in the absence of other information, I don't see how we could ever collectively agree on a better value. The least subjective option, as far as I'm concerned, is "more than 50%". If a is 50.001 and b is 49.999, then a is the one with the most, not b.
That's why terms and laws are written vaguely. If you're obviously on one side of the line then you're obviously on one side of the line. If you're straddling the line, your punishment depends upon how the judge feels that day. So it would be wise not to risk it.
I don't think this is a bad-faith response. Someone might fully retain copyright while using heavy autocomplete. I don't think such a person would be banned by the codeberg policy. This is at least a little bit unintuitive.
They could have tagged these projects with "AI" or "MixedAI", and they'd be removed from the default views. It should also make clear that such projects may have a dubious legal status. Though within businesses no one cares?
If this opens more questions than it answers, then you are simply discovering that you don't understand copyright (which is ok, but not the fault of the Codeberg policy).
> 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.
That is not what their wording says. They could have said "code that is mostly covered by copyright under German law", but they chose to say 'mostly consist of code written by "generative AI"-tools'.
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.
This is good and I am saying this as someone using coding agents full on.
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.
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). "
This makes sense from a copyright perspective -- as long as you maintain 50% or more actual human authorship, in case of a conflict with future copyright laws around LLM generated code, you can at least claim that a majority of the code is human-authored.
They are unable to be more clear, because the law is unclear. The safe assumption would be "if the code would be copyrightable, had it been written by a human, then it must have been written by a human". It is impossible to be more clear than this, because "copyrightable" is a legal term of art decided by judges.
Unsurprising given that a large amount of posts on the front page are about or in support of projects that are entirely vibecoded with near zero human input beyond "hm it feels kind of slow, make faster?".
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.
Why are all the vibe coders so upset? Places that have mostly non-polluted projects will be fantastic for future model training. Codeberg doing this will mean that it will be scraped and part of all future models' learning input, surely? They have never cared and will never care about anyone's rights.
Because it's either plainly stupid virtue signalling aimed at the current strawman (AI bad!) or- if the legal justification stands- it's another symptom of European death by over-regulation, which should be judged as tragic.
I think AI has enabled a sort of inverse impostor syndrome, where many charlatans now beleive they are peers with people who actually understand and care about programming.
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.
> 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.
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.
> Even if the copyright concern was legitimate (highly doubt so), Codeberg would be fully protected under safe harbor laws.
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.
Agreed that it would potentially be a bigger concern but it seems they support both "Free Software" and "Open Source". See my other reply here: https://news.ycombinator.com/item?id=49004912
I don't think they have ulterior motives because I do, I think that because of the loaded language they use (e.g. "taking a stand against AI", "LLM-extrusions") and the dubious reasons they came up with.
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).
> 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.
Codeberg was never really intended as a place for small scale or one person teams to host vibe coded projects. It is more of a platform for opensource projects, medium and large teams, and people who take this work seriously. Many users also moved there after Github began pushing AI features so aggressively.
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?
>Still, one thing should not be forgotten, those who fail to adapt eventually fall behind.
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.
In a way it make sense. I don't think they have the capability/money to scale out the system to support a plethora of vibe coded project like Github does.
The copyright/Anti-ai angle is likely an excuse and they're running
out of money for hosting, hence errors and anti-scraping blocks.
Its completely absurd to think they somehow waited years to ban AI for copyright, while deleting
anything they deem as copyright infringing.
Looks like those vibe-powered opensource projects will mighrate
to somewhere else.
Okay, lets imagine it get enforced and every project is
scanned for compliance, with some false positives
and inability to prevent people from sneaking in AI code.\
Do you agree with it now?
vibe-coders should be pretty happy about the fact that places like this exist, given that's it's ultimately thanks to non-vibe-coded OSS that LLM can improve.
I've actually been thinking it'd be nice to have something like GitHub, but only for hand-written code. If anything, for now I'd like to learn primarily from and perhaps contribute to projects that have human involved at all stages.
But I don't think it will be easy to detect AI written code especially with frequent releases of new models.
It does and it's unfortunate. Perhaps Codeberg might try to block AI crawlers access to the entire website. I know it's all jazz nowadays, but I'm not in favor of using other's people work to train AI without explicit consent.
Claude-written code is quite easy to spot - because it has this particular style of overly verbose "walls of text" comments, that 1. repeat verbatim what the code just below does, 2. include "list of things" that quickly go stale, 3. talk about "how things used to be before we changed it here" (useless to anyone reading the code now), 4. "talk to reviewer" in comments. And many other comment sins.
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.
Claude-written code is quite easy to spot - because it has this particular style of overly verbose "walls of text" comments, that 1. repeat verbatim what the code just below does, 2. include "list of things" that quickly go stale, 3. talk about "how things used to be before we changed it here" (useless to anyone reading the code now), 4. "talk to reviewer" in comments. And many other comment sins.
"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?
They could, for example:
1. Rely on users flagging suspected repos (with human moderators reviewing afterwards).
2. Deploying AI agents to evaluate repos against known AI-generated-code patterns (also with human moderators afterwards). Of course, the irony of the situation - using slop machines to fight slop.
Most people leave the “Co-authored-by: <name>” lines in their Git commits [1]. I’m not sure why exactly, but I suppose it's either laziness or a lack of care. Regardless, if I were Codeberg, that would be the first thing I’d detect, and then I might use a Random Forest Classifier to identify the rest.
As other have mentioned, and I'm the same. Leaving "Co-Authored by:" feels like open disclosure of how the project was created.
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) :)
I feel it's a faux pas to pass off work entirely or partially done with AI as your own. I make sure I leave those lines in specifically for that reason.
AGENTS.MD and CLAUDE.MD also provide easy detection avenues. Trivial measures will probably catch 70-80% or more. You don't need a 100% enforcement rate for a rule to exist, indeed there are basically no rules anywhere in the world that achieve that.
At that point just use whitelist-based approach. I already used gitignore like that. It also easier to avoid committing temp file and secret through whitelist because those file are never in gitignore in the first place.
It does not prove anything. Some could not commit those files to avoid detection, and those files exist in projects in which only some code is AI generated. I have them in projects where, for example, LLMs were used to generate tests and for a few fixes/additions in the rest of the code.
Whilst they may provide a strong signal, I wouldn't call it a guarantee. If I wanted to avoid vibe-coded contributions I might add such files, but with instructions for the agents that they're not allowed to read or touch any of the files in the repository.
It's rare to see terms of use that are all actively enforced. I think it makes sense to state the outlines to justify e.g. banning a user, or in potential legal matters (IANAL, just assuming).
I'd guess that it only matters once shit hits the fan (e.g. a copyright dispute or license violation). E.g. when the project owner claims that he wasn't aware of the violation because LLMs wrote all the code, Codeberg can simply point to the terms of use before deleting the project, which from their point of view is a quick and clean solution to the problem.
It seems the conflict between LLM/AI and programmers is getting hotter and hotter by the week. I wonder where all of this will lead. Our times are uncertain. Many people believe with certainity the future. But only one will come true.
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 know what you mean, but its not about if a certain ideology is right. Its about reality. Reality will not diverge it will stay as one. Like the future it cannot diverge there will only be one future. I didnt want to imply that only one purposed future can be true.
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.
I would argue that reality is what those different tribes perceive as "the truth", and from that pov there can indeed be different realities existing side by side ;) I wish that everybody could agree on the same objective truth, but that seems to a pipe dream, especially in the "information age".
I'm tired of SlopHub but at the same time how do you realistically ban vibe coding? This will cause a bunch of infighting with people accusing each other.
Not sure if a non-deterministic copy-paste is really filling a gap ;)
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.
Mmm, actually the composer is the one that creates music.
A director simply translates music to something human can experience and enjoy, similar to a compiler.
Get your slop machine to write you better analogies.
I understand that as a non-profit, they don't accept vibe code, and there are also server load issues. But once you start using Gen AI code, as you'd know if you've tried it, you can't stop the entire codebase from becoming Gen AI generated, because the machine's abstraction and my abstraction are different.
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'm surprised it took them this long. Good on them.
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 marked all my project on github as closed and added a link to the codeberg one.
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.
All my active repos have both codeberg and github remotes now. Unfortunately codeberg can be very slow or even impossible to push/pull from sometimes which is the only reason I’m still maintaining GitHub repos.
This seems like a wild basis for a hosting policy. It doesn't ban code shown to be infringing, malicious, insecure, or unmaintained. It bans code based on how Codeberg believes it was produced or some individuals at Codeberg believe it was produced.
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.
Maybe this is a fundamental difference in how our societies work, between Germany and the USA. Here, when we go swimming, we leave our phone, keys, and wallets at the beach and just go swimming. It's a trust system, my neighbor won't take my shit because I won't take his shit. So maybe the idea behind the system is to simply ask people who want to upload their slop to not upload their slop, and that that has enough of an impact to be worth doing already.
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.
How do you define vibecoding? Is using LLM-assisted autocomplete vibecoding? Is editing single files via an LLM vibecoding? When do we cross the threshold?
Just saying mostly written by generative AI is extremely broad as that can happen even with “human written” codebases via AI-assisted tools.
Is this "don't vibe" code, or "don't accept" code written with agents? We have a philosophical (and potentially legal) aversion to code written by models.
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>
My hypothesis is that this particular kind of AI pushback originates from people who really are terminally online and can't separate their job (programming) from free time (also programming). Programmers are unusual in that regard, compared to most other professions.
I'm all for banning vibe-coded projects because they are pure turds but I'm not for banning a codebase where AI is used to make the codebase better.
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.
Well, that's their prerogative, I guess. But the burden of proof will be on them, of course. They can't expect users to prove a negative. There are simple tells that would reveal a lot of projects. Claude Code's moronic defaults of littering commit messages with both a CC author trailer as well as a session link will reveal some, but those are easy to disable. Then there is stuff like prolific comments, em-dashes in comments, etc.. but none of those are real proof that a project was built with agents.
Either way, I was never interested in codeberg and this just means that I'm less interested.
That legalistic "I can do whatever I want as long as it's not breaking the law" argumnet lasts until the first time they ban wrongfully ban someone for it and get huge negative PR
Sure. And I'm sure that their users will take kindly to being banned with prejudice and without recourse. By adopting this policy they are committing to policing it and acting as judges. And at some point someone will be banned that shouldn't have and it will go on social media and so on and so forth.
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?
Are they also banning projects whose code was written with AI or just vibe coded projects? Roughly every serious codebase now and forevermore will be at least mostly AI written.
> Roughly every serious codebase now and forevermore will be at least mostly AI written.
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.
It's not strange to ban chess computers at a chess tournament, and it's equally not strange to ban AI code on a code hosting platform.
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.
"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.
I suppose, if your completely insane fantasy ever becomes reality, they could simply reverse the policy. Codeberg currently lives in the real world, not your fantasy world, so it makes sense to determine policy based on the state of the real world.
Concerned Citizen: Someone keeps lighting the town hall on fire, can someone arrest them!
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!
I was kind of thinking of switching to codeberg, but now I am questioning their decisions here.
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.
In Codeberg's case, they have chosen not to distinguish between those. Their bright line is "mostly", as in, if your project "mostly" consists of LLM output then it is not allowed. The definition of "mostly" is, of course, vague, and people are going to argue about where the line should be drawn. But there's at least something vaguely approaching an objective measure: they might look through commits and count how many have an LLM listed as a co-author, or they might count the number of lines of code changed by those commits. Who knows, at this point, what they will end up doing. But they're not trying to judge "how well did you curate the output of the LLM"? They have, instead, decided to judge "how many lines/commits/something created by LLMs have ended up in your code?"
“Vibe coded” does not appear in the patch to the terms of use.
“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.
This appears to be mostly due to fringe / activist views about copyright, rather than anything to do with quality or principle. If it was the latter I could get on board, as in instituting some standards against slop. But in reality it’s just letting activists hijack with their agenda.
All vibe coded projects are in the public domain by definition.
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.
The discussion doesn't appear to have been in public at all; all there is on the linked issue is a bunch of first-time posters who wouldn't have been able to vote.
I don't think American laws such as s230 apply here, as Codeberg is German.
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.
It's absolutely hilarious that vibe coders see themselves as some sort of protected class who are being discriminated against and think they can use the law to force an unrelated third-party to host their files for free. AI psychosis is a trip.
When from what I have understood they can easily deploy fully featured alternative in afternoon or less. Just take a single weekend and they have as good place to host their code ready, tested and secure. So should not in anyway be a big deal for them.
How do they define "vibe coded"? I believe "vibe coded" was originally defined as letting AI write code without human review. But I often see it used for all AI assisted code.
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
They are free to limit what they allow in their service and we are free to host our own code:
- Forgejo: https://forgejo.org/
- Gitea: https://gitea.com/
- GitLab Community Edition: https://about.gitlab.com/
- OneDev: https://onedev.io/
- Gogs: https://gogs.io/
- Sourcehut: https://sourcehut.org/
- GitBucket: https://gitbucket.github.io/
Germany copyright law feels like the relevant thing here - Codeberg is a non-profit based in Germany.
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.
> copyright doesn't suddenly get declared as being of the model owner
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.
Given a German court talked about not having the prompt logs making it hard in a specific case to prove it had sufficient human input, maybe someone could do an open source hosting service where the Claude Code logs were fully uploaded with each commit, so that could be later proven? It would be cool anyway to have such a service, so people could learn how to do LLM coding better from the examples.
Once you look at how German anti-nuclear lobbying set Europe back decades on its energy independence and made us dependent on authoritarian hellholes of the world, it all makes sense. They just want to do the same with AI.
It’s probably pointless to argue, but there was never a lot of nuclear power in Germany to begin with. This is just turning energy policy into culture war.
Not a lot is inaccurate given the carbon intensity of coal and its importance in Germany's energy mix. Anyway, I think the nuclear debate is quite off-topic here.
Yes, when you spend decades arguing "nuclear bad" (or "AI bad"), it turns out you get left in the past as Jurassic Park.
Jurassic Park is science fiction, though.
Unfortunately the truth is far worse, and some of us have to share a union with these fools ;)
Open source code that is AI authored is worthless. If I need an LLM authored, idk, terminal emulator, I will simply ask my AI or use one of the tens of thousands out there.
I use Linux every day, where AI coauthored analysis and code has already contributed to harden its security. Like it or not, any project that categorically rejects AI contributions or engages in vague activist type purity spiral testing over "vibe coding" seems like a security liability at best.
It depends what you mean by "AI authored". If you mean one-shotted vide code, then yes. If you mean someone providing detailed instructions for architecture and functionality and reviewing the code, then it does have a lot of value.
What is a vibe coded project? Where does it start? Cursor autocomplete? One shot Github project copies?
[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.
“You must not share projects that mostly consist of code written by "generative AI"-tools (including services such as Claude, OpenAI Codex).”
Seems pretty clear to me?
What is mostly? >50%? >75%?
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.
It means if the guy running Codeberg doesn't like it. This is his fiefdom, not a public square. Make your own fiefdom for AI coding. You could call it Vibeberg.
> I personally don't find this clear at all and see many unhappy discussions in the future for Codeberg.
You don't understand their goal in doing this?
Frequently (much more than we'd like to admit) actual contracts leave loopholes in, said loopholes which go against the spirit of the contract. It is not unusual to have a concrete contract that allows more (or less) than the spirit the contract was signed in.
Rather than nitpicking the contract (the TOU), why do you think they need this updated contract in the first place?
I think it's reasonable to assume >50% by default, unless the source clarifies what they mean by "mostly".
>50% of words? Characters? LoC? Files?
Yeah, this is the bit where it really falls down, regardless of how you define "mostly"!
To me majority != mostly.
I agree that they have different connotations. However, in the absence of other information, I don't see how we could ever collectively agree on a better value. The least subjective option, as far as I'm concerned, is "more than 50%". If a is 50.001 and b is 49.999, then a is the one with the most, not b.
It's reasonable not to assume what they meant. They have a clear stance on the subject instead of letting us guess
> What is mostly? >50%? >75%?
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.
If 'mostly' means over 50% you could call this clear.
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?
"Mostly at any given time?"
Yes, old project with 1M LOC of pre-AI code + 100% AI 100k LOC generated coded for the last three months, is that mostly?
Noone can discuss in good faith if the argument is: "If generated in bulk by vibe code" >> "what if I generate word by word with autocomplete".
This is IMHO the reason why we can't have sanity, because there are people ready to hack it.
That's why terms and laws are written vaguely. If you're obviously on one side of the line then you're obviously on one side of the line. If you're straddling the line, your punishment depends upon how the judge feels that day. So it would be wise not to risk it.
I don't think this is a bad-faith response. Someone might fully retain copyright while using heavy autocomplete. I don't think such a person would be banned by the codeberg policy. This is at least a little bit unintuitive.
If you have to ask then the answer is no.
I mean Codeberg will pretty soon start having to ask. How do they determine that and who'll be in charge to find violations there?
It isn't a public service. It is one person offering his resources to help FOSS. He decides if you align with his mission.
They could have tagged these projects with "AI" or "MixedAI", and they'd be removed from the default views. It should also make clear that such projects may have a dubious legal status. Though within businesses no one cares?
> projects that mostly consist of code written by "generative AI"-tools
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.
"I guess if you auto complete line by line and actually read the code it should be gucci."
Not from their wording.
If this opens more questions than it answers, then you are simply discovering that you don't understand copyright (which is ok, but not the fault of the Codeberg policy).
> 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.
That is not what their wording says. They could have said "code that is mostly covered by copyright under German law", but they chose to say 'mostly consist of code written by "generative AI"-tools'.
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.
This is good and I am saying this as someone using coding agents full on.
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.
They are a non-profit with a mission to support free software:
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). "
So 'mostly' is the clearest they get here. Meaning more than 50% of lines of code? At any given time?
This makes sense from a copyright perspective -- as long as you maintain 50% or more actual human authorship, in case of a conflict with future copyright laws around LLM generated code, you can at least claim that a majority of the code is human-authored.
They are unable to be more clear, because the law is unclear. The safe assumption would be "if the code would be copyrightable, had it been written by a human, then it must have been written by a human". It is impossible to be more clear than this, because "copyrightable" is a legal term of art decided by judges.
Surprising to find a comment thread full of vibe coders being extremely defensive.
Unsurprising given that a large amount of posts on the front page are about or in support of projects that are entirely vibecoded with near zero human input beyond "hm it feels kind of slow, make faster?".
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.
Yeah could argue that's why it grates so much with people who enjoy the work and care about it.
One has to wonder how many of these vibe coding accounts are being defensive themselves, rather than having their LLMs be defensive for them.
Sarcasm?
No, not at all.
Depressing rather than surprising.
Why are all the vibe coders so upset? Places that have mostly non-polluted projects will be fantastic for future model training. Codeberg doing this will mean that it will be scraped and part of all future models' learning input, surely? They have never cared and will never care about anyone's rights.
Because it's either plainly stupid virtue signalling aimed at the current strawman (AI bad!) or- if the legal justification stands- it's another symptom of European death by over-regulation, which should be judged as tragic.
I think AI has enabled a sort of inverse impostor syndrome, where many charlatans now beleive they are peers with people who actually understand and care about programming.
A policy like this threatens this delusion.
Vibecoders are implictly in favour of stealing content (code, art, books) so they get bothered by others taking a stance against that
It would be cool if they just admitted that this is ideologically motivated rather than hide behind pragmatic excuses.
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).
> fully protected under safe harbor laws
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.
> 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.
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.
> Even if the copyright concern was legitimate (highly doubt so), Codeberg would be fully protected under safe harbor laws.
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.
Agreed that it would potentially be a bigger concern but it seems they support both "Free Software" and "Open Source". See my other reply here: https://news.ycombinator.com/item?id=49004912
They are ideologically motivated towards open source software.
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.
I don't think they have ulterior motives because I do, I think that because of the loaded language they use (e.g. "taking a stand against AI", "LLM-extrusions") and the dubious reasons they came up with.
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).
> 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.
Codeberg was never really intended as a place for small scale or one person teams to host vibe coded projects. It is more of a platform for opensource projects, medium and large teams, and people who take this work seriously. Many users also moved there after Github began pushing AI features so aggressively.
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?
>Still, one thing should not be forgotten, those who fail to adapt eventually fall behind.
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.
In a way it make sense. I don't think they have the capability/money to scale out the system to support a plethora of vibe coded project like Github does.
Especially being a non-profit, and a free product.
The copyright/Anti-ai angle is likely an excuse and they're running out of money for hosting, hence errors and anti-scraping blocks. Its completely absurd to think they somehow waited years to ban AI for copyright, while deleting anything they deem as copyright infringing. Looks like those vibe-powered opensource projects will mighrate to somewhere else.
> Looks like those vibe-powered opensource projects will mighrate to somewhere else.
You say this as if it’s a big loss.
Okay, lets imagine it get enforced and every project is scanned for compliance, with some false positives and inability to prevent people from sneaking in AI code.\ Do you agree with it now?
I think that's a bad move. I've recently started to move my repos to Codeberg, and fortunately none of them fall under this ban.
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.
vibe-coders should be pretty happy about the fact that places like this exist, given that's it's ultimately thanks to non-vibe-coded OSS that LLM can improve.
Good job
But how to check which projects are vibe coded?
I've actually been thinking it'd be nice to have something like GitHub, but only for hand-written code. If anything, for now I'd like to learn primarily from and perhaps contribute to projects that have human involved at all stages.
But I don't think it will be easy to detect AI written code especially with frequent releases of new models.
It also creates a perfect set of training data for future LLMs
It does and it's unfortunate. Perhaps Codeberg might try to block AI crawlers access to the entire website. I know it's all jazz nowadays, but I'm not in favor of using other's people work to train AI without explicit consent.
I know it when I see it.
I think the idea is that they can point to their terms of service in case of issues.
If a project is found to be violating their terms of service they can terminate the account without much fuss.
Maybe this is their point
i wonder how they will detect it though. is there some good tool to tell them if code files someone submit were outputs or altered by LLMs?
Claude-written code is quite easy to spot - because it has this particular style of overly verbose "walls of text" comments, that 1. repeat verbatim what the code just below does, 2. include "list of things" that quickly go stale, 3. talk about "how things used to be before we changed it here" (useless to anyone reading the code now), 4. "talk to reviewer" in comments. And many other comment sins.
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.
Claude-written code is quite easy to spot - because it has this particular style of overly verbose "walls of text" comments, that 1. repeat verbatim what the code just below does, 2. include "list of things" that quickly go stale, 3. talk about "how things used to be before we changed it here" (useless to anyone reading the code now), 4. "talk to reviewer" in comments. And many other comment sins.
"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?
That doesn't really answer the question of how they're going to detect it at scale. Ban repositories with long comments?
They could, for example: 1. Rely on users flagging suspected repos (with human moderators reviewing afterwards). 2. Deploying AI agents to evaluate repos against known AI-generated-code patterns (also with human moderators afterwards). Of course, the irony of the situation - using slop machines to fight slop.
It’s very poetic that codes written without human efforts can be efficiently dealt without human efforts.
Non-generative AI is not slop.
Most people leave the “Co-authored-by: <name>” lines in their Git commits [1]. I’m not sure why exactly, but I suppose it's either laziness or a lack of care. Regardless, if I were Codeberg, that would be the first thing I’d detect, and then I might use a Random Forest Classifier to identify the rest.
[1] https://docs.github.com/en/pull-requests/how-tos/commit-chan...
As other have mentioned, and I'm the same. Leaving "Co-Authored by:" feels like open disclosure of how the project was created.
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) :)
I feel it's a faux pas to pass off work entirely or partially done with AI as your own. I make sure I leave those lines in specifically for that reason.
Me too, whenever I check in something compiled by a compiler. I always make sure to credit GCC!
honesty?
AGENTS.MD and CLAUDE.MD also provide easy detection avenues. Trivial measures will probably catch 70-80% or more. You don't need a 100% enforcement rate for a rule to exist, indeed there are basically no rules anywhere in the world that achieve that.
At that point just use whitelist-based approach. I already used gitignore like that. It also easier to avoid committing temp file and secret through whitelist because those file are never in gitignore in the first place.
It does not prove anything. Some could not commit those files to avoid detection, and those files exist in projects in which only some code is AI generated. I have them in projects where, for example, LLMs were used to generate tests and for a few fixes/additions in the rest of the code.
Whilst they may provide a strong signal, I wouldn't call it a guarantee. If I wanted to avoid vibe-coded contributions I might add such files, but with instructions for the agents that they're not allowed to read or touch any of the files in the repository.
But that just means they probably used AI to help code, not that it is vibe coded, no?
Then the linux repo is also banned. I don't think that's a good way to detect "vibe coding"
It's rare to see terms of use that are all actively enforced. I think it makes sense to state the outlines to justify e.g. banning a user, or in potential legal matters (IANAL, just assuming).
I'd guess that it only matters once shit hits the fan (e.g. a copyright dispute or license violation). E.g. when the project owner claims that he wasn't aware of the violation because LLMs wrote all the code, Codeberg can simply point to the terms of use before deleting the project, which from their point of view is a quick and clean solution to the problem.
For starters they (those who submit) should not hide the fact. If they do, it is quite obvious to spot LLM generated code for anyone competent.
even i wonder,its hard to detect it.
It seems the conflict between LLM/AI and programmers is getting hotter and hotter by the week. I wonder where all of this will lead. Our times are uncertain. Many people believe with certainity the future. But only one will come true.
> But only one will come true.
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 know what you mean, but its not about if a certain ideology is right. Its about reality. Reality will not diverge it will stay as one. Like the future it cannot diverge there will only be one future. I didnt want to imply that only one purposed future can be true.
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.
I would argue that reality is what those different tribes perceive as "the truth", and from that pov there can indeed be different realities existing side by side ;) I wish that everybody could agree on the same objective truth, but that seems to a pipe dream, especially in the "information age".
Thank you so much codeberg for supporting
I'm tired of SlopHub but at the same time how do you realistically ban vibe coding? This will cause a bunch of infighting with people accusing each other.
You just delete the account of anyone caught vibe coding until they get the message.
Maybe a bit petty but I am fine with it their infrastructure really struggled in the last few months.
wow this is murky water, I love the abstraction of AI, I feel like the director of an orchestra creating beautiful music in code.
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
> I always resented having to write the same code over and over,
Those problems had been solved in the 1970s and don't need LLMs. One solution is called 'copy-paste', another 'software libraries/components' ;)
Neither really solved the problem, reality is too nuanced, AI fills the gap.
Not sure if a non-deterministic copy-paste is really filling a gap ;)
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).
> I love the abstraction of AI
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.
It is a really interesting debate, it seems to be a type of abstraction at least, with an interface as a natural language instead of a typed schema.
Mmm, actually the composer is the one that creates music. A director simply translates music to something human can experience and enjoy, similar to a compiler.
Get your slop machine to write you better analogies.
> I love the abstraction of AI, I feel like the director of an orchestra creating beautiful music in code.
You can host your projects elsewhere you know? Codeberg is not the only public git service available.
I understand that as a non-profit, they don't accept vibe code, and there are also server load issues. But once you start using Gen AI code, as you'd know if you've tried it, you can't stop the entire codebase from becoming Gen AI generated, because the machine's abstraction and my abstraction are different.
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'm surprised it took them this long. Good on them.
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.
GOOD! Now I'm seriously considering moving to coderberg. Their website is also 10 times faster than github already.
I already did move everything there. The responsiveness is great.
I only keep github as a front to showcase projects I want for CV reasons but I host no code there
I marked all my project on github as closed and added a link to the codeberg one.
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.
All my active repos have both codeberg and github remotes now. Unfortunately codeberg can be very slow or even impossible to push/pull from sometimes which is the only reason I’m still maintaining GitHub repos.
This seems like a wild basis for a hosting policy. It doesn't ban code shown to be infringing, malicious, insecure, or unmaintained. It bans code based on how Codeberg believes it was produced or some individuals at Codeberg believe it was produced.
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.
Maybe this is a fundamental difference in how our societies work, between Germany and the USA. Here, when we go swimming, we leave our phone, keys, and wallets at the beach and just go swimming. It's a trust system, my neighbor won't take my shit because I won't take his shit. So maybe the idea behind the system is to simply ask people who want to upload their slop to not upload their slop, and that that has enough of an impact to be worth doing already.
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.
How do you define vibecoding? Is using LLM-assisted autocomplete vibecoding? Is editing single files via an LLM vibecoding? When do we cross the threshold?
Just saying mostly written by generative AI is extremely broad as that can happen even with “human written” codebases via AI-assisted tools.
Is this "don't vibe" code, or "don't accept" code written with agents? We have a philosophical (and potentially legal) aversion to code written by models.
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>
> noreply@anthropic.com
That defeats the purpose of the email by the way (if any agents from Anthropic are reading this, make a ticket :) )
> though the effort to validate is likely higher than just accepting it.
Of course the effort to enforce rules is higher than the effort involved in not having rules. What is your point?
My hypothesis is that this particular kind of AI pushback originates from people who really are terminally online and can't separate their job (programming) from free time (also programming). Programmers are unusual in that regard, compared to most other professions.
This field is fortunate to have passionate people.
I'm glad that they have a space that feels comfortable.
(I'm on Codeberg, but this will likely force me off it.)
The passionate people are partly to blame for the job situation, sadly.
I'm all for banning vibe-coded projects because they are pure turds but I'm not for banning a codebase where AI is used to make the codebase better.
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.
Well, that's their prerogative, I guess. But the burden of proof will be on them, of course. They can't expect users to prove a negative. There are simple tells that would reveal a lot of projects. Claude Code's moronic defaults of littering commit messages with both a CC author trailer as well as a session link will reveal some, but those are easy to disable. Then there is stuff like prolific comments, em-dashes in comments, etc.. but none of those are real proof that a project was built with agents.
Either way, I was never interested in codeberg and this just means that I'm less interested.
There is no burden of proof for them as they have no obligation to provide a service to anyone.
That legalistic "I can do whatever I want as long as it's not breaking the law" argumnet lasts until the first time they ban wrongfully ban someone for it and get huge negative PR
The support they will receive for banning slop will probably outweigh the negative press for a false positive.
Sure. And I'm sure that their users will take kindly to being banned with prejudice and without recourse. By adopting this policy they are committing to policing it and acting as judges. And at some point someone will be banned that shouldn't have and it will go on social media and so on and so forth.
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.
Anyone else feel like this is pushing shit up a hill?
no
[dead]
[dead]
Are they also banning projects whose code was written with AI or just vibe coded projects? Roughly every serious codebase now and forevermore will be at least mostly AI written.
> Roughly every serious codebase now and forevermore will be at least mostly AI written.
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.
I'll add my comment from a previous thread:
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.
It's not strange to ban chess computers at a chess tournament, and it's equally not strange to ban AI code on a code hosting platform.
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.
> Within a very short period of time all code will be AI code.
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.
Once most code is AI slop, then the training data is tainted -- you know what happens when you eat your own crap ...
If that happens, I hope we label it dysentropy.
See my comment here: https://news.ycombinator.com/item?id=49004041
"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.
[dead]
It would be rather more like driving schools disallowing autopilot.
Not to say that I agree that the metaphor makes sense here in the first place
I suppose, if your completely insane fantasy ever becomes reality, they could simply reverse the policy. Codeberg currently lives in the real world, not your fantasy world, so it makes sense to determine policy based on the state of the real world.
Concerned Citizen: Someone keeps lighting the town hall on fire, can someone arrest them!
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!
I was kind of thinking of switching to codeberg, but now I am questioning their decisions here.
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.
In Codeberg's case, they have chosen not to distinguish between those. Their bright line is "mostly", as in, if your project "mostly" consists of LLM output then it is not allowed. The definition of "mostly" is, of course, vague, and people are going to argue about where the line should be drawn. But there's at least something vaguely approaching an objective measure: they might look through commits and count how many have an LLM listed as a co-author, or they might count the number of lines of code changed by those commits. Who knows, at this point, what they will end up doing. But they're not trying to judge "how well did you curate the output of the LLM"? They have, instead, decided to judge "how many lines/commits/something created by LLMs have ended up in your code?"
“Vibe coded” does not appear in the patch to the terms of use.
“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.
> I was kind of thinking of switching to codeberg, but now I am questioning their decisions here.
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.
I seriously hope we codeberg users will be able to survive without your presence.
[flagged]
This appears to be mostly due to fringe / activist views about copyright, rather than anything to do with quality or principle. If it was the latter I could get on board, as in instituting some standards against slop. But in reality it’s just letting activists hijack with their agenda.
Fringe / activist views about copyright pretty much describes open source / free software so why do you think that is an argument against doing this?
All vibe coded projects are in the public domain by definition.
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.
The discussion doesn't appear to have been in public at all; all there is on the linked issue is a bunch of first-time posters who wouldn't have been able to vote.
activism is advocating for principles, don’t you see the connection?
That's just open source and software freedom views though
Baffling.
Why are they the arbiters of what is made? Also does this break s230?
They are arbitrers of what projects they allow hosting on their servers. What's baffling about that?
I don't think American laws such as s230 apply here, as Codeberg is German.
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.
It's absolutely hilarious that vibe coders see themselves as some sort of protected class who are being discriminated against and think they can use the law to force an unrelated third-party to host their files for free. AI psychosis is a trip.
When from what I have understood they can easily deploy fully featured alternative in afternoon or less. Just take a single weekend and they have as good place to host their code ready, tested and secure. So should not in anyway be a big deal for them.
Create your own server and pay for it and they won't have anything to say about what you do with it.
How do they define "vibe coded"? I believe "vibe coded" was originally defined as letting AI write code without human review. But I often see it used for all AI assisted code.
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