On the 22nd of July, the members of Codeberg e.V. voted to stop hosting projects that "mostly consist of code written by 'generative AI'-tools." Another motion passed committing the association never to use hosted code or user data to train models. The blog post explaining it doesn't lead with freedom or with ethics, it leads with the price of a solid state drive: they bought one for about €700 a few years ago and it now costs them about €3,700. Let's dig into what that means, because many of the reactions I've seen seem to get it wrong.
They didn't change a licence. The members of the association that operates Codeberg changed the terms under which Codeberg is willing to host your code. The Open Source Definition governs licences, and clauses five and six constrain what a licence may say about who gets to use software and what for. It doesn't say anything about being forced to host all open source code. I find it very interesting why so many people feel like a rule had been broken when none had been.
The rule exists inside the licence: you may not say no to anyone in particular. Conditions are fine sometimes, copyleft is proof of that, but not refusing a use, or a party, and attempts to put that kind of refusal into the licence itself have consistently fallen outside the FOSS boundary, whether that was the SSPL imposing special conditions on offering software as a service, ethical licences restricting harmful or military uses, or RAIL licences naming categories of use.
Outside a licence, where the Open Source Definition has nothing to say, it seems like saying no is treated as though it were prohibited anyway. Codeberg updated its Terms of Use and got called a censor for it. So there is a place where saying no is against the rules, and a place where saying no is not against the rules but you'll be treated as though it were. It seems like what is being refused is the ability to say no.
What Codeberg used is a freedom that was always available to all of us. Codeberg is infrastructure held and governed by an association whose members have decided what their shared resources are for. And I think that's the answer to why so many people felt a rule had been broken: a constraint written for licence text became, somewhere along the way, a constraint on who we were permitted to be, and we've been enforcing it on ourselves in places it never had any authority.
We're used to reading FOSS as an opposition, and its history makes it seem like it is. But let's examine what a licence does. It doesn't oppose, it disposes. An open source licence is a standing grant made in advance, to persons unknown, for purposes unspecified, without any expectation that they return to ask whether this particular use is acceptable. Openness is a posture of permanent availability. Openness is not neutral.
The thing a disposition can't do is establish a relationship by itself. We built the relationships despite the disposition, and for a long time openness still came with friction. Taking code at scale meant understanding it, integrating it, maintaining it, and employing people who knew something about the communities it came from. That never guaranteed reciprocity, but it made complete detachment difficult.
Then something arrived that could take all of it at once while bypassing much of the labour that had made the FOSS community structurally relevant. Freedom zero isn't what legally makes model training possible, but the open source tradition did deliberately construct software as something whose downstream use would not require renewed permission from its authors. Non-discrimination is part of what guarantees that the licence cannot withhold those freedoms from a particular recipient because of who is asking or what field they work in.
Copyleft is the counterexample, yet it doesn't reach far enough. The GPL is still a disposition, only with conditions attached, and those conditions principally bind people when they redistribute or convey covered software. But LLM training doesn't map neatly onto that trigger. A model trained on GPL code is not automatically a covered derivative work merely because that code was somewhere in the corpus. It's a defence built against one specific enemy hailing back to the 80s, the proprietary fork. But in 2026 the extraction doesn't need to fork the commons at all.
There's an interesting asymmetry in the four freedoms. They protect the user from the developer's power, and there's no matching freedom protecting a developer from a user precisely when that user is exercising those freedoms correctly. A defence here is needed, but it has to come from somewhere else.
Back to Codeberg, the easy reading would be that they built a fort, and conclude the commons needed a wall all along. But a fort is defined by what it keeps out. An association is defined by what it holds together, and it needs an edge for the same reason any commitment needs one, which is that you can't be obliged to everybody. Gardens, not roads, as I've put it before: a garden has a gardener and a boundary.
Free association has never meant association with everyone on demand. It has included the freedom to decide who you're doing this with and on what terms. Nothing in the FOSS definitions ever required us to give that up, so let's stop pretending there was some sort of betrayal of values here.
Maybe the mistake was expecting licences to do political work they were deliberately designed not to do. Open Source is indefensible, but maybe it's also not where we build our defences.
We don't need to defend the freedom to take from everyone all the time. Our freedom to associate, our freedom to refuse, and our freedom to decide over infrastructure we reasonably control all need defending though, and all of that is outside the jurisdiction of licences and code. Thank you Codeberg for reminding us it was there all along.
I'd like to hear from people who host things and pay for SSDs and have had to make one of these calls, which is a thing I've been doing myself lately and finding harder than the writing. What did you have available to you, and what did it cost?