Codeberg's over-reaction to AI
Jul 26, 2026 · 3 minute read · CommentsTags: AIFOSS
Words: 622
Codeberg have effectively banned AI (well LLMs) from their platform. Not just from using it but actually any project that develops code using an LLM. It's their right, of course, but now I have to avoid them to, not due to the AI ban *per se*, but rather the implications of the subtle difference of 'what something is' rather than 'how it came to be'.
A lot has now been written about Codeberg and their decision around LLMs (or what is now, sadly, mislabeled as 'AI'). Their blog post ( Protecting our FLOSS commons from LLMs) is their justification for these bans, but what's curious is that they are banning usage of a tool as part of their ideology of 'AI is bad, human is good' mantra.
Way back in the day software was written in Assembler and that's what Real Programmers(tm) used. Then some bright spark(s) invented a compiler and you could abstract your code away from assembler and program in a higher level language. This also led to the language wars (see this PDF) due to differences in opinions about what was the right way to write software (or is that 'right' software ha). Although it's arguable that this is categorically different to LLMs regurgitating tokens, ... ahem, "software", using an LLM is a little bit like 'compiling a series of prompts into working software'. One difference is that the prompts are discarded and only the compiled artifact is retained; it would be like disposing of the C source, and keeping only the binary. Oh?
Anyway, what's more interesting about the Codeberg ban is that they are banning the use of a tool to create software (e.g. like a specific compiler) rather than just a category of software (e.g. crypto, hacking kits, malware, trojan-ware, etc.). That's what's particularly interesting, and worrying, about this ban.
It's the type of things that lead to witchhunts. "You coded with the wrong sort of tool.", "You didn't use x linter." "We're banning Haskell .." (okay, the last one is a little bit far-fetched, but the absurdity is largely to make the point.)
Another thing that strikes me is just how hard it is going to be to enforce it with anything other than "that code smells of LLMs". i.e. what if you used Supermaven in your code? That's AI autocomplete, and there's no AGENTS.md or CLAUDE.md anywhere to be seen. (Disclaimer: I don't use Supermaven, but that's largely because I don't want to send all my code to it.)
So the 'in group' at Codeberg decided to ban this. I wonder what they'll decide to ban next? We can't know, but this does set a precendent that allows them to arbitrarily ban another category or tool next time around.
I won't be moving any repositories to Codeberg. They'll be pleased about that as I do use a little help from LLMs to do some of the boilerplate coding, to research and suggest functions here and there and to review my code for issues. It's a bit of a pair programmer as all my FOSS code is written mainly by me for me, which also represents the vast quantity of FOSS code.
Once LLMs (or coding harnesses) get the hang of manipulating ASTs directly (i.e. once it goes mainstream) there might not be a need for the 'textual' representation of it, as it may be simply more efficient to just store the AST; who knows where this may lead.
LLMs (even post AI bubble crash) are probably here to stay, and thus another tool in the arsenal to create software. A code forge banning it is likely to make that code forge even more of a target for scraping (i.e. not LLM generated code there...). More likely they will lift the ban over time.
But they have set out their stall and that is "we're happy to ban things". This makes them less reliable as a place to store code, and I'm going to go elsewhere. Radicle looks interesting, as does Tangled. And for private repositories, ironically, self-hosted Forgejo is probably where I will end up. I wonder if they will ever get their federation sorted out ...