SourceHut Bans LLM-Generated Code, Following Codeberg
SourceHut changes its terms of service to prohibit LLM-assisted code, becoming the second major forge to follow Codeberg’s lead. If your project relies on generative AI tooling, prepare a migration to Forgejo or a self-hosted instance.
August 27, 2026. Drew DeVault announces on the SourceHut blog a change to the platform’s terms of service that prohibits code assisted by LLMs and other generative AI. Late July 2026. Codeberg paved the way by limiting LLM use on its platform. September 4, 2026. The Self-Host Weekly confirms the policy is in force. Why it matters: this is the second major forge to draw a clear line on AI-generated code, and it forces self-hosters to decide — document a policy, or migrate.
The decision and its scope
The change is blunt: SourceHut no longer wants “vibe-coded” projects, and it owns the stance openly — the homepage now advertises “No AI features whatsoever” as a selling point. The rule covers LLMs and generative technologies broadly, applied to the production of source code.
Enforcement is deliberately light-touch. SourceHut refuses any automated AI-detection tooling and relies on the honor system — the same one it already uses for its pay-what-you-can pricing. Lying about AI use, however, costs you the presumption of good faith and exposes your account to suspension.
Two guardrails soften the transition: existing projects that have used AI may remain if they adopt a policy consistent with the terms for future development, and there is no retroactive enforcement — nor any tolerance for shaming people who used these tools in the past.
The reliance on the honor system rather than an automated detector is itself a stance. Where a commercial platform would have deployed automated classification — with its false positives and de facto surveillance — SourceHut refuses the tooling and treats trust as an operating principle, just as it does with pay-what-you-can pricing. It is consistent with its total absence of tracking: a forge that does not track its users is not about to start spying on their AI use.
More than resources: an ethical choice
DeVault’s reasoning does not hide behind technique alone. The material incentives are real — vibe-coded projects waste build minutes, produce larger codebases, push frequent commits, and LLM crawlers cause outages and drive up hardware costs — but the author says it plainly: “in truth, the rationale is more about politics and ethics than anything else.”
The compass is SourceHut’s mission: “make free software better.” LLMs, he argues, exploit free software — its training data, its licenses, its platforms — while giving little back, and their models remain proprietary. The chardet library incident, rewritten by AI to dodge copyleft obligations, illustrates the risk. The candor is all the more credible because he concedes that LLMs “find real bugs, including security bugs,” and because he frames the decision within the climate emergency — tools he calls “extremely energy-inefficient” for the problems they solve.
Codeberg first, Debian next: a line is being drawn
SourceHut’s decision is not isolated. It belongs to a sequence that has redrawn the ecosystem in two months.
Codeberg opened the dance in late July by restricting LLMs on its forge — documented in our analysis of this summer’s self-hosting trends. Then Debian ruled in late August by adopting the “Responsible Use of Generative AI” resolution — a vote we covered in detail.
The nuance matters: Debian regulates, SourceHut prohibits. But the underlying movement is the same — free-software forges and distributions are no longer treating AI-generated code as a neutral question, and are taking public positions. Every announcement shrinks the space for implicit ones.
The ecosystem splits into three camps
SourceHut’s decision clarifies a fracture that had been simmering. Three positions now coexist in the forge world. On one side, the conviction forges — Codeberg and now SourceHut — that ban or restrict AI-generated code in the name of code quality and free-software ethics. On the other, the commercial forges — GitHub and GitLab — that put AI assistance at the core of their product, with Copilot and Duo. In between, the neutral forges such as Forgejo and Gitea, which host without taking a position and leave the policy to the instance owner.
For a self-hoster, that map is directly operational: it determines where to migrate depending on the position you want to adopt. Choosing SourceHut means choosing a platform that has already decided for you; choosing Forgejo means keeping the decision in your hands.
What it means for self-hosters
For a homelab or a community, the impact is practical. SourceHut is 100% free software: nothing stops you from self-hosting it and applying your own policy — the restriction binds the managed platform, not the code you run yourself.
For those who want to keep using AI, DeVault explicitly recommends Forgejo — easier to deploy — or a public instance, and notes that proprietary options like GitHub “may also be viewed favorably by projects which support AI use.” Data export goes through the hut tool, with high fidelity toward another SourceHut instance and partial import into Forgejo or GNU Mailman.
There is no hard deadline to migrate, but the invitation is clear: affected projects should have a viable migration plan in place as soon as possible.
Concretely, the move comes down to three steps: export your Git and Mercurial repositories (the bulk of the value), pull your mailing lists and tickets through the hut tool, then replay everything onto a Forgejo instance or another SourceHut instance. Migration to Forgejo is partial on metadata — tickets and lists do not always import cleanly — but complete on code, which remains the most valuable part.
Beyond the immediate migration, the episode is a reminder that a forge is no longer a neutral utility. The platform you pick now carries a default stance on AI, and that stance flows down into the projects hosted on it. For self-hosters running a community forge, that is an argument for Forgejo or Gitea when you want neutrality, and for SourceHut when you want the policy already decided. Either way, the days of hosting code without an opinion on AI are ending.
Verdict
If your project generates code with LLMs and still lives on SourceHut, migrate now to Forgejo or GitHub: the platform no longer fits you, and lying about your usage exposes your account to suspension.
If you self-host a forge for a community, use the opening to decide and document your policy on AI-assisted code — having no position becomes untenable when forges are displaying theirs.
If you share SourceHut’s values, the decision makes the platform more attractive: an “AI-free” environment becomes a selection criterion, not a handicap.