AI Radar · Daily · August 27, 2026
SourceHut, the developer platform run by Drew DeVault, has rewritten its terms of service to prohibit "original content written with or which facilitates the use of LLMs" across the entire service, starting September 10 SourceHut. The ban is not limited to source code: it covers tickets, assets, emails and any software shipping its own generative AI features. The only carve-out is for mirrors of codebases governed by a different upstream policy, such as the Linux kernel.
The reasoning leans on environmental and labor arguments over a technical one. SourceHut calls LLMs "extremely energy-inefficient" and argues planned AI infrastructure will add electricity demand "comparable to the entire country of India," delaying decarbonization and locking in fossil-fuel generation SourceHut. It also frames model training as taking open-source code "by force," disregarding licenses and author intent, then reselling the output, and cites maintainer burnout from low-quality AI-generated contributions.
How enforcement actually works
There is no scanner, no watermark check, no diffing against known model outputs. SourceHut will act on individual reports and judge them case by case, and harassing other users over past AI use is itself a violation. A real ban with no detection mechanism behind it was the most contested part of the policy on Hacker News, where the thread drew 92 points and 33 comments HN discussion. Supporters, including at least one paying customer who called it "an anti-slop policy," backed the stance as consistent with SourceHut's existing "no AI features whatsoever" identity. Critics called the rule "unenforceable" given the lack of tooling, questioned ambiguous cases like hosting security-analysis tools that use AI internally, and noted a tension between DeVault's anti-copyright views and a policy partly justified on licensing grounds. Some said they would cancel paid accounts over it.
Grandfathering and the exit path
Existing projects are not retroactively punished for AI use predating the policy, and SourceHut says there will be no purge of past contributions. Projects can keep their account only by adopting a compliant, AI-free workflow going forward. For maintainers who intend to keep using LLM assistance, SourceHut points them at the door rather than pretending otherwise: the blog post offers guidance on self-hosting, using its hut command-line tool to export repositories and tickets, and migration help, naming Forgejo, self-hosted Mailman and even GitHub as landing spots.
Not the first mover
SourceHut is following, not originating, this line. Codeberg, the donation-funded, community-governed forge, held a members' vote that closed July 22, 2026, passing 358 to 144 to prohibit projects that "mostly consist of code written by" generative AI tools OMG! Ubuntu. The two differ in scope and motive. Codeberg's "mostly" threshold still permits AI-assisted work inside otherwise human-led projects, justified on cost: SSDs rose roughly fivefold, from about €700 to €3,700 per drive, making storage of unused AI-generated repositories unsustainable for a donation-funded service. Codeberg also committed to never training models on hosted code. SourceHut's rule is broader, covering any AI-facilitated content rather than just "mostly" AI code, argued on environmental, labor and licensing grounds, and set unilaterally by DeVault's team rather than by member vote.
Neither platform will scan for violations, so enforcement depends on self-reporting and manual review, an approach neither has published data on. SourceHut has not said how many projects it expects to lose, nor how much of its hosted content already involves LLM assistance. There is also no independent measurement behind the "comparable to India" electricity claim; it is SourceHut's characterization of broader AI infrastructure buildout plans, not a figure specific to code-hosting workloads.
Why it matters beyond SourceHut
Two ideologically driven forges, chosen by developers precisely because they are not GitHub, have now put a hard stop on AI-assisted contributions within two months of each other. Neither is a mass-market platform, but both draw the kind of maintainer who sets norms other projects copy, giving every other forge a concrete template to adopt or reject.