SourceHut could have written “no LLMs” and closed the ticket. Instead, its August 27 post spends several paragraphs on uses that remain acceptable, uses that will not, and uses it has no intention of hunting down.
Start with the permitted side: you may still use an LLM privately to review your work, ask questions or look for security problems. If you then write the code yourself, SourceHut says that use remains compatible with the policy.1
Then comes the red line. Code, assets, tickets and emails produced with LLM assistance will no longer qualify as original project content SourceHut is willing to host.1 The same goes for software whose purpose is to contain or directly enable generative-AI features.
The practical question is therefore not whether somebody opened a chatbot during the day. SourceHut is drawing its line around how the artifact presented to the forge as original work was produced.
Not active yet
On August 28, the public Terms of Service have not caught up yet. They still show the older prohibited-content list: explicit sexual content, malware in certain cases, rights violations, cryptocurrency and so on.2 The new LLM line is not there yet.
SourceHut says it will take effect on September 10, 2026 for new projects, after the usual two-week notice period.1 This is therefore an announced terms change, not a clause already present in the public version today.
The announced clause is short; the blog immediately makes it less abstract with examples: using generative AI to produce or assist with source code, assets, tickets, emails and similar material will no longer be permitted.1
Looking only at Git history will not settle every case. A hand-written tool may still sit outside the policy if the product itself directly enables generative AI.1
Bots can stay
The current terms also make a lazy “SourceHut bans machines” reading impossible.
An account may already be operated by an automated process as long as a human performed the initial registration.2 Bulk automated access is also permitted in some cases, including archival and open-access research, subject to rules around robots.txt, request rates and data use.2
One limit is already explicit: automated tools may not collect SourceHut data to train a machine-learning model.2
By April 2025 this dispute had already become infrastructure: Anubis was sitting in front of some routes to stop aggressive LLM crawlers. and reiterated that SourceHut did not want to provide user data in bulk for commercial or training purposes.3
Until now, the visible boundary mostly concerned what automated actors could take out of the forge. The announced rule adds a boundary around what projects may put in as original work.
The Linux case
The busiest word in the new clause is “original”.
SourceHut says a mirror of the Linux kernel can remain even though the wider Linux project has a more permissive LLM policy than SourceHut wants for its own upstream projects.1 The host is not claiming that bytes become contaminated once an AI tool has touched some distant part of their history.
An individual developer may also host a personal tree and stage patches. A Linux subsystem with its own compatible policy can use SourceHut. But SourceHut says the broader Linux upstream would not be suitable as the primary project on the platform under its more liberal LLM rules.1
A forge is more than a shelf of files. It can be a mirror, a workbench or the primary home of a project. The announced policy mostly cares about that last role.
No detector
No clause makes the awkward forensic question disappear: did this forty-line function come from a person, a completion, a chatbot and then three rounds of human editing?
SourceHut declines the technical premise. It says it will deploy no automated AI-use detector and will not proactively hunt for covert use.1
What SourceHut promises is much less theatrical than automated policing: occasional review, context and room for exceptions. The company compares this trust to the honor system it already uses for financial assistance: users choose support according to their means without producing a dossier of evidence for every claim.1
If somebody lies and the use is later discovered, that presumption of good faith is gone and the account can be suspended.1
This is less mechanical than a storage quota. It avoids a different project, though: building a scanner, investigating every commit, handling false positives and turning every slightly regular writing style into forensic evidence.
Responsibility shifts to the project. SourceHut does not promise to prove where every line came from; users are expected to know their own process and act accordingly. Lying breaks the trust relationship.
Codeberg meters differently
SourceHut explicitly points to Codeberg, which announced its own restrictions in July.14 Their diagnosis overlaps heavily: hungry crawlers, CI and storage spent on projects with little community, maintenance load, licensing and trust between contributors.14
Their operational boundary is different.
Codeberg says in its early guidelines that a tiny LLM-generated side project consuming very few resources may in practice be tolerated. Autonomous projects, heavily LLM-written projects, disproportionate resource users and projects tied strongly to the LLM ecosystem are the cases it says may no longer be welcome.4
SourceHut announces a more categorical rule for original content: even a cheap patch is on the wrong side if a generative tool produced or assisted it.1 At the same time, private consultation remains possible.
That is where the two policies stop being interchangeable. A forge is not merely publishing an opinion about average generated-code quality. It is a decision about which costs and working practices a community agrees to mutualize.
Leaving is part of it
SourceHut does not plan retroactive purges. A project that used AI in the past can remain if future development adopts a compatible policy.1
Projects that intend to keep having generative tools produce code or other artifacts must migrate. There is no rigid deadline stated for these existing projects, but SourceHut asks for a viable migration plan as soon as practical.1
Migration is not left as a vague threat. SourceHut is free software and can be self-hosted; the post also points to Forgejo, Mailman and proprietary forges depending on the workflow a project wants.1 SourceHut's older explanation of its business model makes the same architecture explicit: customers pay the company to host software they are still free to run themselves.5
Migration still costs time. Publishing the forge itself at least leaves users a way to keep the software while rejecting the hosted service’s social rule.
A social rule
The text does not pretend to settle the old fight over authorship, originality or exactly how much assistance turns a tool into a coauthor. SourceHut deliberately leaves interpretation room and promises case-by-case enforcement.1
The result is narrower: the text says where SourceHut places responsibility.
The crawler vacuuming up the commons for model training is already outside the line. A developer privately consulting an LLM may stay. An original artifact co-produced by that LLM may no longer become SourceHut upstream. And the host will not try to infer this with a probabilistic score; it expects users to make the distinction through their behavior.
This is less an anti-AI technology than an admission contract.
For a forge, deciding who and what it wants to host has always been part of the product.