SourceHut Bans Code and Content Created with LLMs

SourceHut, a platform for hosting free software, has announced a change to its terms of service: starting September 10, 2026, its list of prohibited content will include “original content authored with the assistance of, or facilitating the use of, LLMs and other generative AI technologies.” The ban covers source code, assets, tickets and emails produced with the involvement of such tools. The decision was made following a public discussion with the community, where, according to the administration, support for the restrictions prevailed.

Practical reasons and matters of principle

SourceHut also has thoroughly down-to-earth motives. “Vibe-coded” projects, as a class, turn out to be record-setters for resource consumption: they invent bloated CI configurations that burn through build minutes, spawn overgrown codebases, and often push commits frequently, which leads to overloads and outages. Such projects are frequently abandoned, having wasted computing resources for nothing. Add to that the invasion of AI crawlers that wear down system administrators.

The main role, however, the team admits, is played by ethics and politics. LLM training is set up so that models absorb enormous volumes of open code without regard for licenses or the will of authors, and then charge money for code that is essentially a plausible recombination of the training data. The models themselves remain proprietary and give almost nothing back to the community. As an alarming example, they cite the rewriting of the chardet library with the help of an LLM in order to circumvent copyleft obligations.

Separately, the authors speak about the climate: the planned AI infrastructure, by their figures, will require electricity comparable to the consumption of the whole of India, diverting “green” generation to itself and slowing the decarbonization of other industries. Social costs are mentioned too — the exploitation of low-paid labor for data labeling, pressure on the job market, and the “deskilling” of AI users themselves.

How it will work

The rules are formulated with caveats. The word “original” makes it possible to keep hosting mirrors of major projects on SourceHut, such as the Linux kernel, whose policies on AI are more lenient. Private use of LLMs to review your own code, ask questions, or analyze security is not prohibited — provided that you then push hand-written code to the repository.

SourceHut intends to enforce the rules “gently”: there will be no automated AI detectors, and the bet is on honesty. But a user caught cheating will lose the presumption of good faith, and their account will be blocked. No retroactive sanctions against past projects are envisaged; moreover, the team has promised to clamp down on harassment of those who previously used AI.

Those who intend to keep using generative tools are invited to migrate. Among the alternatives named are self-hosting SourceHut, the Forgejo forge, GNU Mailman for mailing lists, and — for AI proponents — “proprietary solutions like GitHub.”

Source: Lobsters