The Umbrellix Logo. Also a civil emblem of the Evdonia micronation. It's a transparent image with bottom-right corner stripes of purple, grey white, celestial, and green.Umbrellix™ Umbrellix Software House

Umbrellix Software AI policy

For the purposes of this document, an Umbrellix project is any project which has decided to use the Umbrellix AI policy, which is this document.

All patch requests to Umbrellix software are subject to our strict policy against LLMs, including self-hosted models.

No Umbrellix project permits contributions generated by or with the help of large language models, be they commercial such as DeepSeek, ChatGPT, or Claude, or self-hosted models trained only on your own code to which you have forever given away the copyright (which obviates the copyright violation concern inherent in other LLMs). No Umbrellix project permits contributions using other forms of generative AI, including, but not limited to, image diffusion, neural network speech synthesis (exceptions can be arranged under very limited circumstances), Markov chains, and Dr. Sbaitso.

Exceptions to this policy only apply to upstreams of tracking forks, where it is anticipated that we cannot enforce our no-AI-patch policy consistently. There will however be a best effort made to systematically reduce and replace AI-generated code, documentation, and artwork. This does not soften our prohibition on introducing our own AI-created code.

The detectable use, or declaration of use, of an LLM or other generative AI technology in any portion of your email to a maintainer using the Umbrellix AI policy requesting that a patch be contributed will result in your patch being rejected, and attempts to promote LLM-enhanced forks of a project through its channels will result in immediate and permanent banishment. Those with a pattern of submitting generative AI patches will be forbidden from sending patches even if they vow never to use generative AI again, as nuisance or exhausting users.

Umbrellix projects with open source or USL licences cannot, legally, prevent you from ingesting the project into a self-hosted LLM, or training an LLM on the project (though we can legally prevent you from making transformative uses that would violate the licence). However, because you must know that most commercial LLMs will record data you feed them and use them for further training, by doing so, you are enabling copyright infringement and become, jointly with your AI purveyor of choice, liable to potentially unlimited damages payable to us and our upstreams.

As we cannot prevent LLM ingestion, it is inevitable that someone will, by coincidence as LLMs are prediction-biased random number generators, find a bug by using an LLM. If you somehow manage to find a bug using an LLM, we expect complete collateral construction - rearrive at the conclusion without any bot assistance. Chat, off. Claude, off. Perplexity, off. Copilot, off. All of it off. Refind the bug yourself, build your model of the program and the problem it’s trying to solve in your mind, try to figure out why it’s a bug yourself, and create your own fix. The ideal result, and the only one we will accept, is that we would be unable to tell that an AI tool has even been used. Ideally you come away from the experience emboldened to find bugs yourself, using fuzzers, static analysers, and your own one or two eyes (or ears, or up to ten fingers - we do not willingly discriminate against screen reader and Braille users, and that’s a promise.), not using the AI tool.

Purely technical rationale

According to our friend Atax1a, but given in our words, LLM-generated code is, in a sense, random noise. It is difficult to pick out any kind of mental model of the problem which the code is supposed to be trying to solve.

This is predictable from how LLMs work. They are, essentially, playing a prediction game - which symbol (a character or group of characters) comes next?

This is not an unpopular position.

Our position (or more precisely, its more extreme analogue that prohibits reporting «LLM-found» bugs completely) is not unpopular. Multiple projects have sprouted up to document projects that permit (or, on the positive side, prohibit) AI, to permit more conscious software usage and ultimately discourage further use of AI across open source.

Lists, where linked, may never be able to satisfy particular standards for completeness (reference to Wikipedia hatnote).