AI Policy

1. 1. Issues

The use of large language models as part of the process of writing programs comes with several concerns; the closer the end result is to the LLM's output the greater the concern:

• Quality: Although models are improving rapidly, even the latest (which happen to also be the most resource-consuming and therefore expensive) are estimated *by their makers* to be wrong 25-30% of the time. Sometimes that manifests with the model going off and solving a problem that isn't real, sometimes it means that the model writes code that doesn't work, sometimes it means that it writes code that does work most of the time, or that works all of the time but has poor performance.
• Provenance: Unless the model is open-source with a well documented training history there is no way to know where the content produced by a model came from. Sometimes it's completely novel and sometimes it's regurgitated from the model's training data. If republished the latter case could expose the publisher to legal liability for infringement.
• Copyrightability: work written by humans can be copyrighted. Open Source licenses depend on the code being copyrighted to be enforceable.
• Skill development: If all of the grunt work of writing code is delegated to AI models then there is no path for new programmers to gain the experience and skills needed to evaluate the output of those AI models.

1. 1. Other Projects:

• The Apache Foundation allows use of AI tools subject to the constraints that "The terms and conditions of the generative AI tool do not place any restrictions on use of the output that would be inconsistent with the Open Source Definition." and recommends disclosure of any AI tools used in a submission.
• The Fedora Foundation allows using AI tooling provided that the human contributor takes full responsibility for the submission and discloses the use of the specific AI tool used with a git message trailer.
• While the Gnome Foundation has not articulated a policy, the Gnome Shell has written and Gnome Circle has adopted the following:

While it is not prohibited to use AI as a learning aid or a development tool (i.e. code completions), app developers should be able to justify and explain the code they submit, within reason. Submissions with large amounts of unnecessary code, inconsistent code style, imaginary API usage, comments serving as LLM prompts, or other indications of AI-generated output will be rejected

• The Linux Kernel policy is very permissive, requiring only that a human developer review all of the code in a submission and certify that it is correct, meets the licensing requirements and assumes full responsibility. It also recommends tagging submissions with an "assisted-by" tag with the particulars of the AI tool used.
• The LLVM compiler project allows use of AI tooling provided that "there must be a human in the loop. Contributors must read and review all LLM-generated code or text before they ask other project members to review it. The contributor is always the author and is fully accountable for their contributions. Contributors should be sufficiently confident that the contribution is high enough quality that asking for a review is a good use of scarce maintainer time, and they should be able to answer questions about their work during review."
• OpenInfra allows AI tool use with similar provisos about disclosure, responsibility, and understanding.
• OpenTelemetry allows AI tools but insists that the majority of a contribution should be written by a human.
• The QEMU emulator project disallows AI-generated content but allows other uses of AI by contributors who then create the content provided that use is disclosed in a commit trailer.
• The Zig language project completely disallows 'any' use of AI, but that has more to do with their intentions for a community than licensing or code quality concerns.
• How open source projects are drawing the line on AI contributions provides a nice discussion of policies and implications with some recommendations.