Free and open source software developers us AI coding assistants such as Claude Code, Copilot CLI, Antigravity, and OpenCode in their daily work. The Software Freedom Conservancy responded to that trend with a set of recommendations for contributors who use these tools, which it groups under the label LLM-gen-AI, meaning generative AI systems backed by LLMs.

Best practices, ordered by priority

The document contains 14 recommendations presented in the authors’ order of importance. The Conservancy describes the items as best practices that guide contributors who decide to use these systems.

The top items address how projects treat people. The FOSS community should actively support contributors who reject LLM-gen-AI systems. Every contributor deserves self-determination over whether to use them, and the Conservancy calls on the industry to keep such use optional and to adopt non-discrimination policies for those who opt out.

The recommendations point to reports of technology workers ordered by management, sometimes under penalty of termination, to use these systems for all of their work. Projects that hold a zero-tolerance policy should still welcome contributors who submit AI-assisted work and respond with the same courtesy given to an inadequate first patch.

Review, disclosure, and records

Contributors must spend substantial time reviewing AI-assisted and AI-generated contributions before submission. Disclosure of how and when an LLM-gen-AI system assisted a contribution stands as a moral imperative, and that information belongs in commit logs in a machine-readable format, with the system name, its version, and a description of its role.

Prompt-generated contributions that lack human vetting belong only in areas a project designates for them. Contributors should keep records of their prompts and interactions and archive these meta-artifacts as part of the Corresponding Source, regardless of license.

Licensing and strategic use

The Conservancy advises caution around the legal significance of generated contributions. Changes made to a codebase under a copyleft license must carry that project’s license, due to copyright and the contractual terms of the license. “Copyleft Everything” remains the safest approach, since widely used LLM-gen-AI systems were trained on well-known copylefted code.

The Conservancy will lend staff time to the copyleft-next project to develop a license suited to LLM-gen-AI outputs. Use of these tools, including proprietary ones, counts as an appropriate strategic compromise when they can accelerate FOSS improvements. The recommendations also caution against overuse and skill atrophy, and ask developers to weigh resource consumption, including the environmental impact of these systems.