Open Source Project Adopts AI Policy After Finding Docs Generated

The Archimedes Foundation, an open source infrastructure project with seventeen thousand GitHub stars, announced a comprehensive policy on generative AI use on Tuesday, three weeks after discovering that its entire documentation suite, including all governance documents, had been written by Claude 3.5 Sonnet. The policy, which runs to twelve thousand words across seven sections, was approved unanimously by the foundation's steering committee and ratified through the same large language model that generated the documentation it was meant to oversee.

The discovery came during a routine audit by the project's technical lead, Marcus Chen, forty-one, of Portland, Oregon, who was reviewing the provenance of the project's policy library. "We wanted to make sure we were being thoughtful about this," he said via email, adding that the committee had decided to publish the policy immediately and address any inconsistencies during the next quarterly review cycle.

We wanted to make sure we were being thoughtful about this.

The new guidelines prohibit using generative AI for code review, security assessments, and architectural decisions, while permitting its use for documentation, examples, and the drafting of policy itself. The policy also establishes a review process requiring human approval for all AI-generated content before publication, though the approval checklist was itself generated by the same system and has not yet been manually verified. A spokesperson for the foundation declined to specify which sections of the policy were written by humans versus which were generated, citing the complexity of tracking attribution across multiple editing sessions.

The policy recommends that projects using the foundation's infrastructure adopt similar restrictions, though it does not mandate them. An implementation guide for other projects, published the same day, was generated in its entirety by GPT-4 and has been downloaded forty-three thousand times.

At press time, the foundation's technical steering committee was preparing to vote on whether to subject the policy itself to the AI audit process it had just created. The vote will take place over email, with the voting reminder drafted by the same Claude instance that wrote the original policy documents.