No LLM Code in Dependencies
Joey Hess has implemented a policy for his open source projects that refuses to use or depend on any code generated by large language models (LLMs). He argues that LLM code introduces low-quality, hard-to-maintain contributions that degrade software, comparing it to a contaminant that should be kept out of the dependency chain to preserve code quality and developer accountability.
Background
Joey Hess is a veteran open-source developer, best known as the creator of `git-annex` and a former Debian maintainer. This blog post announces a policy for his software projects: they will not accept dependencies (i.e., libraries or packages relied upon to build/run the code) that contain code written by a large language model (LLM, e.g., ChatGPT, GitHub Copilot). The reasoning is that LLMs produce code with "hallucinated" errors, subtle bugs, and security holes that no human has actually reviewed or vouched for. By enforcing this rule at the dependency level, Hess aims to prevent the widespread, unchecked proliferation of AI-generated code from undermining the reliability and trustworthiness of the software supply chain. This reflects a growing tension in the open-source world between the push for rapid AI-assisted development and the traditional values of human accountability and careful code review.