FreeBSD policy: AI-generated source code ? No, thanks!

In the current status report, the FreeBSD developers have clearly spoken out against incorporating LLM-generated code into the classic open source operating system. They do see advantages of LLMs when generating or revising documentation and other texts. Translations, for example, can be generated much faster and in sufficient quality with an LLM than by purely manual work. However, LLMs may not be used to generate source code. The report cites "license concerns" as the primary obstacle, which is certainly a problem, as the output of any LLM comes from content from all kinds of sources that has previously been copied without consent.

However, if you regularly read the mailing lists of the various BSD projects, you will notice another problem in particular: Time and again, questionable ideas, improvements to problems that do not exist and supposed patches that do not work at all are submitted by people who have not previously appeared on the lists. This spamming with the help of ChatGPT and co. is always relatively easy to recognize, but it costs the developers, who are not abundant anyway, valuable time. It is therefore no wonder that at least in the BSD systems camp, this type of development work is viewed very critically and with displeasure.

The official wording in the current status report for the second quarter of 2025 can be found under "Policy on generative AI created code and documentation": "Core is currently investigating the establishment of a policy for the use of LLM (including, but not limited to, the generation of code). The result will be added to the Contributors Guide in the document repository. AI can be useful for translations (which seem to be faster than manual work), explaining long/unintelligible documents, finding bugs or understanding large codebases. Currently we tend not to use it for code generation due to license concerns. The discussion will continue at the Core session of the BSDCan 2025 developer summit, and Core continues to gather feedback and work on the policy."

NetBSD and Gentoo don't want LLM code either

If other open source projects welcome the use of ChatGPT & Co as a source for security-critical code, among other things, then this must be accepted. But it is perhaps a good idea if at least part of the open source world tends to rely on conservative programming methods and prefers human intelligence. This creates methodological diversity and is fundamentally better than an LLM-generated monoculture of source code spanning all systems, which at some point nobody will be able to penetrate.

Concrete examples: NetBSD changed its commit guidelines in May 2024. LLM-generated code is considered "tainted" and may not be submitted without explicit permission. Gentoo explicitly prohibits any code contributions created with LLM tools. The Gentoo Council justified this with concerns about copyright, quality and ethics. And the QEMU project also has a clear commitment: code that is known or even suspected to have been created with LLM will be rejected. Reason: Unclear license origin and possible non-compliance with the DCO (Developer Certificate of Origin). The Linux Foundation, on the other hand, generally allows the use of AI tools, but relies on the responsibility of the developers.

Quarterly report with a delay

The FreeBSD project usually publishes a detailed status report in the month following each quarter (i.e. in April for Q1). The reports provide information about the current state of development, i.e. progress in system development, important updates to ports and packages, changes to the infrastructure and contributions from the community. In this way, the FreeBSD project shows transparently what has been worked on, what plans there are and how the project is developing.

The current status report for the second quarter of 2025 should have been published in July, but was delayed. According to the developers, more and more important reports have been received that still had to be included in the report. At least this shows that the development of FreeBSD is now picking up speed again.

FreeBSD admins and users find interesting information in the reports and like to skim through them. The current report, for example, refers to the now available time-controlled ZFS snapshots or CPU pinning for Bhyve VMs. Also amusing is the announcement that the porting of FreeBSD to the PinePhone Pro is making progress and that there is a network driver for "headphone-to-USB-serial" adapters – that works via PPP and is somewhat romantically reminiscent of computer communication in the last century.

(mack)