Latest

Technology

Libreboot gives up: its anti-AI policy lasts only a month

Libreboot withdraws its anti-AI policy a month after approving it and admits it cannot reject generated code if its upstream projects accept it.

Translated from Spanish by an automated system. The Spanish original prevails. Read the Spanish original

Libreboot gives up: its anti-AI policy lasts only a month
Libreboot gives up: its anti-AI policy lasts only a month

A policy against code generated by artificial intelligence lasted a month at Libreboot, the free boot firmware project that inherits from coreboot. The new text, titled "AI Slop Reduction Policy" and published on November 20, 2026, does not ban anything: it commits to reducing AI's influence "where viable", which includes reverting patches from upstream projects, rewriting them or, in the worst case, forking them. The project's own note explains the change without embellishment and acknowledges that the previous stance was untenable.

The case matters because it is the first time a free firmware project explicitly backs down in this debate, and because the argument it uses is not technical but structural. Libreboot cannot reject AI-generated code as long as it depends on projects that accept it, and those projects are the foundation on which everything else is built.

What does the new document say?

The text is neither a full surrender nor a change of sides. Libreboot maintains that AI in its current form is a problem, that its author would prefer it did not exist and that the project will continue to prefer human work when it has the option to choose. What changes is the scope: from prohibiting to discouraging.

The tools it reserves for itself are concrete. Reverting a patch that came from an upstream project, rewriting it or maintaining its own fork are three ways of removing the code it does not want, and all three carry a maintenance cost. In a small project, that cost is what decides.

The policy also acknowledges a limit that many similar documents avoid mentioning. coreboot, the project Libreboot depends on, does not reject AI-generated contributions, and neither does EDK2, the reference implementation of UEFI boot firmware. If the base accepts that code, a derivative cannot declare itself clean without rewriting everything.

The patch that was reverted

The example the project itself documents serves to understand the problem in practice. A patch that set automatic battery charging on ThinkPad laptops was reverted because it was optional and because it had been generated with AI. The result is that charging thresholds no longer adjust themselves, although they remain accessible through the ectool utility.

The story has a second layer that makes the case more uncomfortable. The upstream project, coreboot, had added support for the kernel module that manages those laptops, and that addition broke mobile broadband and Bluetooth on some machines. The fix for that regression was, precisely, AI-generated code.

There is the dilemma in a single decision. Libreboot could have kept the patch that restored functionality and accepted its origin, or dropped support for the module to avoid depending on it. It chose the latter, which is the conservative option in policy terms and also the one that leaves users with fewer features.

Why a downstream cannot object

The structural argument takes up half the document and is the part worth reading calmly. A project like Libreboot does not write firmware from scratch: it assembles pieces from coreboot, from the bootloader, from the blobs it manages to liberate and from the build tools. Each of those pieces has its own policy on AI, and none of them fully coincides.

If the project wanted to be consistent with a total ban, it would have to review the entire tree, detect the generated code and rewrite it, something beyond the reach of a volunteer team that already devotes its time to reverse engineering. The text's conclusion is almost a logical surrender: one cannot oppose something that is part of all one's dependencies.

That audit work cannot be automated with the same tool one wants to exclude, and there lies the practical trap of the debate: whoever reviews needs human judgment and time, and time is precisely what is lacking in a volunteer maintenance project with new boards arriving every month. The automatic detection of bugs with AI in the Linux kernel shows the same pattern from the other side, with models finding errors humans did not see and projects accepting that help because the cost of rejecting it is higher than that of using it.

The author's own contradiction

The most honest part of the document is the one that exposes the contradiction without trying to resolve it. Its author writes that AI is "the worst thing we have built" and that he would like to see it destroyed, and a few lines below he acknowledges that reverse engineering with a decompiler and a conversational assistant multiplies productivity by orders of magnitude.

That acknowledgment has practical consequences the text does not hide. Much of Libreboot's work consists of understanding proprietary firmware in order to replace it, and the tools that speed up that understanding mean more motherboards can be liberated sooner. If the result is that more people escape proprietary software, the author argues that the end justifies the means.

It is an uncomfortable stance and it is formulated as such, without the solemnity of corporate statements. It is also the reason why this document reads better as a description of the community's real state than as a declaration of principles.

The cost of purity

What has changed is the conversation. A month ago the project could present itself as an example of resistance to AI, and today it is the example of why that resistance does not scale when one depends on someone else's code. The lesson applies to any small distribution, not just firmware.

What Libreboot has discovered in a month is that consistency has a price and that the price is paid in features users lose. The previous policy was clear and probably impossible to comply with, and the new one is ambiguous and executable, which in a volunteer project is usually the only real option.

The debate it leaves open will not be resolved with documents. It will be resolved when someone publishes a tool capable of reliably attributing which lines a person wrote and which a machine wrote, and until then each project will decide case by case, as this one has, with the policy in one hand and the release schedule in the other.

José Maldonado

Author

José Maldonado

Activist and blogger on technology, free software and blockchain.

The latest
Linux 7.4: NVMe namespaces in parallel, one second less

Keep reading