AI License Laundering: Why Asahi's Switch to the GPL Doesn't Close the Real Gap

What's at stake

Asahi Linux is the volunteer project that brings Linux to Macs with Apple chips. On November 7, 2026, in its Linux 7.3 progress report, it announced a change of course: all repositories are moving from the permissive MIT license to the GNU General Public License, version 2 or later. Author James Calligeros cites multiple attempts to exploit the MIT license in order to replace the project and monetize Linux on Apple Silicon. In extreme cases, he says, code for unfinished features was taken from team members' private forks. It was either published without attribution or "sloppily rewritten with LLM assistance" "to obscure its provenance." Asahi doesn't say who was behind it.

Our question: does the GPL help against exactly this trick? Copyleft means modified versions must be distributed under the same open terms. But that only applies if the result still legally counts as a derivative of the original. Our thesis: the GPL makes AI-assisted closed forks more expensive, but it doesn't prevent them. The real gap is the unresolved copyright status of AI rewrites.

What Asahi announced and what has already switched

Asahi had deliberately released a lot of code under MIT so that BSD systems could use it too. According to the report, that is "unfortunately no longer tenable." From now on, all Asahi code is to be released "exclusively under strong copyleft licenses." Asahi points to Wine as a model, which switched from MIT to the LGPL in 2002. The tech press is missing three details: a list of affected repositories, a timeline, and a statement on how existing code will be handled.

A look at the repositories shows that some have already switched:

The key line is in the m1n1 README: code up to and including 1.9.9 remains available under MIT. So anyone who wants to can keep developing from the last MIT version at any time.

Which forks might be meant

The report names no names. It's FOSS cites two projects as context. The first is Aurora Silicon with its Vulkan driver Borealis. By its own account, it's derived from HoneyKrisp, runs on M1 through M5 as well as the A18 Pro, and is maintained in its own Mesa fork. According to Phoronix, the project replaced Mesa's AI rules with its own AI-friendly policy. At the end of October, Khronos, the body behind the Vulkan graphics standard, assigned Borealis its own driver ID. The second project is Gravity Linux, a Fedora derivative from Cody Ho, who previously worked at Apple and OpenAI. According to The Register, Gravity split off over Asahi's LLM policy. It built its M4 GPU driver in about a month, mostly with coding agents. The m1n1 license commit landed five days after The Register's September 25 report. Asahi doesn't say whether the two are connected.

Both forks are open source themselves. Gravity's bootloader, a fork of m1n1, now even puts its own contributions under the GPL. The README contradicts itself, though, on whether it's "v2 only" or "v2 or later."

What the GPL can and can't do

One point often gets lost in the debate: MIT also requires the copyright and license notice to be preserved in all copies. Anyone taking code without attribution is already violating MIT. That's an enforcement problem, not a flaw in the license. Here's how the individual scenarios break down:

The Wine precedent shows what you can realistically expect. TransGaming's proprietary fork WineX, later Cedega, was based on the last Wine version under the X11 license and kept running after the switch. According to Wikipedia, TransGaming initially stopped taking Wine code afterward but later integrated individual LGPL parts. The service wasn't shut down until February 2011, partly because Wine caught up on features. So the license didn't prevent the fork. It cut it off from new code.

Mesa is where Asahi has the least room to maneuver. Mesa's core is MIT-licensed, and contributors must agree to the license of the respective component. Besides Asahi, Collabora and Broadcom also hold rights to the shared Asahi UAPI header. Our assessment: Asahi can't unilaterally relicense upstream code, and GPL-only drivers in Mesa would break with established practice there. That makes the GPL switch weakest precisely for the GPU drivers, which is where Borealis comes in.

Criticism comes mainly from the BSD camp. OpenBSD developer Bryan Steele called relicensing m1n1 and the drivers without the consent of all previous contributors "disingenuous and misguided." OpenBSD's Apple Silicon support depends on m1n1. According to runtimewire, DHH and Peter Steinberger have also criticized the project's AI ban. We couldn't find their original statements about Asahi. DHH, however, has long mocked Linux projects' AI policies, likening them to rejecting the railroad because you like petting donkeys.

chardet: a license switch via Claude, with the analysis still pending

The Python library chardet shows just how open the core question is. Maintainer Dan Blanchard had it rewritten with Claude in five days in late February 2026 and released version 7.0.0 under MIT instead of the LGPL in early March. The package name and API stayed the same. According to Blanchard, the plagiarism tool JPlag found at most 1.29 percent similarity to earlier versions. Between previous versions, the figures ranged from 43 to 93 percent. In his own log, however, Blanchard concedes that Claude read old code, including the entire universaldetector.py file. The legal expert behind shujisado.org therefore sees a clear dependence on the original. Whether protected expression was carried over, he says, is an open question. Blanchard later switched to 0BSD, a license with almost no conditions, to sidestep the question of whether AI code can be copyrighted at all.

Zoë Kooyman, executive director of the Free Software Foundation, told The Register:

"There is nothing 'clean' about a Large Language Model which has ingested the code it is being asked to reimplement."

Clean room refers to the classic process in which one team writes only a specification and another team, which has never seen the original code, reimplements it from that specification. In March, Bradley Kuhn announced an analysis on behalf of the Software Freedom Conservancy and warned it would take a long time. Nothing has appeared on the organization's blog to date. License attorney Heather Meeker, on the other hand, considers it probably correct that 7.0.0 doesn't have to be under the LGPL, since nobody has identified any copied protected material. She describes copyleft as "languishing" anyway. That this isn't a thought experiment is shown by the Malus service: it offers "clean room as a service" and, according to 404 Media, actually sells AI-built clones.

Our assessment: With chardet, the logs are public, and still, eight months later, there's no solid answer. With covert rewrites like the ones Asahi describes, even the evidence would be far weaker.

The legal situation in Germany, the EU, and the US

Under Section 69a of the German Copyright Act (UrhG), a program is protected if it is its author's "own intellectual creation." Only a human can be an author. According to the law firm CMS, code generated entirely by AI is therefore in the public domain, free for anyone to use. Under Section 69c No. 2, adaptations of a program require the rights holder's consent. In the 2012 SAS Institute case, the CJEU ruled that functionality, programming languages, and data file formats aren't protected. An AI reimplementation that only replicates functionality therefore infringes nothing. If, however, it carries over structure and organization, it may count as an adaptation. Where the line falls has to be decided case by case. We found no German ruling on AI code.

The burden of proof lies with the plaintiff. A test described by Taylor Wessing in 2025 also shows that explicitly instructing a model not to use protected code doesn't prevent copying. Closed forks have another catch: their code doesn't even need to be copyrighted. Anyone distributing only binaries can rely on trade secret protection.

In the US, the Ninth Circuit Court of Appeals ruled on September 16, 2026, in the Copilot case Doe v. GitHub that the AI outputs are new works, not copies stripped of copyright management information. The court explicitly left open whether they infringe copyright. Claims under the license agreements are still pending at the trial level.

Codeberg bans instead of licensing

Others rely on house rules rather than licenses. Since July 2026, Codeberg's terms of use have included, under § 2 (1) No. 7, a ban on projects consisting "predominantly of code from generative AI tools," explicitly naming Claude and OpenAI Codex, for example. Members voted 358 to 144 in favor, with 14 abstentions. According to The Register, Codeberg cites the unclear copyright status and "license laundering" as reasons. The rule is only enforced case by case. What "predominantly" means isn't defined. We found no reports of projects being deleted. Asahi also bans generative AI for substantial contributions, arguing among other things that LLMs could violate the clean-room rules of reverse engineering. But such rules only apply to your own project. They don't govern what others do with the code.

Our take

Our assessment: Asahi's move is understandable, but it only targets part of the problem. The GPL changes the math for commercial, closed derivatives. Anyone who wants to use Asahi code has to either disclose the source or rewrite it far enough that it no longer counts as a derivative work, and carry the legal risk of doing so. That complicates due diligence reviews and investor pitches. Against open competing forks like Borealis or Gravity, though, the GPL doesn't help. Against missing attribution, MIT would have sufficed, if the violation were enforced. And the old MIT version stays usable, just as Cedega built on Wine's last X11 version.

The real gap is the question of when an LLM rewrite still counts as a derivative work. There's no court ruling on it, no Software Freedom Conservancy analysis, and no clear line in EU law. Asahi's allegations remain unproven so far. The project hasn't publicly named any examples or anyone responsible. As long as that stays the case, the GPL switch is mainly a political signal: anyone "laundering" copyleft code through AI should at least know they're on shaky ground.