The Free Software Foundation Europe published an analysis on 25 August addressing a question that has become urgent as LLM-generated code enters production codebases: if an AI writes the code, who owns the copyright, and can that code be licensed as Free Software at all?
The answer, under current law across most jurisdictions, is straightforward for fully AI-generated output and murky for everything in between. And the implications for Free Software are direct, because copyright is the legal mechanism that makes copyleft licenses work.
Copyright requires a human author
Copyright vests in the original author the instant a work is created. That author must be a natural person. The US Copyright Office stated this clearly in 1965 and reaffirmed it in January 2025, even as AI technologies advanced. The position is that copyright covers an author's creative implementation, their specific expression of an idea, not the idea itself.
The EU follows the same principle, though without harmonized legislation across all member states. A 2024 Council of the EU policy questionnaire found a majority of member states agreed that existing copyright principles already address LLM-generated works, and that wholly LLM-generated output cannot be copyrighted. The Court of Justice of the EU has ruled in multiple cases that copyright requires a work to represent the intellectual creation of a natural person, reflecting the author's personality through free and creative choices.
Courts in individual member states have applied this reasoning directly. The Municipal Court of Prague ruled in 2023 that an LLM-generated image could not be copyrighted unless the creator demonstrated it was the result of their unique creative contribution. The Munich District Court dismissed a copyright claim over three logos created with an LLM, finding that the person who generated them could not claim them as their own work. The court did not rule out copyrightability for LLM-assisted works in general, only that the specific outputs at issue did not qualify.
What this means for copyleft
If a piece of software is not copyrightable, it is in the public domain. That means no one, including the person who generated it, has the right to apply license terms to it. No Free Software license, no copyleft provision, no distribution conditions of any kind. The FSFE's analysis frames this as a direct problem for developers who use LLMs to generate code and then attempt to apply GPL, MIT, or other licenses to the output.
The distinction matters because Free Software depends on the copyright holder's ability to grant permissions. If there is no copyright holder, there is no one with standing to grant those permissions, and the code sits in the commons without any legal framework governing its use.
The gray zone of assisted creation
Fully LLM-generated code is the clearest case. But most real-world usage involves a developer directing an LLM, selecting and editing its output, and integrating the result into a larger project. Existing case law on machine-assisted creation offers some guidance.
The US Supreme Court's 1884 Burrow-Giles decision on photography established that using a machine does not negate copyrightability. A photographer's creative choices in framing, arranging subjects, and selecting what to capture constituted sufficient authorship, even though the camera performed the mechanical work. The court treated the machine's contribution as minimal because it was rote and directly mechanical.
A 1991 Third Circuit decision, Andrien, addressed a compilation of maps that a plaintiff had directed a printer to rescale and print. The court ruled that authorship belonged to the person who translated an idea into a fixed expression, provided the executing party did not make intellectual modifications. The court warned that this principle has limits: a process that is rote or mechanical in a way that requires no intellectual modification can be disregarded as a contributing author.
The CCNV v. Reid case before the US Supreme Court clarified that commissioning a work and providing detailed instructions constitutes only the provision of ideas, which are not copyrightable. The sculptor who executed the work was the author, not the organization that described what it wanted.
Applied to LLMs, these precedents suggest that a developer who writes detailed prompts and gets code back may be providing non-copyrightable ideas. The developer who selects, arranges, and substantially modifies that code into a working program is doing something closer to authorship. The line between the two is not drawn clearly enough in current law to give confident guidance.
The UK and Ireland diverge
Section 9(3) of the UK Copyright, Designs, and Patents Act of 1988 assigns authorship of computer-generated works to "the person by whom the arrangements necessary for the creation of the work are undertaken." Ireland has a similar provision in Section 21(f) of its Copyright and Related Rights Act 2000. Both provisions predate modern AI, and neither has been tested in court against LLM-generated content.
The UK and Ireland frameworks mean that computer-generated works must have an author, regardless of whether a human directly created the expression. This puts them at odds with the general EU position that only natural persons can hold copyright. For developers in those jurisdictions, the question of who "made the arrangements" to generate code with an LLM is legally relevant but unresolved.
Where this leaves developers
The FSFE's analysis does not resolve every question, but it identifies the core tension. Free Software licensing requires copyright. LLM-generated code, when the human contribution is limited to prompts, may not have copyright to license. Developers who want to release AI-generated code as Free Software need to ensure their own creative contribution to the work is substantial enough to establish authorship, or they need to accept that the output may be unlicensed and unenclosable.
The practical advice for developers working with LLMs is to treat the tool as a starting point. Edit, restructure, and transform the output. Make creative choices about implementation. The more the final code reflects your decisions rather than the model's first draft, the stronger the copyright claim, and the more meaningful any Free Software license applied to it becomes.