Copyrightability of LLM-generated code: Can we license “vibe code”
into Free Software?
The use of large language models (“LLMs”) has exploded in recent
years, including in the generation of source code. But even as their
usage gains popularity, these trends bring with them pressing legal
questions as well: if code is generated by an LLM, is that code
copyrightable? And if so, who owns the copyright to that generated code?
It is important for developers to know how to answer these questions,
because copyright is a foundational pillar of the legal framework that
supports Free Software. In this article, we aim to help our readers
understand how copyright law affects your ability to create Free
Software when programmes are made with the help of LLMs, by providing a
breakdown of the various legal principles and court decisions regarding
copyright and machine generated content.
Table of contents
- Copyright and Licensing in Free Software
- Copyright only applies to implementation
- Only a natural person can be a copyright owner
- Public Domain and the level of human authorship
- Outliers for authorship requirements: The UK and Ireland
- Principles of copyrightability in Assisted Works from the USA
- Using prompts alone is insufficient for copyright
- Machine modification of original input
- Modification and/or rearrangement of generated output
- Copyright infringement
- Formalized recognition of copyright in assisted work
- Copyfraud and concealment of AI usage
- Recommendations for AI usage in Free Software contributions
Copyright and Licensing in Free Software
Copyright is a legal construct that grants a person exclusive rights
over a piece of creative work: only the copyright holder is allowed to
reproduce a work, give copies of their work to others, and modify the
work. Copyright comes about by default; it vests inherently in the
original author of the work in the instant that work is created,
including in software developers when they write code.
One problem with copyright is that it makes software illegal to share
by default. To overcome this, a license is used in order to define the
terms under which the copyright holder allows the recipient of the
license to use the software. If that license is drafted in such a way as
to allow the recipient to enjoy the Four Freedoms, then
that license is a Free Software license.
If a piece of software is not copyrightable, that software has no
rights reserved and is in the public domain. This also means that no one
has the rights to apply any kind of license terms to that software,
including copyleft
license terms that work to maintain the Four Freedoms in any
downstream distributions and derivatives of that software.
Copyright only applies to implementation
The use of a machine or computer to generate works that are
traditionally created by human beings does not fit neatly into the
traditional understandings of copyright principles described above. This
is because copyright traditionally covers an author’s creative
implementation and their specific expression of an idea via execution,
rather than the ideas themselves. This gets complicated when the
implementation and execution is taken over by machines.
Indeed, the underlying assumption in copyright law is that the
implementation of an idea in order to produce a creative work to
fruition is the difficult part, which is why this
implementation is deserving of legal status. As an example, the idea of
lions loosely doing Hamlet is not copyrightable, but the implementation
of that idea in the form of the film “The Lion King” is. In
legal terms, this separation of what is and is not covered by copyright
is generally referred to as the “idea-expression
dichotomy”.
The idea-expression dichotomy means that copyright has traditionally
been understood to exist only for human-created work, and copyright has
been similarly understood to be owned only by human beings. Up until
recent technological developments, creative implementation has for the
most part always had to be executed by humans. The introduction of
generative LLMs however disrupts this traditional workflow of
idea-expression, as (in the case of wholly LLM-generated work) a human
being can now contribute merely the idea, with the implementation and
expression of that idea performed by the LLM.
Nevertheless, we can see that copyright law across multiple
jurisdictions remains mostly clear: the copyright owner has to be a human
being, and the copyright should only extend to human-created work.
Only a natural person can be a copyright owner
This idea that only a human being (or “natural person”) may be
considered an author of copyrightable work has been a historical feature
of copyright law. Indeed, the US Copyright Office stated
back in 1965 that:
“[t]he crucial question appears to be whether the “work” is
basically one of human authorship, with the computer merely being an
assisting instrument, or whether the traditional elements of authorship
in the work (literary, artistic, or musical expression or elements of
selection, arrangement, etc.) were actually conceived and executed not
by man but by a machine.”
This understanding has persisted since then, with the US Copyright
Office releasing
a report in January 2025 reaffirming their 1965 stance, even in
light of the advent of AI technologies.
Such understanding can be seen as well in the European Union (“EU”).
While there is a current lack of harmonised, specific legislation
throughout all member states on the copyrightability of LLM-generated
works, there are nonetheless strong indications that copyright does not
apply to purely LLM-generated works within the EU, and only a natural
person can be considered an author. This was outlined in the results of
a policy
questionnaire conducted by the Council of the EU in 2024 to explore
the relationship between AI and copyright, where a majority of EU member
states agreed that current copyright principles within the EU and
national legislations already adequately address the copyrightability of
LLM-generated work.
Specifically, member states acknowledged that wholly LLM-generated
work cannot be copyrightable, but partly LLM-generated work may be so if
it can be shown that the human input in the creative process was
significant, similar to the stance from the US Copyright Office.
Existing case law at both the Court of Justice of the European Union
(“CJEU”) and at the member state level further support this. For
example, the CJEU has clarified in a number of cases that, in order for
authorship and copyright to be applicable, a creative work has to represent
the expression of the intellectual creation of a natural person. In
other words, the work has to be able to reflect the author’s
personality, in the sense that the author was able to express
their creative abilities in the production of the work by making free
and creative choices.
At the member state level, the Municipal Court of Prague in the Czech
Republic stated in 2023 in the specific context of LLMs that a work of
authorship must be the unique result of the creative activity of a
natural person. Unless a creator can demonstrate that an LLM-generated
image is the result of their unique creative contribution, authorship
cannot be claimed.
More recently, the Munich
District Court dismissed a copyright claim over three logos after
finding that a person who created them with an LLM could not claim them
as their own work. In line with the previous reasoning of the CJEU and
the Municipal Court of Prague, the decisive factor of copyrightability
here is whether or not personal creative work of the natural person is
directly reflected in the resulting implementation, despite the
automated creative process. Nevertheless, the Munich District Court here
did not generally dismiss the idea of LLM-generated or -assisted works
to be copyrightable, but merely ruled that the specific LLM outputs at
hand were not.
Public Domain and the level of human authorship
The implication of these principles is that wholly LLM-generated
outputs are not copyrightable as they have no human author, and will
therefore by default be considered to be in the public domain, as
affirmed by an EU
wide study commissioned by the European Parliament.
Nevertheless, while most jurisdictions agree on the above, this
understanding is overly-simplistic. In practical usage, LLM-generated
output in many cases only forms just part of the creative work, with
human input forming the other part. The laws regarding copyrightability
therefore tend to be insufficiently clear and specific on what happens
when LLMs are used together with human contributions in the current
creation of creative works, including when producing software code.
Outliers for authorship requirements: The UK and Ireland
It is also worth mentioning that not all jurisdictions fully recognise
that only the works of natural persons are copyrightable.
Outside of the EU in the United Kingdom (“UK”), Section 9(3) of the Copyright,
Designs, and Patents Act of 1988 provides that:
“[i]n the case of a literary, dramatic, musical, or artistic work
which is computer-generated, the author shall be taken to be the person
by whom the arrangements necessary for the creation of the work are
undertaken."
This copyright framework results in a situation in the UK in which
all creative works must and will have an author, regardless of the lack
of human contribution in their creation. That being said, this was
formulated before current technological advancements, and the practical
application of Section 9(3) to LLM-generated content has not yet been
tested and questioned in the courts.
Ireland has a provision similar to the UK in Section 21(f) of their
Copyright
and Related Rights Act 2000, which hands authorship and copyright
ownership to, “in the case of a work which is computer-generated,
the person by whom the arrangements necessary for the creation of the
work are undertaken”. As an EU member state, this unfortunately
puts Ireland at odds with the general EU position requiring human
creative contributions for authorship and copyright to be
applicable.
Principles of copyrightability in Assisted Works from the USA
Nevertheless, we can take some guidance from existing case law and
legislation that currently deal with assisted creations, either by
machine or other means. For this, US case law can be useful to
conceptualise and establish principles of what constitutes a creative
element that is copyrightable in light of emerging technologies.
The copyrightability of photographs, for example, was the subject of
considerable debate at the time when cameras were new inventions, as can
be seen in a US Supreme
Court case from 1884 (“Burrow-Giles”). Here, it was argued
that photographs were the products of machines and therefore lacked the
requisite human authorship to be copyrightable.
When making their decision, the court considered that taking a
photograph requires various choices to be made by the photographer,
including selecting and arranging objects to be included in the frame,
arranging the subject, and other decisions to evoke the desired
expression or emotional impact, before using the machine itself to
capture the image. In the court’s opinion, once the choices of the
photographer have been made,
“the remainder of the process is merely mechanical, with no place
for novelty, invention, or originality. It is simply the manual
operation, by the use of these instruments and preparations, of
transferring to the plate the visible representation of some existing
object, the accuracy of this representation being its highest
merit."
The court therefore rejected an argument that photographs lacked
human authorship and were the product of a machine, instead ruling that
the use of a machine does not negate copyrightability by default.
Rather, the work is copyrightable if it contains sufficient human
authored expressive elements. Because of the rote and directly
mechanical nature of the machine in producing the image, its creative
contribution was greatly diminished to the point where almost all
creative authorship should be credited to the human who took the
photograph.
A Third
Circuit US Court of Appeals case from 1991 (“Andrien”)
establishes a similar important principle. This case concerned copyright
over a compilation of maps that the plaintiff had requested a third
party printer rescale and print. The plaintiff argued that he had
expressly directed the preparation of the copies in specific detail, so
that the compilation only required a simple and mechanical process to
achieve its final tangible form. Because the printer did not change the
substance of the plaintiff’s original expression, the court ruled that
authorship belonged wholly to the plaintiff, as he was:
“the person who translates an idea into an expression that is
embodied in a copy by himself or herself, or who authorises another to
embody the expression in a copy."
Nevertheless, the court was quick to point out that this definition
is subject to limits. Similar to the reasoning in Burrow-Giles,
a process that is rote or mechanical in a manner that does not require
intellectual modification or highly technical enhancement can be
disregarded as a contributing author.
The US Supreme Court has also considered the degree of creative
contribution necessary to qualify for authorship and copyright. In Community
for Creative Non-Violence v Reid (“CCNV”), a
non-profit organisation commissioned a sculpture, giving the sculptor
detailed instructions on what it should look like. Both parties
eventually got into a dispute over who should be considered the legal
author of the resulting sculpture. In awarding authorship to the
sculptor, the court reiterated that the author of a copyrighted work is
the person who translates an idea into a fixed, tangible expression.
Upon remanding the case back to the trial courts for further
consideration, the lower
court specified that the acts of commissioning the sculpture and
providing detailed instructions constituted only ideas, which as
explained by the idea-expression dichotomy, is not enough to warrant
authorship and copyrightability.
A general trend we see across these cases is that a distinction needs
to be made between using machinery as a tool to assist in the creation
of works, and using it as a stand-in for human creativity. From these
three cases, we can see some legal principles relating to copyright that
are useful for an analysis of the use of LLMs:
- The use of a machine to create a work does not negate
copyrightability for the human creator (Burrow-Giles);
- When a person hires someone or uses something to execute their
established expression, the executing party or thing has no claim to
authorship if they use a process that does not require intellectual
modification or highly technical enhancement of the expression
(Andrien); and
- Providing detailed instructions for the creation of the work by
someone else constitutes only the provision of non-copyrightable ideas,
and that person providing instructions has not done enough to warrant
authorship (CCNV).
There are certain features in the workflow of using LLMs that raise
the question of whether or not they count as human contribution to a
level that warrants authorship and copyright. Using these established
principles can therefore help us better understand the copyrightability
of an AI-generated work.
Using prompts alone is insufficient for copyright
Advocates for human authorship of LLM-generated work often argue that
the required human creativity for the output exists in prompts,
comparing prompting to commissioning a creative work, or providing an
artist with instructions. This opinion is however not generally
reflected in case law and the general principles of copyright law, where
prompts are unable to provide sufficient human control to allow users of
an AI tool to be recognised as authors of the resulting output. Just
like the detailed instructions given by the non-profit to the sculptor
in CCNV, prompts can be seen to merely be instructions that
convey ideas, which are not covered by copyright law.
As a side note, prompts themselves can be copyrightable if they are
sufficiently creative to pass the threshold of
originality. However, even in situations where specific prompts are
copyrightable, they would still be considered a distinct and separate
entity from the output.
The US Copyright Office additionally points
out key differences between prompting an LLM and commissioning
creative work:
- Potential for involvement in the creative process
In the case of human to human commissioned works, the hiring party is
able to oversee, direct, and/or understand the contributions and process
of the commissioned artist. This allows for a commissioning party to
potentially be considered a joint author in some circumstances. With AI
tools, the prompter receives the completed output after the generative
process has been completed, and plays no role in the creative process of
the output after the prompt is received.
- Lack of control over conversion of ideas into a fixed
expression
In general, prompts do not adequately determine the expressive
elements produced, or control how the system translates them into a
specific output. In other words, while prompts reflect a user’s idea,
the user does not control the specific way in which that idea is
expressed.
The gaps between prompts and the resulting outputs demonstrate that
the user lacks control over the conversion of idea into expression, and
that the system is largely responsible for the expressive elements in
the output. Additionally, where no instructions were given, including
for elements in the work that are necessary but may have been overlooked
by the prompter, the system works to fill in the gaps. Indeed, the fact
that identical prompts may on separate occasions generate different
outputs further indicates a lack of human control.
- Effort required to revise and refine prompts is irrelevant
Prompting can often take repeated revisions and refinements (aka
“prompt engineering”) in order to get a desirable outcome for the user.
However, such effort is not relevant as copyright extends only to
original authorship, regardless of the effort or work put in.
Moreover, inputting a revised prompt does not currently seem to be
materially different from inputting a single prompt. This can be viewed
as simply “re-rolling the dice”, which causes the system to select more
outputs for the user, but does not represent the required degree of
control over the creation of the output that is required for authorship.
Affirming this principle, the aforementioned case in the Municipal
Court of Prague also stated that simply writing a prompt to wholly
generate a work cannot generally amount to authorship under copyright
law.
Nevertheless, the principles regarding prompts may evolve in the
future depending on how law and policy view prompting and outputs. For
example, in the aforementioned copyright claim in the Munich District
Court, the court considered that the progressive attempts of the user
during the prompting process may lead to output that reflects their
personality, giving more weight to the idea of output authorship.
Machine modification of original input
There are several generative tools where inputs are able to be
substantially retained as being part of the output. This allows for
users to use such tools to amend, adapt, translate, or otherwise modify
their own human created works. For example, a software developer may use
such a tool to search for bugs or syntax errors in their code, which
will be automatically fixed. This is analogous to what happened in
Andrien, where the plaintiff already had implemented his idea,
resulting in a copyrightable expression with his collection of maps,
before seeking the help of the printer to finalize and refine his
expression.
When a person A inputs their own human-created, copyrightable content
directly into an LLM, modified outputs of that input may still be
credited to A and be copyrightable, provided that the modified output
retains enough features of the original that it retains identifiable and
perceptible features of its original human authorship. A’s own creative
expression identifiable in the modified output may be copyrightable, in
a similar manner that copyright law currently extends to derivative
works. Especially in the case of source code, human created inputs tend
to result in a limited range of what the modified output can look like
when run through LLMs for amendments or modifications, allowing the
output to retain the characteristics of its human authorship.
Modification and/or rearrangement of generated output
The non-copyrightability of individual elements in an overall work
does not prevent that work from being copyrightable under the law. This
is particularly relevant for copyright over software, as software
elements produced as output by LLMs are often used as building blocks in
the assembly of a larger work.
Generating output with AI tools is also often an initial or
intermediate step, with human-authored contributions being added to, or
modified into, the AI-generated content before it is presented as a
finished work. The non-copyrightability of individual elements in an
overall work does not prevent that work from being copyrightable under
the law, if the presence of human creative elements in the entirety of
the work passes a certain threshold, as seen in Burrow-Giles.
This principle has been demonstrated by the copyright dispute over a
comic book titled “Zarya of the Dawn” by writer Kris Kashtanova
in the USA. Although Kashtanova had initially applied for and received
formal copyright recognition from the US Copyright Office in 2022, the
Copyright Office later partially
revoked such recognition, upon their discovery that she had utilised
the image generation software Midjourney to produce the images used in
the book.
In their analysis, the Copyright Office reiterated that the images
that were wholly produced by Midjourney were non-copyrightable.
Nevertheless, they found that the text in the book was written entirely
by Kashtanova herself, and was therefore copyrightable due to it being a
product of human authorship. In addition to the text, the Copyright
Office also found that the manner in which Kashtanova had chosen to
select and arrange the non-copyrightable images together with her text
was sufficiently creative, and reflected her human authorship.
Accordingly, copyright was considered to cover Kashtanova’s authorship
of the overall text and compilation of the book, but not to each of the
individual generated images that make up the book.
Image by Kris Kashtanova using Midjourney AI, Public domain, via Wikimedia Commons
This case is particularly relevant for copyright over software, as
software elements produced as output are often used as building blocks
in the assembly of a larger work. Zarya of the Dawn can be
viewed analogously with the use and/or linking of various functions and
modules in a software project: developers often reuse elements that they
do not own copyright over, and their use of such elements does not mean
that they can claim authorship over them. They nevertheless are entitled
to general authorship and copyright over the overall program that they
have created. Similar to the creative choices made by Kashtanova in the
arrangement and compilation of images in Zarya of the Dawn, the
compilation, arrangement, and internal structure of a software program
can be considered creative enough if it was implemented by a human
developer.
In practical terms, this means that developers should structure their
programs independently if they wish to claim authorship over such
programs and license them as Free Software. They should not be using
LLMs to produce a complete program for them, but rather only use such
tools to support them in the implementation of their ideas.
Copyright infringement
However, legal problems can also arise when the tool reproduces
something that is already copyrighted, or at least closely enough to be
legally actionable. This can be a concern with tools that have been
trained on datasets that include copyrighted materials, and the
likelihood of it happening shifts depending on a number of factors,
including the ways in which the tool has been prompted, as well as the
parameters used for the tool’s training model.
In such situations, the reproduction of copyrighted material (as well
as its distribution or subsequent publication) constitutes copyright
infringement. Generally in most jurisdictions, copyright infringement is
a strict liability act: this means that the intention to commit
copyright infringement is not necessary to establish whether or not the
infringer is at fault. In other words, “I did not know that the LLM
copied/reproduced this work” is not a valid defence against copyright
infringement.
Liability for copyright infringement, if discovered, generally falls
on the party who distributes it, which would be the user and not the
creator of the LLM in question. Nevertheless, the traditional
understanding of copyright infringement is also being challenged by the
rise of AI technologies. For example, in the ongoing case of
Author’s Guild et al v OpenAI, the US courts are still
evaluating arguments about whether the production of copyrighted works
in outputs should be considered “derivative works”, or merely
replications. Until we have more legal certainty, it is worthwhile to
keep these considerations in mind when using the various LLMs available
to the public.
Having said that, there are also efforts to produce LLMs trained on
“safe” datasets. For example, the datasets used in GPT-NL Public Corpus LLMs from the
Netherlands have been curated for proper compliance, using public and
permissively licensed content to train the GPT-NL model, in an attempt
to establish a model that mitigates the risk of copyright infringement.
Formalized recognition of copyright in assisted works
As we can now see, the general principles of copyright establish that
LLM-generated output is not copyrightable, which therefore places such
work in the public domain. However, if there exists a certain amount of
human-authored creative content together with the generated output, the
overall work can be considered to be LLM-assisted or -modified, and be
considered deserving of copyright under the law. Some jurisdictions are
now taking steps to codify such recognition of copyrightability into
their national laws.
For example, in Italy, a new
law touching on AI technologies entered into force in October 2025.
Specifically, it functions to include the adjective “human” to the
definition of intellectual works in Article 1 of the Italian
Copyright Act of 1941, as well as clarifying that copyright applies
to works of human ingenuity created with the aid of LLMs, provided that
such work can also be established to be the result of the human author’s
own intellectual work.
In Ukraine, legislation also now provides an alternative framework
for LLM-generated and -assisted work, in the form of what is known as a
“sui generis right”. In legal terms, a sui generis
right refers to a type of right that extends to things that are so
unique that they cannot fit into traditional forms of categorisation,
and therefore need to be in their own one-of-a-kind classifications.
In cases of unique output generated by a computer program, Article
33(2) of the Ukrainian Law 2811-IX
on Copyright and Related Rights grants sui generis rights
to the authors of the computer program, their heirs, persons to whom the
authors or their heirs transferred economic rights to the computer
program, or the lawful users of the computer program. Traditional
copyright principles will nevertheless still apply to any parts of the
work that was created by humans.
In other words, Ukrainian law now additionally allows developers or
proprietary owners of LLM systems and their users to enjoy rights
traditionally held by copyright holders, thereby giving them control and
licensing power over such generated content, as well as the ability to
restrict unauthorised use. The Ukrainian courts have thus far not yet
dealt with the nuances of the application of this law, and such rights
have also not been accepted by the majority of member states of the
EU.
Copyfraud and concealment of AI usage
As things currently stand, the existing copyright rules across all
jurisdictions require a great amount of effort for developers to
determine the copyrightability of their work when they use AI-generated
outputs. Additionally, it is currently next to impossible to accurately
determine whether a particular line of code is AI-generated, or whether
it is human written. This creates a situation where the awareness of
whether or not there is AI-generated code in a particular repository
depends almost entirely on the goodwill of its contributors, and how
forthcoming they are.
Because of these factors, some developers might feel encouraged or
pressured to conceal their own uses of AI, in order to avoid
complications in how they license their work, as well as how they
present notices for accurate copyright and licensing information within
their project repositories. It is nevertheless still important to keep
in mind that even if doing so may be easier in the short term, it can
cause serious legal problems downstream. Even if AI-generated content is
not covered by copyright, such content may still infringe on existing
copyrights, in instances when the output resembles or duplicates
copyrighted training data.
Indeed, as noted by the Software Freedom Conservancy (“SFC”), some
Free Software projects' leaders have taken a zero-tolerance approach to
AI-generated contributions to their projects, in order to simplify the
increasingly burdensome responsibilities that maintainers have to
shoulder to analyse incoming contributions for their legal requirements
in light of generative AI.
Recommendations for AI usage in Free Software contributions
With these issues in mind, the recently published Recommendations
When Using LLM-Backed Generative AI Systems for FOSS
Contributions by the SFC outlines some of the legal
difficulties for copyright and licensing that generative systems have
created for software developers who wish to write, maintain, or
contribute to Free Software projects.
Notably for the purposes of this article, the SFC recommends the full
recording and disclosure of how and when an AI tool was used to assist
in the creation of a contribution. As stated by the SFC:
“FOSS project leaders cannot make good decisions about LLM-gen-AI
policy if they cannot survey which contributions were assisted, and how
much they are assisted. Part of the contribution process should (at
least) include a disclosure of what LLM-gen-AI system was used, its
version (as these system change over time), and a brief description of
how the system assisted the contributor. This information should be
included in a machine-readable format in commit logs."
Indeed, such disclosure can be an important foundational step to
allow for the accurate assessment of the copyrightability of code that
has been assisted or generated by AI tools, in order to assess their
licensability into Free Software. Open and clear disclosure is a helpful
step for the Free Software community to maintain a healthy licensing
ecosystem, which is currently threatened by the legal uncertainties that
come with the advent of generative AI.
Additionally, it is worthwhile for developers to document in some
capacity the extent of human work that they have done in their software
projects, whether it be the writing of code, the selection and
arrangement of components within the project, or the extent of human
modification of machine generated content.
Support FSFE
Anmelden
Agentenlose Konfigurationsverwaltung mit Ansible gilt als Alternative zu Puppet, Chef und Saltstack. Ein Workshop vermittelt Playbooks und Rollen für den Praxiseinsatz. (
