# Copyleft

> Source: https://aiwiki.ai/wiki/copyleft
> Updated: 2026-11-18
> Fact-checked: 2026-11-19
> Categories: AI Policy & Regulation, Open Source AI, Software Development
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Copyleft." aiwiki.ai, 18 Nov 2026. https://aiwiki.ai/wiki/copyleft
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**Copyleft** is a licensing method that uses copyright to keep a work, and every modified version of it, available on the same free terms: recipients may use, change and redistribute the work, provided that what they pass on carries the same license.[1][2] The [Free Software Foundation](https://aiwiki.ai/wiki/free_software_foundation) (FSF) describes it as "a general method for making a program (or other work) free (in the sense of freedom, not "zero price"), and requiring all modified and extended versions of the program to be free as well."[1] The [GNU General Public License](https://aiwiki.ai/wiki/gnu_general_public_license) (GPL) calls itself "a free, copyleft license for software and other kinds of works", and the ShareAlike [Creative Commons licenses](https://aiwiki.ai/wiki/creative_commons_licenses) attach a similar condition to adaptations of licensed material.[3][4] Licenses without the condition, such as the [BSD licenses](https://aiwiki.ai/wiki/bsd_licenses), are called non-copyleft or non-reciprocal by the [Open Source Initiative](https://aiwiki.ai/wiki/open_source_initiative) (OSI) and are often called permissive.[2][5]

How copyleft licenses apply to machine learning is unsettled. Whether a model trained on copyleft code or text is a derivative of it, whether the Affero GPL reaches a hosted model, and whether share-alike terms can bind model [weights](https://aiwiki.ai/wiki/weight) are questions on which free software organizations, Creative Commons and legal scholars have published tentative and sometimes conflicting positions. The Software Freedom Conservancy (SFC) wrote in June 2026 that "The question of licensing obligations for material passed through the process called "training" remains undecided."[6][7] This article summarizes license texts, court rulings and published positions; it is not legal advice.

## Origin of the term

The word is older than the GNU licenses. The source listing of Li-Chen Wang's Tiny BASIC for the Intel 8080, printed in volume 1 of *Dr. Dobb's Journal* and dated 10 June 1976, carries the lines "@COPYLEFT" and "ALL WRONGS RESERVED" under the author's name.[8]

Richard Stallman, who started the GNU Project, has written that in 1984 or 1985 Don Hopkins mailed him a letter whose envelope carried several sayings, one of them pairing the word "Copyleft" with the phrase "all rights reversed", and that he used the word to name the distribution concept he was developing at the time.[9] Sam Williams's biography *Free as in Freedom* tells the story with different details: Hopkins was returning a borrowed manual for the 68000 microprocessor and decorated the envelope with stickers from a science-fiction convention, one of which read "Copyleft (L), All Rights Reversed."[10]

Williams traces the legal bargain to the GNU Emacs License, which debuted when Stallman released GNU Emacs in 1985: users could modify the program so long as they published their modifications, and derivative works had to carry the same license.[10] Stallman created the FSF in the same year.[9] Version 1 of the GNU GPL is dated February 1989, version 2 June 1991 and version 3 29 June 2007.[11][12][3]

The FSF explains the name this way: "Proprietary software developers use copyright to take away the users' freedom; we use copyright to guarantee their freedom. That's why we reverse the name, changing "copyright" into "copyleft."" It adds that copyleft does not mean abandoning copyright, that "left" is not a reference to the verb "to leave", and that a backwards C in a circle has no special legal significance.[1]

## How copyleft works

### A license built on copyright

In the FSF's description, "To copyleft a program, we first state that it is copyrighted; then we add distribution terms, which are a legal instrument that gives everyone the rights to use, modify, and redistribute the program's code, or any program derived from it, but only if the distribution terms are unchanged."[1] Stallman's stated reason for not relying on a simple permissive license is that such licenses allow proprietary modified versions. His example is the X Window System, developed at MIT and released under a permissive license, which computer companies then added to their proprietary Unix systems in binary form only.[9]

The operative clause of GPLv2 is section 2(b): "You must cause any work that you distribute or publish, that in whole or in part contains or is derived from the Program or any part thereof, to be licensed as a whole at no charge to all third parties under the terms of this License."[12] GPLv3 section 5(c) says: "You must license the entire work, as a whole, under this License to anyone who comes into possession of a copy."[3] The FSF notes that copyleft licenses generally require source code to be made available and do not let others add requirements.[13] GPLv3 defines source code as "the preferred form of the work for making modifications to it" and states: "You may not impose any further restrictions on the exercise of the rights granted or affirmed under this License."[3] In April 2026 the FSF applied that rule to OnlyOffice, which had added a term requiring licensees to "retain the original Product logo" to software under the Affero GPL; the FSF called the term a further restriction that licensees are permitted to remove.[14]

### Acceptance, termination and enforcement

GPLv2 explains why a licensee is bound without signing anything: "You are not required to accept this License, since you have not signed it. However, nothing else grants you permission to modify or distribute the Program or its derivative works."[12] Under GPLv2 section 4, an attempt to copy, modify, sublicense or distribute the program outside the license "will automatically terminate your rights under this License".[12] GPLv3 keeps automatic termination but reinstates the license provisionally when the violation stops, permanently if the copyright holder gives no notice within 60 days after it stops, and permanently for a first-time violator who cures within 30 days of receiving notice.[3]

Each recipient of a GPLv3 work "automatically receives a license from the original licensors".[3] On enforcement, the FSF's FAQ says: "Since the GPL is a copyright license, it can be enforced by the copyright holders of the software."[15]

### What the condition does not reach

The GPL does not require anyone to release a modified version. The FSF's FAQ says an organization can make a modified version and use it internally without ever releasing it; the source obligation arises when the modified version is released to the public.[15]

The condition applies to derivatives, not to unrelated works stored or shipped alongside. GPLv2 states that "mere aggregation of another work not based on the Program with the Program" on a storage or distribution medium does not bring the other work under the license.[12] The OSI's FAQ puts it this way: "Copyleft provisions apply only to actual derivatives".[2] Where an aggregate ends and a single combined program begins is, in the words of the FSF's FAQ, "a legal question, which ultimately judges will decide."[15]

According to the FSF's FAQ, the output of a copylefted program is generally not covered either. Asked whether a developer can require that a program's output be GPL-licensed, the FAQ answers that "copyright law does not give you any say in the use of the output people make from their data using your program", with an exception for output that copies substantial text from the program itself. It adds that when a program translates its input into some other form, the copyright status of the output inherits that of the input.[15] These limits come up again in the AI questions below.

## Strong, weak and library copyleft

Copyleft licenses differ in how far the same-terms condition extends. Mozilla's FAQ for the Mozilla Public License (MPL) summarizes three scopes: "MPL: The copyleft applies to any files containing MPLed code. LGPL: The copyleft applies to any library based on LGPLed code. GPL: The copyleft applies to all software based on GPLed code."[16] The FSF's license list uses the terms "strong copyleft", "weak copyleft" and "partial copyleft" for these gradations.[17]

| License | Steward | Scope of the same-terms condition | Version cited |
|---|---|---|---|
| GNU GPL | FSF | All software based on the covered code[16] | 3 (June 2007)[3] |
| GNU Affero GPL (AGPL) | FSF | As GPLv3, plus a source offer to remote network users of a modified version[18][19] | 3 (November 2007)[18] |
| GNU Lesser GPL (LGPL) | FSF | The library itself; linking with nonfree modules is permitted[17] | 3 (June 2007)[20] |
| Mozilla Public License | Mozilla Foundation | Files containing MPL-licensed code[16] | 2.0[21] |
| GNU Free Documentation License | FSF | Derivative works of a document[22] | 1.3 (November 2008)[22] |
| CC BY-SA | Creative Commons | Adaptations of the licensed material, when shared[4][23] | 4.0 (November 2013)[24] |
| Open Database License | Open Data Commons | Publicly used derivative databases[25] | 1.0[25] |
| CDLA-Sharing | [Linux Foundation](https://aiwiki.ai/wiki/linux_foundation) | Published data and enhanced data, but not computational results[26] | 1.0 (announced October 2017)[27] |

The AGPL is the GPL with one added requirement, written for programs that run on servers. The FSF's rationale is that a developer could otherwise modify a GPL program, run it on a server for the public and never release the changes.[19] Section 13 of AGPLv3 reads: "if you modify the Program, your modified version must prominently offer all users interacting with it remotely through a computer network (if your version supports such interaction) an opportunity to receive the Corresponding Source of your version by providing access to the Corresponding Source from a network server at no charge".[18] The FSF also says the AGPL "does not address the problem of Service as a Software Substitute (SaaSS)."[19]

The FSF describes the LGPL as a "compromise form of copyleft" and says it is "not a strong copyleft license, because it permits linking with nonfree modules."[1][17] Its advice to library authors is that the choice is strategic: the Lesser GPL permits use of a library in proprietary programs, while the ordinary GPL makes it available only for free programs.[28] The MPL's "file-level" copyleft is narrower still: new files that contain no MPL-licensed code are not covered, and MPL code can be combined into a larger work distributed under terms of the distributor's choice, provided the MPL's requirements are still met for the covered files.[16][21]

## Share-alike licenses for content and data

Creative Commons published its first licenses in December 2002 and version 4.0 in November 2013.[24] Creative Commons says its ShareAlike licenses "are known as "copyleft" licenses".[29] Section 3(b) of CC BY-SA 4.0 says of anyone who shares adapted material: "The Adapter's License You apply must be a Creative Commons license with the same License Elements, this version or later, or a BY-SA Compatible License."[4] Creative Commons has declared two outside licenses compatible with BY-SA 4.0: the Free Art License 1.3, on 21 October 2014, and GPLv3, on 8 October 2015. The GPLv3 compatibility is one-way only, meaning BY-SA material can be adapted into a GPLv3 work but not the reverse.[30] Under the Wikimedia Foundation's Terms of Use, contributors license the text they submit to Wikimedia projects under both CC BY-SA 4.0 and the GNU Free Documentation License, unless a project or feature requires a different license.[31]

The FSF's own license for manuals says of itself: "This License is a kind of "copyleft", which means that derivative works of the document must themselves be free in the same sense."[22]

Two licenses carry the idea to data. The Open Database License has a "Share alike" clause for derivative databases that are publicly used.[25] The FSF classifies it as a free and copyleft license for data but does not recommend it, citing "inconvenient requirements about signing contracts which try to create an effect like copyleft for data that is not copyrightable".[17] The Linux Foundation's Community Data License Agreement, announced on 23 October 2017 in a Sharing and a Permissive version, was presented partly for people "looking for datasets to help kick off training an AI system".[27] CDLA-Sharing-1.0 requires published data and enhanced data to stay under the same agreement, but it defines the "Results" of computational use separately and says: "This Agreement imposes no obligations or restrictions on Your Use or Publication of Results."[26]

## Enforcement in court

| Year | Dispute | Court | Outcome |
|---|---|---|---|
| 2004 | netfilter/iptables v. Sitecom | Landgericht München I, Germany | Sitecom's appeal against a preliminary injunction, obtained over its distribution of the software without complying with the GPL, was turned down, as reported by Harald Welte on gpl-violations.org[32] |
| 2017 | Artifex Software v. Hancom | US District Court, N.D. California | Motion to dismiss contract and copyright claims over Ghostscript denied[33] |
| 2021-2022 | Stockfish v. ChessBase | District Court Munich I, Germany | Settled; ChessBase agreed to stop selling products containing Stockfish code[34] |
| 2021- | Software Freedom Conservancy v. Vizio | Superior Court of California, Orange County | Summary judgment for Vizio denied in December 2023; listed by SFC as ongoing[35][36] |

In *Artifex*, decided on 25 April 2017, Magistrate Judge Jacqueline Scott Corley denied Hancom's motion to dismiss. She rejected its argument that reliance on an unsigned GNU GPL could not plausibly show the existence of a contract, and its argument that the contract claim was preempted by copyright law.[33] A US appellate court had earlier addressed open-source license conditions in *Jacobsen v. Katzer* (Federal Circuit, 13 August 2008), which concerned the Artistic License and held its terms to be "enforceable copyright conditions"; the [open-source license](https://aiwiki.ai/wiki/open_source_license) article covers that case.[37]

The Stockfish settlement extended to neural networks. The developers of the GPLv3 chess engine sued ChessBase in 2021 to enforce the license's termination clause, saying its Fat Fritz 2 and Houdini 6 products were Stockfish derivatives.[38] Under the settlement announced on 18 November 2022, ChessBase agreed to stop selling products containing Stockfish code, with its license to be reinstated after one year, and to appoint a Free Software Compliance Officer. The Stockfish team also reported that "neural networks offered by ChessBase for use with Stockfish will be provided under GPL or a compatible license".[34]

*SFC v. Vizio*, filed in October 2021, tests whether a purchaser of a device, rather than a copyright holder, can enforce the GPL as a contract. On 29 December 2023 Judge Sandy Leal denied Vizio's motion for summary judgment, ruling that "Plaintiff's claim for breach of contract is not preempted by the Copyright Act".[35] As of November 2026, SFC's case page, last updated on 27 February 2026, listed the case as ongoing with a trial date in late 2026.[36] The GPL article describes the later rulings.

## Adoption and objections

RedMonk analyst Stephen O'Grady wrote in March 2026 that the industry has been in a long-term shift from copyleft to permissive licenses and estimated that permissive licenses became the majority at some point between 2014 and 2017. He also reported that the permissive share in his GitHub sample fell from 82% in 2022 to 73% in 2025, with the caveat that this could be a sampling issue, and noted that Elastic and Redis had moved from [source-available](https://aiwiki.ai/wiki/source_available) terms back to an OSI-approved license by adopting the AGPL.[5]

Some organizations restrict copyleft code in their own products. Google's open-source documentation states that AGPL-licensed code must not be used at Google, describing the license's network terms as "extremely difficult for Google to comply with".[39] The Apache Software Foundation places the GNU GPL, AGPL and LGPL in its "Category X" of licenses that may not be included in Apache products.[40]

Stallman's defense of the mechanism is practical. In "Copyleft: Pragmatic Idealism" he argues that GNU C++ exists as free software only because the GPL required it: the consortium that wrote the C++ front end normally kept its work proprietary but had to release code that was meant to be linked with GCC under the GPL.[41] The OSI objects to one common piece of vocabulary: its FAQ says that using "permissive" as the opposite of "copyleft" is best avoided because it "misrepresents copyleft conditions as restrictions."[2]

## Copyleft and artificial intelligence

### AI software under copyleft licenses

The AI programs in the table below are distributed under copyleft licenses. The licenses were read from each repository's license file in November 2026.

| Project | What it is | License |
|---|---|---|
| Ultralytics [YOLO](https://aiwiki.ai/wiki/yolo) | Object detection and other computer vision models and library | AGPL-3.0[42] |
| [AUTOMATIC1111](https://aiwiki.ai/wiki/automatic1111) Stable Diffusion web UI | Web interface for [Stable Diffusion](https://aiwiki.ai/wiki/stable_diffusion) | AGPL-3.0[43] |
| [ComfyUI](https://aiwiki.ai/wiki/comfyui) | Node-graph interface and back end for diffusion models | GPL-3.0[44] |
| TextGen (formerly [text-generation-webui](https://aiwiki.ai/wiki/text_generation_webui)) | Desktop app for local LLMs | AGPL-3.0[45] |
| [KoboldCpp](https://aiwiki.ai/wiki/koboldcpp) | Runs GGUF models locally with a KoboldAI interface | AGPL-3.0[46] |
| SillyTavern | LLM front end | AGPL-3.0[47] |
| Fooocus | Image generation software | GPL-3.0[48] |
| Stockfish | Chess engine | GPL-3.0[49] |
| Leela Chess Zero (lc0) | Neural network chess engine | GPL-3.0[50] |
| Coqui TTS | Deep learning toolkit for text-to-speech | MPL-2.0[51] |
| Mozilla DeepSpeech (archived) | Speech-to-text engine | MPL-2.0[52] |

The [Hugging Face](https://aiwiki.ai/wiki/hugging_face) Hub also accepts copyleft identifiers in model and dataset metadata, including agpl-3.0, gpl-3.0, lgpl-3.0, mpl-2.0, cc-by-sa-4.0, cdla-sharing-1.0 and odbl.[53]

### Whether trained models are covered

GPLv2 defines a "work based on the Program" as "either the Program or any derivative work under copyright law", tying the license's reach to copyright law.[12] Published positions differ on whether a trained model is such a work.

- **Software Freedom Conservancy.** In February 2022 SFC's Bradley M. Kühn wrote of [GitHub Copilot](https://aiwiki.ai/wiki/github_copilot) that GitHub had avoided "the question of whether the trained model is a "work based on" the input. We contend that it probably is." He added that "only a court decision can settle the matter adequately."[54] A committee SFC convened in 2022 produced an aspirational statement in October 2024 whose principle is "FOSS in, FOSS out, and FOSS throughout": all software, weights and training data under a free and open-source license. SFC said the statement was not meant to be prescriptive or definitional.[55]
- **Free Software Foundation.** In July 2021 the FSF called Copilot "unacceptable and unjust" because it requires nonfree software and is a service substituting for software, and it listed the legal questions as open, among them whether a trained model is "a compiled version of the training data, or is it something else".[56] In February 2022 it published five of the 22 white papers submitted, while stating that it might or might not agree with them and was not elaborating its own developing positions.[57] In October 2024 it said that an ML application cannot be called free unless its training data and processing scripts respect all users, while noting that "training data is not the "source code" of model parameters in the usual sense."[58] In March 2026, after learning that *Free as in Freedom*, published under the GNU Free Documentation License, was in datasets at issue in [Bartz v. Anthropic](https://aiwiki.ai/wiki/bartz_v_anthropic), the FSF wrote that if it took part in such a lawsuit and found its license violated, "we would certainly request user freedom as compensation."[59]
- **Open Source Initiative.** The [Open Source AI Definition](https://aiwiki.ai/wiki/open_source_ai_definition) permits terms that "require any modified version to be released under the same terms as the original", and its FAQ says the definition "does not take a stance on the legal nature of Parameters".[60][61]
- **Creative Commons.** In August 2023 Creative Commons wrote that "there are strong arguments that, in most cases, using copyrighted works to train generative AI models would be fair use in the United States", and that where training is covered by an exception such as fair use, the user does not need to rely on the CC license.[62] Its 2025 guidance says that "in many cases, neither the AI model nor its outputs would be considered to be derivative works of training data under copyright law".[63]
- **US Copyright Office.** Its May 2025 pre-publication report on generative AI training says: "Model weights that have memorized protectable expression from training data may also infringe the derivative work right." It adds that courts addressing infringement claims about model weights "have reached varying conclusions."[64]
- **Scholars.** Shanklin and co-authors at Yale write that, on the assumption that training is not [fair use](https://aiwiki.ai/wiki/fair_use), generative AI models trained on code "could plausibly be considered derivative works of the training data, although this interpretation remains legally unsettled", and that if training is fair use, a license's conditions could not apply to the model.[65] Peter Henderson and Mark Lemley argue that model weights and outputs "are largely not copyrightable, making it unclear whether there is even anything to be licensed"; their paper concerns the terms of use attached to AI models, not copyleft specifically.[66]

Two European rulings from November 2025 addressed whether weights contain copies of training works, in cases that did not involve copyleft licenses. The Landgericht München I, in a suit by the collecting society GEMA against [OpenAI](https://aiwiki.ai/wiki/openai), held that song lyrics reproducibly contained in two models were reproductions fixed in the model parameters; the court's press release notes the judgment was not final.[67] In [Getty Images v. Stability AI](https://aiwiki.ai/wiki/getty_v_stability_ai), the High Court in London held that "the model weights are not themselves an infringing copy and they do not store an infringing copy".[68]

### The Copilot litigation

*Doe v. GitHub* is a proposed class action filed on 3 November 2022 in federal court in San Francisco against [GitHub](https://aiwiki.ai/wiki/github), [Microsoft](https://aiwiki.ai/wiki/microsoft) and OpenAI over Copilot and [OpenAI Codex](https://aiwiki.ai/wiki/openai_codex). The plaintiffs' lawyers described it as alleging violation of "A set of 11 popular open-source licenses that all require attribution of the author's name and copyright, including the MIT license, the GPL, and the Apache license".[69] The allegation quoted concerns attribution, and the list includes licenses that are not copyleft: the FSF classes the Expat license, which it notes some people call the MIT License, as non-copyleft.[17]

On 16 September 2026 the US Court of Appeals for the Ninth Circuit affirmed the dismissal of the plaintiffs' claims under section 1202(b) of the Digital Millennium Copyright Act. Judge Miller wrote that Copilot and Codex "do not "remove or alter" copyright management information (CMI) from a copy of an existing protected work but instead create new works that never contained that information."[70] The court noted that the complaint had been "whittled down to three claims: one claim under the DMCA and two claims for breach of contract", and that the district court had denied a motion to dismiss the contract claims, "which remain pending." It treated a theory based on the training input as forfeited, and on output that resembles existing code it wrote: "We express no view on whether that similarity would allow plaintiffs to assert a claim for copyright infringement."[70] The opinion does not decide whether training on GPL-licensed code, or emitting code similar to it, breaches the license.[70]

### AGPL and hosted models

The FSF's 2021 list of open questions included: "If Copilot learns from AGPL-covered code, is Copilot infringing the AGPL?"[56]

Ultralytics, which licenses its YOLO software under AGPL-3.0 and sells an Enterprise License as the alternative, takes a broad position on its own licensing page: "All Ultralytics YOLO trained models fall under the AGPL-3.0 License by default. The AGPL-3.0 License covers the training code and the models produced by that training code." The page says an Enterprise License is needed for "SaaS platforms, APIs, or cloud systems that use YOLO behind the scenes" unless the whole project is open-sourced.[71] That is the licensor's reading. The FSF's general FAQ, quoted above, says the output of a program is not in general covered by the copyright on the program's code, and it describes the AGPL's remote-interaction clause as aimed at programs "expressly designed to accept user requests and send responses over a network". The FAQ does not discuss trained models.[15]

### Share-alike training data and model weights

Creative Commons' 2025 guidance on AI training says there are many scenarios in which compliance with license conditions is not required: the licenses have no effect where copyright law requires no permission, and the attribution, ShareAlike and NoDerivatives conditions are triggered only by publicly sharing the work or an adaptation of it. For those who want a conservative approach, it says that where training data is under ShareAlike, "model outputs and the model itself, if shared publicly, should be made available under the same CC license as the original works".[63] A preliminary analysis of [ChatGPT](https://aiwiki.ai/wiki/chatgpt) published by the Wikimedia Foundation's legal team said that "it is not clear yet whether massively copying content from these sources may result in a violation of the Creative Commons license if attribution is not granted".[72]

Share-alike licenses are applied to datasets and to some model weights. On Hugging Face, as of November 2026, the databricks-dolly-15k dataset from [Databricks](https://aiwiki.ai/wiki/databricks) is tagged cc-by-sa-3.0 and Wikimedia's Wikipedia dataset cc-by-sa-3.0 and gfdl.[73][74] [Stability AI](https://aiwiki.ai/wiki/stability_ai) released its StableLM-Base-Alpha checkpoints under CC BY-SA 4.0, with the accompanying code under the [Apache License](https://aiwiki.ai/wiki/apache_license) 2.0.[75] CDLA-Sharing, by contrast, exempts the results of computational use of the data from its share-alike terms.[26]

### Copyleft-style licenses written for models

The [OpenRAIL](https://aiwiki.ai/wiki/openrail) licenses apply a similar propagation mechanism to use restrictions. The Hugging Face post that introduced them describes "copyleft-style behavioral-use clauses" under which derivatives must carry at least the same use restrictions, and the RAIL FAQ compares this to copyleft and share-alike clauses while stating that the licenses are not open source.[76][77] The FSF called RAIL licenses "nonfree and unethical" in April 2026 because they restrict how software may be used.[78]

ModelGo, a license family anchored at the NUS Institute of Data Science, includes MG-BY-SA, a share-alike model license.[79] Version 2.0 defines the licensed model to include learned weights, treats offering the model as a hosted service or through an API as distribution, and requires licensees to "license Derivative Materials and Extracted Models under the terms and conditions of this License", where extracted models include those made by [distillation](https://aiwiki.ai/wiki/knowledge_distillation) from the model's output.[80] The project reported in February 2026 that it had withdrawn MG-BY-SA from its OSI submission "as it may not be compatible with the Open Source Definition."[79]

In a July 2025 paper, Shanklin and co-authors proposed a Contextual Copyleft AI (CCAI) license for code, under which "Any verbatim copy, modified version, derivative work, or model trained on this code must also be wholly licensed under CCAI", with a network clause modeled on the AGPL.[65] SFC's June 2026 recommendations say it will support the copyleft-next project in producing a license "extremely suitable as a copyleft for LLM-gen-AI outputs."[6]

### Copyleft code in training sets and outputs

SFC reported in 2022 that, by GitHub's own account, Copilot's model had encountered a copy of the GPL more than 700,000 times in training.[54] Academic studies have since examined copyleft code in training data and model output:

- Katzy and colleagues examined the datasets of 53 code [language models](https://aiwiki.ai/wiki/large_language_model) and found that "every dataset we examined contained license inconsistencies"; of 514 million code files, 38 million were exact duplicates of files in their strong-copyleft dataset.[81]
- The LiCoEval benchmark tested 14 models and found that even the top performers produced code strikingly similar to existing open-source implementations in 0.88% to 2.01% of cases, and that "most LLMs fail to provide accurate license information, particularly for code under copyleft licenses."[82]
- Colombo and colleagues generated more than 70,000 method implementations with ChatGPT and found that a larger context increases the likelihood of reproducing copyleft code, while higher temperature settings can mitigate it.[83]
- A study of 1.86 million models on Hugging Face found that licenses in [fine-tuning](https://aiwiki.ai/wiki/fine_tuning) lineages drift "towards permissive or copyleft licenses, often in violation of upstream license's terms".[84]

### AI-assisted rewrites

Coding agents have raised a further question: whether code regenerated by a model escapes the original license. In March 2026 version 7.0.0 of the chardet library was released as a "Ground-up, MIT-licensed rewrite" of a project previously under the LGPL.[85] A GitHub user who introduced himself as Mark Pilgrim, the library's original author, objected in the project's issue tracker that the maintainers had no right to relicense: "Adding a fancy code generator into the mix does not somehow grant them any additional rights." Maintainer Dan Blanchard replied that he had started in an empty repository, had instructed [Claude](https://aiwiki.ai/wiki/claude) not to base anything on LGPL or GPL code, and that the JPlag similarity tool showed at most 1.29% similarity between version 7.0.0 and the previous release.[86] In a separate issue about which license can cover AI-generated code, Richard Fontana wrote that he did not "currently see any basis for concluding that chardet 7.0.0 is required to be released under the LGPL".[88] Blanchard, who described Fontana as one of the authors of LGPLv3, cited that comment when he closed Pilgrim's issue on 26 March 2026 without restoring the LGPL.[86] On 22 March the project had changed the license of the 7.x series from MIT to the Zero-Clause BSD license (0BSD).[89]

Masayuki Hatta, in a paper accepted at AGI-26, argues that capable AI systems can rewrite licensed source into functionally equivalent code "stripped of their original obligations, a form of laundering against which copyleft has no effective defense", and proposes reproducible builds as a substitute guarantee.[87] SFC's June 2026 recommendations advise contributors to "Avoid jumping to conclusions about the legal significance of generated contributions" while the legal questions remain unanswered.[6]

## See also

- [GNU General Public License](https://aiwiki.ai/wiki/gnu_general_public_license)
- [Creative Commons licenses](https://aiwiki.ai/wiki/creative_commons_licenses)
- [Open Source Initiative](https://aiwiki.ai/wiki/open_source_initiative)
- [Open Source AI Definition](https://aiwiki.ai/wiki/open_source_ai_definition)
- [OpenRAIL](https://aiwiki.ai/wiki/openrail)
- [Open-Weight LLM License Comparison](https://aiwiki.ai/wiki/open_weight_license_comparison)
- [AI and Copyright](https://aiwiki.ai/wiki/ai_copyright)

## References

1. Free Software Foundation. "What is Copyleft?". GNU Project, updated 2022-01-02. https://www.gnu.org/licenses/copyleft.html
2. Open Source Initiative. "Frequently Answered Questions". opensource.org, accessed November 2026. https://opensource.org/faq
3. Free Software Foundation. "GNU General Public License, version 3". GNU Project, 29 June 2007. https://www.gnu.org/licenses/gpl-3.0.html
4. Creative Commons. "Attribution-ShareAlike 4.0 International: Legal Code". creativecommons.org. https://creativecommons.org/licenses/by-sa/4.0/legalcode.en
5. Stephen O'Grady. "The State of Open Source Licensing in 2026". RedMonk tecosystems, March 25, 2026. https://redmonk.com/sogrady/2026/03/25/open-source-licensing-2026/
6. Software Freedom Conservancy. "Recommendations When Using LLM-backed Generative AI Systems for FOSS Contributions". sfconservancy.org, June 2026. https://sfconservancy.org/llm-gen-ai/llm-backed-generative-ai-recommendations.html
7. Software Freedom Conservancy. "LLM Backed Generative AI Recommendations". sfconservancy.org, June 18, 2026. https://sfconservancy.org/news/2026/jun/18/llm-backed-generative-ai-recommendations/
8. Li-Chen Wang. "Palo Alto Tiny BASIC". Dr. Dobb's Journal of Computer Calisthenics & Orthodontia, vol. 1, no. 5, 1976 (scan of volume 1 at the Internet Archive). https://archive.org/details/dr_dobbs_journal_vol_01
9. Richard Stallman. "The GNU Project". GNU Project, updated 2025-01-20. https://www.gnu.org/gnu/thegnuproject.html
10. Sam Williams. "Free as in Freedom: Richard Stallman's Crusade for Free Software", chapter 9, "The GNU General Public License". O'Reilly, 2002. https://www.oreilly.com/openbook/freedom/ch09.html
11. Free Software Foundation. "GNU General Public License, version 1". GNU Project, February 1989. https://www.gnu.org/licenses/old-licenses/gpl-1.0.html
12. Free Software Foundation. "GNU General Public License, version 2". GNU Project, June 1991. https://www.gnu.org/licenses/old-licenses/gpl-2.0.html
13. Free Software Foundation. "Categories of Free and Nonfree Software". GNU Project, updated 2025-06-05. https://www.gnu.org/philosophy/categories.html
14. Krzysztof Siewicz. "You cannot use the GNU (A)GPL to take software freedom away". Free Software Foundation, April 15, 2026. https://www.fsf.org/blogs/licensing/agpl-is-not-a-tool-for-taking-freedom-away
15. Free Software Foundation. "Frequently Asked Questions about the GNU Licenses". GNU Project, updated 2026-01-13. https://www.gnu.org/licenses/gpl-faq.html
16. Mozilla. "MPL 2.0 FAQ". mozilla.org, accessed November 2026. https://www.mozilla.org/en-US/MPL/2.0/FAQ/
17. Free Software Foundation. "Various Licenses and Comments about Them". GNU Project, updated 2026-04-22. https://www.gnu.org/licenses/license-list.html
18. Free Software Foundation. "GNU Affero General Public License, version 3". GNU Project, 19 November 2007. https://www.gnu.org/licenses/agpl-3.0.html
19. Free Software Foundation. "Why the GNU Affero GPL". GNU Project, updated 2022-01-02. https://www.gnu.org/licenses/why-affero-gpl.html
20. Free Software Foundation. "GNU Lesser General Public License, version 3". GNU Project, 29 June 2007. https://www.gnu.org/licenses/lgpl-3.0.html
21. Mozilla. "Mozilla Public License, version 2.0". mozilla.org. https://www.mozilla.org/en-US/MPL/2.0/
22. Free Software Foundation. "GNU Free Documentation License v1.3". GNU Project, 3 November 2008. https://www.gnu.org/licenses/fdl-1.3.html
23. Creative Commons. "Frequently Asked Questions". creativecommons.org, accessed November 2026. https://creativecommons.org/faq/
24. Creative Commons. "License Versions". CC Public Wiki, last edited 4 January 2016. https://wiki.creativecommons.org/wiki/License_Versions
25. Open Data Commons. "Open Data Commons Open Database License (ODbL) v1.0". opendatacommons.org. https://opendatacommons.org/licenses/odbl/1-0/
26. Linux Foundation. "Community Data License Agreement - Sharing, Version 1.0". cdla.dev. https://cdla.dev/sharing-1-0/
27. Linux Foundation. "Linux Foundation Debuts Community Data License Agreement". Press release, October 23, 2017. https://www.linuxfoundation.org/press/press-release/linux-foundation-debuts-community-data-license-agreement
28. Free Software Foundation. "Why you shouldn't use the Lesser GPL for your next library". GNU Project, updated 2022-01-02. https://www.gnu.org/licenses/why-not-lgpl.html
29. Creative Commons. "ShareAlike interpretation". CC Public Wiki, last edited 30 October 2014. https://wiki.creativecommons.org/wiki/ShareAlike_interpretation
30. Creative Commons. "Compatible Licenses". creativecommons.org, accessed November 2026. https://creativecommons.org/share-your-work/licensing-considerations/compatible-licenses/
31. Wikimedia Foundation. "Wikimedia Foundation Terms of Use". Wikimedia Foundation Governance Wiki, effective June 7, 2023. https://foundation.wikimedia.org/wiki/Policy:Terms_of_Use
32. Harald Welte. "Sitecom appeal against preliminary injunction turned down". gpl-violations.org, 2004-05-19. https://gpl-violations.org/news/20040519-iptables-sitecom/
33. United States District Court, Northern District of California. "Artifex Software, Inc. v. Hancom, Inc., Case No. 16-cv-06982-JSC, Order re: Defendant's Motion to Dismiss". April 25, 2017. https://www.govinfo.gov/content/pkg/USCOURTS-cand-3_16-cv-06982/pdf/USCOURTS-cand-3_16-cv-06982-0.pdf
34. Stockfish. "ChessBase GmbH and the Stockfish team reach an agreement and end their legal dispute". Stockfish blog, 2022-11-18. https://stockfishchess.org/blog/2022/chessbase-stockfish-agreement/
35. Superior Court of California, County of Orange. "Software Freedom Conservancy, Inc. vs. Vizio, Inc., Case No. 30-2021-01226723-CU-BC-CJC, Minute Order". December 29, 2023. https://sfconservancy.org/docs/Order_Denying_Vizio_Motion_for_Summary_Judgement_12-29-23.pdf
36. Software Freedom Conservancy. "Software Freedom Conservancy v. Vizio" (case page). sfconservancy.org, last updated February 27, 2026. https://sfconservancy.org/copyleft-compliance/vizio.html
37. United States Court of Appeals for the Federal Circuit. "Jacobsen v. Katzer, No. 2008-1001". August 13, 2008. https://www.cafc.uscourts.gov/opinions-orders/08-1001.pdf
38. Stockfish. "Our lawsuit against ChessBase". Stockfish blog, 2021-07-20. https://stockfishchess.org/blog/2021/our-lawsuit-against-chessbase/
39. Google. "AGPL Policy". Google Open Source documentation, accessed November 2026. https://opensource.google/documentation/reference/using/agpl-policy
40. Apache Software Foundation. "ASF 3rd Party License Policy". apache.org, accessed November 2026. https://www.apache.org/legal/resolved.html
41. Richard Stallman. "Copyleft: Pragmatic Idealism". GNU Project, updated 2021-08-28. https://www.gnu.org/philosophy/pragmatic.html
42. Ultralytics. "LICENSE" (GNU Affero General Public License, version 3). GitHub repository ultralytics/ultralytics, accessed November 2026. https://github.com/ultralytics/ultralytics/blob/main/LICENSE
43. AUTOMATIC1111. "LICENSE.txt" (GNU Affero General Public License, version 3). GitHub repository AUTOMATIC1111/stable-diffusion-webui, accessed November 2026. https://github.com/AUTOMATIC1111/stable-diffusion-webui/blob/master/LICENSE.txt
44. Comfy-Org. "LICENSE" (GNU General Public License, version 3). GitHub repository Comfy-Org/ComfyUI, accessed November 2026. https://github.com/Comfy-Org/ComfyUI/blob/master/LICENSE
45. oobabooga. "LICENSE" (GNU Affero General Public License, version 3). GitHub repository oobabooga/textgen, accessed November 2026. https://github.com/oobabooga/textgen/blob/main/LICENSE
46. LostRuins. "LICENSE.md" (GNU Affero General Public License, version 3). GitHub repository LostRuins/koboldcpp, accessed November 2026. https://github.com/LostRuins/koboldcpp/blob/concedo/LICENSE.md
47. SillyTavern. "LICENSE" (GNU Affero General Public License, version 3). GitHub repository SillyTavern/SillyTavern, accessed November 2026. https://github.com/SillyTavern/SillyTavern/blob/release/LICENSE
48. lllyasviel. "LICENSE" (GNU General Public License, version 3). GitHub repository lllyasviel/Fooocus, accessed November 2026. https://github.com/lllyasviel/Fooocus/blob/main/LICENSE
49. Stockfish. "Copying.txt" (GNU General Public License, version 3). GitHub repository official-stockfish/Stockfish, accessed November 2026. https://github.com/official-stockfish/Stockfish/blob/master/Copying.txt
50. Leela Chess Zero. "COPYING" (GNU General Public License, version 3). GitHub repository LeelaChessZero/lc0, accessed November 2026. https://github.com/LeelaChessZero/lc0/blob/master/COPYING
51. Coqui. "LICENSE.txt" (Mozilla Public License, version 2.0). GitHub repository coqui-ai/TTS, accessed November 2026. https://github.com/coqui-ai/TTS/blob/dev/LICENSE.txt
52. Mozilla. "LICENSE" (Mozilla Public License, version 2.0). GitHub repository mozilla/DeepSpeech, accessed November 2026. https://github.com/mozilla/DeepSpeech/blob/master/LICENSE
53. Hugging Face. "Licenses". Hugging Face Hub documentation, accessed November 2026. https://huggingface.co/docs/hub/repositories-licenses
54. Bradley M. Kühn. "If Software is My Copilot, Who Programmed My Software?". Conservancy Blog, Software Freedom Conservancy, February 3, 2022. https://sfconservancy.org/blog/2022/feb/03/github-copilot-copyleft-gpl/
55. Software Freedom Conservancy. "SFC Announces Aspirational Statement on LLM-backed generative AI for Programming". sfconservancy.org, October 25, 2024. https://sfconservancy.org/news/2024/oct/25/aspirational-on-llm-generative-ai-programming/
56. Donald Robertson. "FSF-funded call for white papers on philosophical and legal questions around Copilot". Free Software Foundation, July 28, 2021. https://www.fsf.org/blogs/licensing/fsf-funded-call-for-white-papers-on-philosophical-and-legal-questions-around-copilot
57. Craig Topham. "Publication of the FSF-funded white papers on questions around Copilot". Free Software Foundation, February 24, 2022. https://www.fsf.org/news/publication-of-the-fsf-funded-white-papers-on-questions-around-copilot
58. Free Software Foundation. "FSF is working on freedom in machine learning applications". fsf.org, October 22, 2024. https://www.fsf.org/news/fsf-is-working-on-freedom-in-machine-learning-applications
59. Krzysztof Siewicz. "The FSF doesn't usually sue for copyright infringement, but when we do, we settle for freedom". Free Software Foundation, March 13, 2026. https://www.fsf.org/blogs/licensing/2026-anthropic-settlement
60. Open Source Initiative. "The Open Source AI Definition 1.0". opensource.org. https://opensource.org/ai/open-source-ai-definition
61. Open Source Initiative. "OSAID FAQs". opensource.org, accessed November 2026. https://opensource.org/ai/faq
62. Kat Walsh. "Understanding CC Licenses and Generative AI". Creative Commons, August 18, 2023. https://creativecommons.org/2023/08/18/understanding-cc-licenses-and-generative-ai/
63. Creative Commons. "Using CC-licensed Works for AI Training". creativecommons.org, 2025. https://creativecommons.org/wp-content/uploads/2025/05/Using-CC-licensed-Works-for-AI-Training.pdf
64. U.S. Copyright Office. "Copyright and Artificial Intelligence, Part 3: Generative AI Training (Pre-Publication Version)". May 2025. https://www.copyright.gov/ai/Copyright-and-Artificial-Intelligence-Part-3-Generative-AI-Training-Report-Pre-Publication-Version.pdf
65. Grant Shanklin, Emmie Hine, Claudio Novelli, Tyler Schroder, Luciano Floridi. "The Case for Contextual Copyleft: Licensing Open Source Training Data and Generative AI". arXiv:2507.12713, July 17, 2025. https://arxiv.org/abs/2507.12713
66. Peter Henderson, Mark A. Lemley. "The Mirage of Artificial Intelligence Terms of Use Restrictions". arXiv:2412.07066, December 10, 2024 (forthcoming, Indiana Law Journal). https://arxiv.org/abs/2412.07066
67. Landgericht München I. "Urteil GEMA gegen Open AI". Pressemitteilung 11/2025, 11 November 2025. https://www.justiz.bayern.de/gerichte-und-behoerden/landgericht/muenchen-1/presse/2025/11.php
68. High Court of Justice, Business and Property Courts of England and Wales. "Getty Images (US) Inc and others v Stability AI Limited, [2025] EWHC 2863 (Ch)". 4 November 2025. https://www.judiciary.uk/wp-content/uploads/2025/11/Getty-Images-v-Stability-AI.pdf
69. Matthew Butterick and Joseph Saveri Law Firm. "GitHub Copilot litigation". githubcopilotlitigation.com, November 3, 2022. https://githubcopilotlitigation.com/
70. United States Court of Appeals for the Ninth Circuit. "Doe v. GitHub, Inc., No. 24-7700, Opinion". September 16, 2026. https://cdn.ca9.uscourts.gov/datastore/opinions/2026/09/16/24-7700.pdf
71. Ultralytics. "Ultralytics License". ultralytics.com, accessed November 2026. https://www.ultralytics.com/license
72. Wikimedia Foundation. "Wikilegal/Copyright Analysis of ChatGPT". Meta-Wiki, last edited 21 March 2026. https://meta.wikimedia.org/wiki/Wikilegal/Copyright_Analysis_of_ChatGPT
73. Databricks. "databricks-dolly-15k" (dataset repository metadata). Hugging Face, accessed November 2026. https://huggingface.co/datasets/databricks/databricks-dolly-15k
74. Wikimedia. "wikipedia" (dataset repository metadata). Hugging Face, accessed November 2026. https://huggingface.co/datasets/wikimedia/wikipedia
75. Stability AI. "StableLM: Stability AI Language Models" (README). GitHub repository Stability-AI/StableLM, accessed November 2026. https://github.com/Stability-AI/StableLM
76. Carlos Muñoz Ferrandis. "OpenRAIL: Towards open and responsible AI licensing frameworks". Hugging Face blog, August 31, 2022. https://huggingface.co/blog/open_rail
77. Responsible AI Licenses (RAIL). "FAQ". licenses.ai, accessed November 2026. https://www.licenses.ai/faq-2
78. Krzysztof Siewicz. "RAIL: Nonfree and unethical". Free Software Foundation, April 22, 2026. https://www.fsf.org/blogs/licensing/rail-are-nonfree-and-unethical
79. NUS Institute of Data Science. "ModelGo Licenses". modelgo.li, accessed November 2026. https://www.modelgo.li/
80. ModelGo. "ModelGo Attribution-ShareAlike License Version 2.0". GitHub repository Xtra-Computing/ModelGo, December 2025. https://github.com/Xtra-Computing/ModelGo/blob/main/MGL/V2/MG-BY-SA/LICENSE
81. Jonathan Katzy, Răzvan-Mihai Popescu, Arie van Deursen, Maliheh Izadi. "An Exploratory Investigation into Code License Infringements in Large Language Model Training Datasets". arXiv:2403.15230 (accepted to FORGE 2024), March 22, 2024. https://arxiv.org/abs/2403.15230
82. Weiwei Xu, Kai Gao, Hao He, Minghui Zhou. "LiCoEval: Evaluating LLMs on License Compliance in Code Generation". arXiv:2408.02487 (ICSE 2025), August 5, 2024. https://arxiv.org/abs/2408.02487
83. Gaia Colombo, Leonardo Mariani, Daniela Micucci, Oliviero Riganelli. "On the Possibility of Breaking Copyleft Licenses When Reusing Code Generated by ChatGPT". arXiv:2502.05023 (accepted at ICPC 2025), February 7, 2025. https://arxiv.org/abs/2502.05023
84. Benjamin Laufer, Hamidah Oderinwale, Jon Kleinberg. "Anatomy of a Machine Learning Ecosystem: 2 Million Models on Hugging Face". arXiv:2508.06811, August 9, 2025. https://arxiv.org/abs/2508.06811
85. chardet. "7.0.0" (release notes). GitHub repository chardet/chardet, March 4, 2026. https://github.com/chardet/chardet/releases/tag/7.0.0
86. Mark Pilgrim. "No right to relicense this project" (issue 327, with reply by Dan Blanchard, March 5, 2026). GitHub repository chardet/chardet, March 4, 2026. https://github.com/chardet/chardet/issues/327
87. Masayuki Hatta. "Reproducibility is the New Copyleft: Defining AGI-oriented Reproducible Builds". arXiv:2606.03019 (accepted at AGI-26), June 2, 2026. https://arxiv.org/abs/2606.03019
88. Richard Fontana. Comment on "What does the MIT license cover in chardet?" (issue 334). GitHub repository chardet/chardet, March 20, 2026. https://github.com/chardet/chardet/issues/334#issuecomment-4098524555
89. Dan Blanchard. "Change 7.x license from MIT to 0BSD" (pull request 349). GitHub repository chardet/chardet, March 22, 2026. https://github.com/chardet/chardet/pull/349

