- Joined
- Jan 20, 2026
- Messages
- 345
- Reaction score
- 2,243
Closed models are profiting from the open code - and now they have come up with a legal trap for them.
Genetically AI has grown on someone else’s open source, but it rarely responds with the same openness. Developers of free and open software for years have published libraries, tools, databases, browsers, server components and infrastructure code that can be studied, modified and distributed. Now large AI companies use such materials in training models, but do not disclose what data got into training, how the model works and what rules determine its work.
The researchers proposed to apply the logic of "copyleft" to AI (this is such a play of words - the opposite of the term "copyright"). In the world of free software, the copyplay works as a special option for copyright: if someone creates a derivative work based on open source, the result should also remain free and open. This approach underlies licenses that prevent the adoption of a common development, close it within a proprietary product and deprive the community of access to improvements.
The authors работыoffer Contextual Copyleft AI License, or CCAI, a contextual copyleft license for AI. The idea here is built around one question: is a generative model a derivative work, if the developer taught it on an open source. If yes, then the model should be subject to responsibilities from the license of the source materials.
According to the authors, CCAI would force AI developers to disclose the model architecture and training if the training used open source under such a license. Then the company would not be able to take the community materials, train a closed system on them and leave developers without understanding how their work affected the final product. For the FOSS community, this is a way to regain control over how the code is used in the AI industry.
Free and open software has long ceased to be a niche for enthusiasts. It holds operational systems, web browsers, databases, scientific tools, development tools, cloud platforms, mobile devices and a significant part of the Internet infrastructure. AI developers also depend on this layer: open libraries, frameworks and repositories help to assemble models, train them, test code and deploy services.
The problem arises where openness breaks down. The company can use open materials, but then call the model open only because it publishes weights, displaying access, or part of the documentation. At the same time, training data, accurate architecture, filters, settings and training procedures remain closed. Outside the product looks friendly to the open source, but it is almost impossible to check its origin and limitations.
The authors call it open operation, that is, pseudo-openness. In such cases, the company receives reputational advantages from communication with the open community, but does not give users and developers those rights that are usually associated with free software: study the system, play the result, change components and spread improvements. CCAI should narrow this gap between the open marketing word and the actual testability of the model.
The legal part of the study is built around the doctrine of fair use - the fair use in American law. If AI training on copyrighted materials is recognized by fair use, it will be more difficult for licenses to impose additional responsibilities on model developers. If the courts do not recognize training with such free use, the copyelft approach receives more grounds: the owner of the code will be able to specify in advance the conditions on which his materials are allowed to be included in the training sets.
Researchers consider CCAI legally possible under current copyright, but only with this reservation. The license does not solve the fair use dispute on its own. Rather, it proposes a mechanism in case the law confirms that the use of open source code for training may create responsibilities to the authors of the source materials.
A separate argument is about safety. Geneative AI differs from conventional software in that it can directly create text, code, images and instructions. The open model is able to help researchers, speed up program checking and support development, but the same system can generate convincing phishing emails, harmful content or fraud materials. Therefore, full openness in AI raises more questions than the publication of a regular library.
The authors do not urge to ignore these risks. The study mentions regulatory measures such as EU rules, which prohibit AI systems from using hidden, manipulative or deceptive techniques that affect human behavior and decisions. According to the authors, the CCAI license should work on an equal footing with the following rules: the right limits dangerous methods of use, and the copyleft license is responsible for the transparency of the origin and device of the model.
Of course, the proposed scheme has several practical consequences. Open source software developers get more levers when their code gets into training sets. There is an incentive to create really open generative models, because closed companies will not be able to use code under CCAI without conditions. The community gains access to training data and technical details, not just a ready-made interface or advertising description.
The approach can change the relationship between AI companies and FOSS developers. Now the open code often turns into raw materials: it is taken for training, but the result does not return in the form of an understandable, verifiable and reproduced system. Contextual copyeft tries to transfer the old free software rule to a new field: if the model is trained on a common resource, then the model itself should remain open enough for the community.
The authors of the work - Grant Shanklin, Claudio Novelli, Emmy Hain, Luciano Floridy and Tyler Schroeder - do not claim that CCAI will automatically solve the conflict around generative AI. But their offer gives the FOSS community a legal language to argue with companies that use open source and then close the most important parts of the model. Further, the issue will be resolved not only in university articles, but also in licenses, courts, regulatory documents and practical rules for the publication of AI systems.
Genetically AI has grown on someone else’s open source, but it rarely responds with the same openness. Developers of free and open software for years have published libraries, tools, databases, browsers, server components and infrastructure code that can be studied, modified and distributed. Now large AI companies use such materials in training models, but do not disclose what data got into training, how the model works and what rules determine its work.
The researchers proposed to apply the logic of "copyleft" to AI (this is such a play of words - the opposite of the term "copyright"). In the world of free software, the copyplay works as a special option for copyright: if someone creates a derivative work based on open source, the result should also remain free and open. This approach underlies licenses that prevent the adoption of a common development, close it within a proprietary product and deprive the community of access to improvements.
The authors работыoffer Contextual Copyleft AI License, or CCAI, a contextual copyleft license for AI. The idea here is built around one question: is a generative model a derivative work, if the developer taught it on an open source. If yes, then the model should be subject to responsibilities from the license of the source materials.
According to the authors, CCAI would force AI developers to disclose the model architecture and training if the training used open source under such a license. Then the company would not be able to take the community materials, train a closed system on them and leave developers without understanding how their work affected the final product. For the FOSS community, this is a way to regain control over how the code is used in the AI industry.
Free and open software has long ceased to be a niche for enthusiasts. It holds operational systems, web browsers, databases, scientific tools, development tools, cloud platforms, mobile devices and a significant part of the Internet infrastructure. AI developers also depend on this layer: open libraries, frameworks and repositories help to assemble models, train them, test code and deploy services.
The problem arises where openness breaks down. The company can use open materials, but then call the model open only because it publishes weights, displaying access, or part of the documentation. At the same time, training data, accurate architecture, filters, settings and training procedures remain closed. Outside the product looks friendly to the open source, but it is almost impossible to check its origin and limitations.
The authors call it open operation, that is, pseudo-openness. In such cases, the company receives reputational advantages from communication with the open community, but does not give users and developers those rights that are usually associated with free software: study the system, play the result, change components and spread improvements. CCAI should narrow this gap between the open marketing word and the actual testability of the model.
The legal part of the study is built around the doctrine of fair use - the fair use in American law. If AI training on copyrighted materials is recognized by fair use, it will be more difficult for licenses to impose additional responsibilities on model developers. If the courts do not recognize training with such free use, the copyelft approach receives more grounds: the owner of the code will be able to specify in advance the conditions on which his materials are allowed to be included in the training sets.
Researchers consider CCAI legally possible under current copyright, but only with this reservation. The license does not solve the fair use dispute on its own. Rather, it proposes a mechanism in case the law confirms that the use of open source code for training may create responsibilities to the authors of the source materials.
A separate argument is about safety. Geneative AI differs from conventional software in that it can directly create text, code, images and instructions. The open model is able to help researchers, speed up program checking and support development, but the same system can generate convincing phishing emails, harmful content or fraud materials. Therefore, full openness in AI raises more questions than the publication of a regular library.
The authors do not urge to ignore these risks. The study mentions regulatory measures such as EU rules, which prohibit AI systems from using hidden, manipulative or deceptive techniques that affect human behavior and decisions. According to the authors, the CCAI license should work on an equal footing with the following rules: the right limits dangerous methods of use, and the copyleft license is responsible for the transparency of the origin and device of the model.
Of course, the proposed scheme has several practical consequences. Open source software developers get more levers when their code gets into training sets. There is an incentive to create really open generative models, because closed companies will not be able to use code under CCAI without conditions. The community gains access to training data and technical details, not just a ready-made interface or advertising description.
The approach can change the relationship between AI companies and FOSS developers. Now the open code often turns into raw materials: it is taken for training, but the result does not return in the form of an understandable, verifiable and reproduced system. Contextual copyeft tries to transfer the old free software rule to a new field: if the model is trained on a common resource, then the model itself should remain open enough for the community.
The authors of the work - Grant Shanklin, Claudio Novelli, Emmy Hain, Luciano Floridy and Tyler Schroeder - do not claim that CCAI will automatically solve the conflict around generative AI. But their offer gives the FOSS community a legal language to argue with companies that use open source and then close the most important parts of the model. Further, the issue will be resolved not only in university articles, but also in licenses, courts, regulatory documents and practical rules for the publication of AI systems.
