A group of researchers from Yale University has proposed (PDF) a new type of open license called CCAI (Contextual Copyleft), which broadens the application of copyleft to generative AI models. The idea behind CCAI is that using content licensed under this agreement as training data for machine learning disseminates the copyleft conditions to the resulting generative AI models. It is believed that the new license could help reduce abuses in AI projects and prevent the emergence of fake AI models that are formally presented as open but are tied to the manufacturer due to the concealment of original data and training tools.
CCAI stipulates that any distribution and publication of exact copies or modified derivative works distributed under the CCAI license must be done under the same licensing terms without imposing additional restrictions. This requirement applies to any AI model, dataset, or AI system that was trained using software licensed under CCAI or the results of its work. In the context of training generative AI models, CCAI requires the disclosure of the model's source code, a detailed description of the data used during training, as well as the parameters, weights, and architecture of the model.
The CCAI license can also be used as an additional requirement attached to existing copyleft licenses like AGPLv3. This requirement extends the license's scope to include training datasets, code, and model weights according to the openness criteria for AI systems formulated by the Open Source Initiative (OSI). Code distributed under this license can only be used to train an AI model if all users are provided with a description of the training dataset, the code for training the model, and the trained AI model.
Text of the attached additional requirement: "When using software for training, optimizing, or creating any machine learning model or generative AI system, any resulting model, dataset, or system must be published under terms that do not tighten the requirements of this license. Such a requirement includes providing access to the code for training, a description of the data used during training, model settings, and architecture. Providing access to the model or the output of its work over the network is considered distribution and forms the basis for fulfilling the obligations imposed by the license."
Source: opennet.ru
