Development of openness of AI models and rights to AI generated content

The OECD experience

In August 2025, the OECD presented a report on “AI openness”, describing how countries can implement frameworks for freely distributed components of AI models. The OECD recommends that regulators define the concept of “open-source AI” or “AI source¹ code” in AI regulation by introducing levels of² openness for AI systems into legislation. The OECD notes that the term “open-source AI” does not accurately describe which components³ of AI are open: model weights, code, or data. For example, disclosing only the weights can be labeled “open source,” but in fact provides little practical value without the underlying source code. Disclosing these components is a key tool for transparency enabling third parties to verify a model's quality and risks, identify errors and biases, and better explain its outputs. The OECD suggests describing levels of openness based on the accessibility of components: “open model” (weights and basic description), “open tools” (training and evaluation codes, key datasets are added), “open science” (the entire development cycle and materials are disclosed).

It is important to introduce “open licenses” into regulation, allowing the free use of intellectual property (including AI systems). Licenses vary in purpose. For example, permissive licenses allow the use of open AI elements when the developer is specified. These accelerate development and implementation by enabling free experimentation and the sale of solutions, provided the license terms are met. However, due to the risk of abuse, they need to be supplemented with checks for malware distribution, copyright infringement, or illegal data use. Another type is copyleft licenses, which require that any product using opensource AI elements must be distributed under the same conditions.

In introducing such licenses, the OECD emphasizes that regulators must ensure the legality and security of published AI components and data. Requirements for descriptions are recommended: what elements are published and what data sources are used. Components must be tested for security and compliance with their declared properties.

The US experience

In August 2025, a law on AI content⁴ rights came into force in the state of Arkansas. By default, if a person gives instructions to a generative AI system, supplies data, and receives content (text, images, code, etc.), they hold the rights to the output. If they provide data for training, they become the owner of the version trained on that data, provided the data was obtained legally and the rights must not be transferred to the developer/provider by contract. In the case of employees, if working with AI is part of their job responsibilities, the rights to the content belong to the employer.

The law specifies that it is not possible to appropriate anything that infringes on the copyright or other rights of others. If someone else's copyrighted material or personal data is used when requesting or transferring data, the person does not acquire intellectual property rights to the generated content. The question of who owns a model trained on data from many individuals remains open: the law does not resolve this issue.

The law clarifies rights concerning generated content, the user's prompt, the provided data, and the model trained on that data. If a company legally provides data for training, it obtains ownership rights to the trained model. The law reduces the risk of conflicts in joint projects: the parties can establish a different ownership order in advance—the terms of the contract apply. For example, when retraining a supplier's model on customer data, rights can be divided: the “weights” and their updates remain with the supplier, and the customer receives a license for internal use.

The user owns the copyright to the content if the data used belongs to them on a legal basis. This resolves the issue of rights to AI content in the event of copyright infringement by the user. IP rights to such content do not arise, and fragments (e.g., text or code) similar to someone else's IP object will belong to the copyright holder in the event of a proven infringement.

The experience of Russia

Russia has introduced certain measures to stimulate the development of open-source AI. For example, expenses incurred in developing open-source platforms for “smart assistants” can be counted twice toward reducing corporate income tax. However, Russia has not adopted regulations that take into account the specifics of open AI licensing. When developing such regulations, it is worth paying attention to the OECD's recommendations on the requirements for descriptions of AI components distributed under open licenses, as well as on indicating data sources and ensuring that such components can be verified for security.

Russia also lacks specific regulations on intellectual property rights for AI-generated content and AI models trained on data belonging to third parties. Discussion of these issues in the⁵ State Duma is scheduled for fall 2025. During the discussion, attention should be paid to the following questions: whose data was used to train the AI model that generated the content; does the user who formed the request for the AI to generate content provide their own data?

From the monitoring issue No. 8 (20), August 2025. Download the full issue (PDF) · issue page at the Gaidar Institute