It should not. But this is not a punitive position. It is a recognition of a structural reality.
Copyright arose from authors’ need to protect their works from being copied without their consent. Copyright relies on the clear identification of the specific work to be protected and of the specific author as the initial right holder. AI outputs can be neither a specific work, because they may or may not be recreated in the same AI tool, nor a specific author, because if an output can be recreated by someone else rather than by you at any point, how can you claim to be its author?
AI outputs are created through the combined contribution of the following elements, coming from independent authors: billions of training data items, millions of text and code instructions, a user prompt and the AI’s technologically unpredictable behaviour.
Therefore, the origin of an AI output is untraceable in practice. It is not practical to grant multi-decade legal protection to something that cannot be clearly separated from other works which, unlike AI outputs, have all the required evidence of origin.
Most importantly, AI developers themselves acknowledge this in the terms of service they publish.
For AI developers, the value of AI does not depend on whether AI outputs can be protected by copyright. AI technology has become so widespread and powerful that the absence of copyright protection has not stopped users from deploying it in their projects.
To illustrate the point, despite the lack of internationally recognised copyright protection for AI outputs, by 2025 AI-generated content had flooded the web: AI-written text had become the majority of newly published web articles[1], and generative AI had produced on the order of 15 billion images[2].
For the past three years, terms of service, not copyright, have been the main source of rights in AI outputs. Contract law applies to every model, every user and every use. Copyright as a topic of discussion arises only when an AI output violates someone else’s rights, not because the output itself can be protected by copyright.
Copyright law should not be asked to protect outputs produced by technologies built without regard for copyright, because doing so would demand more from the technology than it can deliver, namely a guarantee that the AI output is free from any copyright other than that of the specific user providing the prompt. This cannot be achieved in practice, as shown by the many copyright lawsuits and by the opaque approach AI developers take to transparency regarding training data.
The practical conclusion is simple: if you want to understand who gets what rights, be realistic and read the applicable terms of use.
