The United States Copyright Office issued its long-awaited comprehensive guidance on artificial intelligence and copyright on Thursday, providing the most detailed official statement yet on how the existing copyright framework applies to content created using AI systems and drawing a series of lines that will have significant practical consequences for artists, writers, musicians, technology companies and anyone who creates or uses AI-generated content professionally or commercially. The guidance, which builds on several earlier statements and court decisions but goes further than any previous official pronouncement in addressing the nuanced questions that AI generation raises for copyright doctrine, has been welcomed by some stakeholders and criticised by others – a division that reflects the genuine difficulty of applying a legal framework designed for human creativity to outputs generated by machine learning systems trained on vast amounts of human-created work.

The core principle established by the Copyright Office’s guidance is that copyright protection requires human authorship – a principle that has been settled law in the United States for decades and that the guidance confirms applies to AI-generated content. A photograph taken autonomously by a camera with no human compositional decisions is not copyrightable; a painting generated by an AI system from a text prompt, with no further human creative input beyond the prompt itself, is similarly not protected by copyright. The human whose creative choices are reflected in a work can claim copyright in those choices, but the portions of a work that are attributable to the AI system rather than to human creative decisions are not protected. This framework, while conceptually clear, raises immensely complex practical questions about how to determine which elements of a given work reflect human creativity and which reflect the AI system’s autonomous generation.

The Human Authorship Spectrum

The Copyright Office’s guidance introduces a spectrum framework for analysing AI-assisted works – works in which a human uses an AI tool as part of a creative process – that recognises the reality that most AI-generated content in commercial practice involves varying degrees of human creative input rather than pure autonomous AI generation. At one end of the spectrum is a simple text prompt that generates an image: if the human’s creative contribution is limited to the prompt (‘a cat sitting in a window’) and all compositional decisions are made by the AI system, no copyright attaches to the AI’s output. At the other end is a work in which a human uses AI tools for specific technical tasks – colour correction, text generation for placeholder content, background removal – while retaining full creative authorship of the work’s essential elements: this work is fully copyrightable as a human-authored work.

• Works created through extensive iterative prompting – where a human artist writes dozens or hundreds of prompts, selects from the AI’s outputs and assembles elements into a coherent final work – may qualify for copyright protection of the ‘selection and arrangement’ of AI-generated elements, a more limited form of protection than full copyright in the underlying content.
• AI-assisted works where a human provides significant creative direction at a granular level – specifying composition, colour palette, style references and making creative choices throughout the generation process – are more likely to qualify for copyright protection in the elements that reflect those human choices.
• Works that incorporate AI-generated elements alongside original human-authored content receive copyright protection for the human-authored portions, with the AI-generated portions in the public domain.
• The guidance explicitly declines to address whether the training of AI systems on copyrighted works without licence constitutes copyright infringement – a question that is the subject of multiple ongoing lawsuits and that the Copyright Office indicates will be addressed in a separate proceeding.
• Registration of AI-assisted works will require disclosure of the AI system used and a description of the human creative contributions, with the Copyright Office reserving the right to deny registration where the human creative contribution is insufficient to support protection.

Training Data and the Open Questions

The Copyright Office’s guidance deliberately leaves unresolved the most commercially significant question in AI and copyright: whether the use of copyrighted works to train AI systems constitutes copyright infringement. This question is the subject of major ongoing litigation, including lawsuits brought by the New York Times against OpenAI and Microsoft, by the Authors Guild against multiple AI companies, and by a coalition of visual artists against Stability AI and Midjourney. The Copyright Office has indicated that it will address the training data question in a separate report expected later this year, a decision that reflects both the complexity of the issue and the desire to avoid prejudging the outcomes of ongoing litigation in federal courts.

The training data question is legally distinct from the copyright ownership question addressed in Thursday’s guidance, but the two issues are intertwined in ways that make their separation somewhat artificial. If training on copyrighted works without licence is ultimately found to be infringement, the legal position of AI companies that have built large commercial businesses on models trained on unlicensed content will be dramatically altered, and the downstream copyright analysis of AI-generated outputs will be affected by questions about whether those outputs are infringing derivatives of the training data. If training is found to be protected as fair use, as the AI companies argue, the analysis of output copyright becomes cleaner but the compensation question for creators whose work contributed to the training remains unresolved.

International Dimensions

The United States Copyright Office’s guidance takes effect in a rapidly evolving international legal landscape where different jurisdictions are adopting different approaches to the same fundamental questions. The European Union’s AI Act, which began full enforcement earlier this year, includes provisions specifically addressing copyright and AI that differ in several respects from the approach outlined in Thursday’s US guidance. The UK Intellectual Property Office has consulted on a framework that would create a text and data mining exception that would permit AI training on copyrighted works without licence under certain conditions – a provision that, if enacted, would create a significant divergence between UK and US law on the training data question.

For companies operating internationally – which includes virtually every significant AI company given the global nature of the market for AI tools and services – the patchwork of emerging national frameworks creates compliance complexity that will require careful legal analysis of which jurisdiction’s law applies in any given context. The international dimensions of AI copyright are likely to be a subject of discussion at upcoming international intellectual property forums, and pressure for greater harmonisation is expected to grow as the commercial significance of the issue becomes clearer through the resolution of the major ongoing litigation. For now, the practical guidance for creators and companies using AI is to document carefully the human creative contributions to any work that incorporates AI-generated content, to consult with intellectual property counsel about the copyright status of specific AI-assisted works, and to watch developments in the ongoing litigation and regulatory proceedings that will ultimately provide the definitive answers that Thursday’s guidance, for all its detail, has left open.