In a long-anticipated report from the U.S. Copyright Office providing guidance on the copyrightability of works created by and/or with the assistance of Artificial Intelligence, the Office reaffirmed its previous guidance that "[w]here AI merely assists an author in the creative process, its use does not change the copyrightability of the output. At the other extreme, if content is entirely generated by AI, it cannot be protected by copyright. Between these boundaries, various forms and combinations of human contributions can be involved in producing AI outputs."

The problem, of course, is determining what will be deemed sufficient to warrant copyrightability in the vast gray area “between these boundaries.” Numerous commenters who provided input to the Office in connection with the preparation of the report argued that a prompt that is sufficiently detailed to itself be protected by copyright should result in copyright protection for the AI output generated from the prompt. But others, focusing on the “black box” nature of the AI system, the fact that a single prompt can result in limitless outputs, and that the prompts are more akin to ideas rather than true expression, argued that AI-generated creations are not sufficiently human-authored to warrant copyright protection regardless of the nature of the prompt.

Focusing on the level of control (or lack thereof) over the AI system, the Copyright Office came down on the side of the latter, noting that “[w]hile highly detailed prompts could contain the user’s desired expressive elements, at present they do not control how the AI system processes them in generating the output.” While future AI technology could give the human operator more control over the output, the AI systems presently available simply, in the Copyright Office's opinion, take too much out of the operator's hands.

This is not to say that all works that have any AI-generated elements will not be protectable. The Office noted that the use of AI systems as a tool to modify existing copyrightable expression likely results in copyrightable output. And human reworking and/or modification of AI-generated output may also result in a copyrightable work, even where such reworking employs AI to create revisions, so long as such revisions are selected and arranged by a human in producing the final work. The key in both cases is that the human input (i) amounts to copyrightable expression and (ii) is directly identifiable in the final work. The Office went on to note that in such situations, the copyright protection would be limited to the identifiable human-controlled features and aspects of the final work, and a registration would require a disclaimer of the AI-produced content.

This is not the last word we expect to hear from the Copyright Office. Having previously released a report discussing deepfakes and suggesting guidance for legislation on the subject, this report on the copyrightability of AI will be followed by a third report focusing on the legal issues inherent in the use of copyrighted works in the training of AI systems.