The US Copyright Office denies purely AI-made work protection

A registration rule, not a training-data ruling

The US Copyright Office's registration guidance, effective 16 March 2023, states that copyright "can protect only material that is the product of human creativity" and that an applicant must disclose AI-generated content in a work and describe the human author's own contribution. Where a prompt alone produces an image or text, the guidance holds that "the traditional elements of authorship are determined and executed by the technology - not the human user", and that material must be excluded from a registration claim. The guidance addresses registration practice specifically; it does not rule on whether training a model on copyrighted works infringes anyone's copyright, a separate question the Office said it would take up later.

What the January 2025 report added

Nearly two years later, the Office's Part 2 report on copyrightability, dated January 2025, reaches the same conclusion after reviewing more than 10,000 public comments: "copyright does not extend to purely AI-generated material, or material where there is insufficient human control over the expressive elements". The report is more specific than the 2023 guidance on prompts, concluding that "based on the functioning of current generally available technology, prompts do not alone provide sufficient control" to make a prompter the output's author, even when a prompt is long, detailed, or repeatedly revised. It also confirms that a human's own expressive input - a drawing fed into an image model, for instance - can carry copyright into the parts of an output that are recognisably derived from it.

A case-by-case test, not a bright line

Both documents describe copyrightability as a fact-specific inquiry rather than a rule that can be checked mechanically: whether a human "selects or arranges" AI output creatively, or modifies it enough to meet the ordinary originality standard, remains a judgment call the Office says it will keep making application by application. Neither document addresses the separate, contested question of whether training itself requires a licence - that dispute is being argued in litigation such as the New York Times' case against OpenAI, described elsewhere in this archive, not settled by this registration guidance.

• Does the asset in question contain any human-authored expression that is separable from, and perceptible in, the AI-generated portion?
• Has the AI-generated content been disclosed and excluded in any copyright registration filed for a work that includes it?
• If a court eventually rules on training-data use, would that ruling change anything about the registration guidance summarised here?

The two documents together draw a narrow but firm line under US law as it currently stands: a prompt is an idea, not an execution, and only the human-authored execution around it is protected. Everything else - what training on copyrighted material requires, and how other countries will draw their own lines - remains open, and the Office says explicitly that it is still watching both.