Skip to content
Adrythm
Book 20 minutes
AI and automation

Human authorship requirement

can AI-generated work be copyrighted / copyright for AI images / disclosing AI in a copyright application

In short

The human authorship requirement is the US Copyright Office position that copyright protects only material made by a person, leaving AI output from prompts alone unprotected. It decides which parts of AI-assisted logos, images and website copy the business can register, and what the application must disclose.

The US Copyright Office says it is well established that copyright can protect only material that is the product of human creativity. Its registration guidance, published in the Federal Register on March 16, 2023, applies that position to work made with generative AI.

The test is who determined the expressive elements. Suppose an AI system receives only a prompt and produces a complex written, visual or musical work. The Office says the technology then determined the traditional elements of authorship. It compares prompts to instructions given to a commissioned artist. A prompt may itself be creative enough for protection, and that still does not make the generated material copyrightable. Part 2 of the Office's AI report, published January 29, 2025, reached the same conclusion about prompts alone.

Including AI material in a larger human work does not bar copyright. A person may select or arrange AI material in a sufficiently creative way, or modify it enough to meet the standard. Copyright then covers only the human-authored aspects. In a graphic novel with human-written text and Midjourney images, the Office found the work copyrightable and the individual images unprotected. Editing in Photoshop leaves the artist as author, though its AI-powered features get the same analysis.

Applicants have a duty to disclose AI-generated content and briefly explain the human contribution. They must use the Standard Application, describe the human authorship in the Author Created field, and exclude AI-generated content that is more than de minimis. The AI tool or its maker should not be listed as an author. An unsure applicant can state generally that the work contains AI-generated material.

A filing that left the AI material out can be fixed. A pending one goes through the Public Information Office, and an issued one needs a supplementary registration. Otherwise the Office may take steps to cancel the registration. A court may disregard it if the applicant knowingly gave inaccurate information and the accurate facts would have led to refusal.

The Office refused a 2018 application for a visual work described as autonomously created by a computer algorithm, a decision challenged in Thaler v. Perlmutter. As of September 2026, the Office's AI page lists a D.C. Circuit decision affirming that refusal. In January 2025 the Office said it plans to supplement the 2023 guidance, so the steps may change.

In practice

A bakery writes the text of a recipe booklet in house and illustrates it with images generated from prompts. Under the Office's guidance, the text can be claimed and the illustrations cannot. The bakery would file the Standard Application, describing the text under Author Created and excluding the illustrations as generated by artificial intelligence. If the booklet was already registered without that disclosure, a supplementary registration corrects the record.

Why it matters to you

Material generated from prompts alone gives the business nothing to register. The human part can still be protected, including creative selection, arrangement or modification of AI output. An application that leaves AI material undisclosed risks losing the benefits of the registration.

What to ask or check

  1. 01Which parts of your logo, images and website copy were generated by AI, and which were made or substantially changed by a person?
  2. 02Can the designer or agency describe in writing the human contribution to each piece made for the business?
  3. 03Did your copyright applications for AI-assisted work disclose the AI material and exclude it from the claim?
  4. 04If an existing registration left AI material out, has a supplementary registration been filed to correct it?

What people get wrong

That a detailed prompt makes the AI output the prompter's copyrighted work. The Copyright Office compares prompts to instructions given to a commissioned artist, and its 2025 report found the mere provision of prompts insufficient.

Red flags

  • A copyright application for AI-assisted work that names the AI tool or its company as an author.
  • An existing registration for AI-assisted work that never disclosed the AI-generated material.

Keep reading

Work made for hire

Work made for hire is the legal category that decides who owns something you paid to have made. Copyright starts with whoever created the work. For commissioned work it only becomes yours through one of two narrow routes in the statute, and software fits neither by default.

Statutory damages

Statutory damages are a copyright award a court sets within ranges fixed by law, without the owner proving any loss: $750 to $30,000 per work, and up to $150,000 if the infringement was willful. In federal court they generally apply only to works registered in time, so registration dates shape what one unlicensed website photo can cost.

Trademark

A trademark is a word, phrase, symbol or design that identifies your goods or services and indicates where they come from. It is not ownership of the word itself. The USPTO is blunt about that: rights attach to how the word is used with your specific goods or services, not to the word in general.

AI content disclosure

AI content disclosure is labeling material as machine-made where a platform requires it. Google's product listings require AI-generated images to carry a specific IPTC metadata code, and IPTC defines three different codes: created with generative AI, edited with it, and edited by a person.

Training data

Training data is what a model learned from. The question owners ask is whether their own data joins it, and there is no single answer: OWASP names three separate stages, pre-training, fine-tuning and embedding. A commitment worth having names the stage and comes in writing.

Retrieval augmented generation

Retrieval augmented generation is how an AI answers from your documents without being trained on them. OWASP describes it as combining a pre-trained model with external knowledge sources at answer time. So the documents sit in a store the system reads from, and who can read that store is the question.

Want this explained against your own numbers?

Twenty minutes, a straight answer, and no follow-up sequence if you decide not to work with us.

Book 20 minutes