When Your Ghostwriter Is a Ghost in the Machine: AI and Copyright Authorship

Ghostwriting traditionally involves a human collaborator who shares copyright with the credited author. But what happens when that “ghost” is an AI? U.S. copyright law grapples with the concept of authorship, and current guidance draws a firm line: only humans can be “authors” in the eyes of the law. This post explores the implications of collaborating with AI as a ghostwriter—examining U.S. Copyright Office (sometimes “USCO”) guidance, key developments like Zarya of the Dawn, and the unsettled gray areas where human creativity meets machine output.

Of great interest (maybe only to me) is that the “COMPENDIUM OF U.S. COPYRIGHT OFFICE PRACTICES” is in its third edition — released and effective December 22, 2014 and last modified January 28, 2021. That means more than four years have passed since it was updated. AI tech time runs even faster than internet company time. Unpacking that, it is an observation that the rate of change in the underlying technology for AI is very fast.

In January, 2021, OpenAI had just announced their DALL-E image generation technology. Generative text AI was starting to come into its own with GPT-3’s 175 billion parameter model. OpenAI hasn’t released numbers or their latest model, but today competitor Grok has 1.7 trillion parameters and estimates by third parties put OpenAI’s current model at a minimum of several trillion parameters. In the four years since the last US Copyright Office Compendium update, there has been incredible improvement to generative AI’s ability to kick out human-like works.

The US Copyright Office released a 3 part report analyzing copyright law and policy issues. A critical finding in the January 2025 “Part 2: Copyrightability” portion of the report found that “Based on the functioning of current generally available technology, prompts do not alone provide sufficient control” to make a work created by prompts copyrightable. We’ll explore that more below.

Human Authorship: The Primary Rule

The U.S. Copyright Office has made its stance clear: human creativity is required for copyright protection. The USCO position is that works entirely generated by AI (with no human input) are not copyrightable. As noted above, critically the Office found that merely providing prompts to an AI isn’t enough to create a protectable work — typing detailed instructions into a model does not by itself make the human a joint author. If a work contains more than a de minimis amount of AI-generated content, the applicant must disclose it and identify the human contributions. In essence, the law views AI outputs as akin to uncopyrightable material (like facts or mechanical processes) unless a human’s creativity meaningfully shapes the final work.

This principle was underscored in Thaler v. Perlmutter, where a researcher sought to register a copyright for an image generated by an AI system. In 2023, a federal court upheld the Copyright Office’s denial, ruling that a work created “absent any guiding human hand” cannot be copyrighted. The court affirmed that an AI, not being human, cannot be an author under U.S. law, echoing a longstanding rule that copyright protects only creations of human intellect. This outcome aligns with earlier precedents (for example, the “monkey selfie” case, Naruto v. Slater, where a photograph taken by a monkey was deemed unprotectable for lack of human authorship). The bottom line: regardless of how impressive or creative an AI’s output seems, without human creative input it lies outside copyright’s umbrella.

Gary’s Dog: Iterative Prompting

There is little question that a complex prompt itself can be copyrighted. The phrase “please create an image of a cute dog that is 1/4 collie and 3/4 german shepherd and is jumping off the diving board of a swimming pool, trying to catch a frisbee in her mouth; the pool is relatively narrow at the point where the diving board is located.; behind the diving board is a wooden fence with about 1 meter of ivy planting in front of it; it is a nice sunny day, and the color temperature should reflect that; in the background are three judges, a gorilla, a bigfoot, and a chimp holding up signs rating the dive (like at the olympics) with a score of 9.8, 9.9, and 8.8 respectively; there are a few scattered leaves floating in the pool; the frisbee is about 20 cm away from the dog’s mouth, and she is about to catch it” is copyrightable. The prompt has human creativity and is expressive enough to qualify. It follows that the image generated by that prompt would be a work of joint authorship, with portions generated by my prompt, and other portions filled in by the AI on its own. Let’s have a look at what Gemini 2.5 Pro does with that prompt:

Let’s consider the copyright implications. The diving board, the dog and the frisbee came from my prompt. However, the grass in the background did not. It follows, then, that I hold a copyright in parts of the image but not others. You can see how this can quickly become nightmarishly complex. The Copyright Office’s guidance that prompts do not provide enough control to make the output reflect the human creativity is pretty clearly wrong in practice (as the bizarre image above shows). Let’s try something else probative: We’ll ask AI to describe “The Scream” and see if it can create the drawing. The prompt is ** really long ** and reproduced at the end of this post.

The image above is obviously (to my eye) a derivative work of “The Scream”, yet it was created solely using a prompt that does not mention the name of the painting at all. If the U.S. Copyright Office is correct that no copyright exists in AI output regardless of the prompt, it raises the question of whether an adequately described work could recreate a copyrighted image without infringing. The answer should be an obvious “no”, but the guidance seems to hint that the answer is “yes”.

Now let’s start with a prompt that probably wouldn’t be copyrightable: “Make a dog”. Nice, image 1 is a golden retriever. Now let’s start iterating. I want the dog to be on the beach. “Please put the dog on the beach”. Kind of boring. Let’s have the dog wearing a birthday party hat: “Please have the dog wear a birthday party hat”. Nice, but not festive enough. “Please have other dogs at the party”. Let’s make it more interesting. “Add a volcano erupting in the distance.” Still not strange enough for me. “Please add a clown riding a unicycle in the background”. Perfect! My artwork is complete:

At some point between the first image — clearly not copyright — and the last, copyright (probably, but not according to USCO guidance that says no prompt is sufficient to guide an AI to reflect human creativity) arose. Rules addressing this situation in a realistic way have yet to be written in copyright offices and legislatures throughout the world, but it will almost necessarily involve a subjective assessment: At what point has the creative input via prompting resulted in a copyrightable work? § 102(a) of the U.S. Copyright Act separately lists “literary works” and “pictorial” works, but this is a hybrid. The human contribution was entirely textual. The artistic work was entirely machine-driven. It isn’t even clear what category this work falls under.

This seems like a new problem, but it has historical analogues. For example, the characters and universe of “Star Trek” are copyrighted. I can’t just write a script set in “The United Federation of Planets” and have Captain Picard lead a team of other known characters through adventures without infringing that copyright. It would be a derivative work, with copyright to the original portions of the Star Trek universe held by that copyright owner, and the original elements held by me. Is it the same in the case of my “Dog Birthday Party With Unicycle Clown, Party Hats, and Erupting Volcano” image? It should be, but it would be far better for legislatures to clarify the law than to have courts try to apply laws written for paintings to AI-assisted creations.

For educational purposes, let’s look at how the Star Trek example might play out as an image. I added “Now replace the dogs with Captain Picard, Worf, Commander Data and Commander LaForge, all celebrating Captain Kirk’s 100th birthday.”. It is pretty clear that Gemini has some guardrails, because it did not create the exact characters I asked for, but it still harkened back to Star Trek, and includes the copyrighted look and feel of the characters:

The actual rendering of the image was done with AI, so there is no copyright to the flourishes it created without prompting, like the footprints in the foreground. The characters are mine (the dog and the clown on the unicycle) and Star Trek’s. If we added in a version of Mickey Mouse that is still copyright, we’d have a triple derivative work. The complexity of the problem isn’t hard to see.

Zarya of the Dawn: A Live Example

A historical example of this mishegas in practice is the case of “Zarya of the Dawn.” In 2022, author Kristina Kashtanova used the AI tool Midjourney to generate artwork for a graphic novel and initially obtained a copyright registration. But the U.S. Copyright Office later revisited that registration, ultimately canceling it in part. The Office concluded that while Kashtanova could claim authorship over the text and the selection, coordination, and arrangement of elements in the novel, the AI-produced images did not meet the requirement of human authorship. The Office reissued a certificate covering only the portions of the work Kashtanova created without Midjourney’s help. In other words, the overall comic book could be partially protected (for its story and the human creative choices in how images and text were combined), but the individual AI-generated illustrations themselves were denied protection as original works.

“Zarya” highlights the nuanced outcomes when a human and AI “co-create.” The human’s contributions (writing the narrative, arranging panels, etc.) remain protected, but the AI’s contributions fall into the public domain. This raises practical questions: if an author uses AI for a draft or an image and then revises or incorporates it, how much revision is needed before the result is considered human-authored? The Copyright Office has indicated that minimal editing of AI output (like touch-ups or selection among options) may not suffice – the human must contribute their own creative expression. In Kashtanova’s case, their curation and story composition were protectable, but not the raw AI art. The situation creates a patchwork of rights: parts of a collaborative human–AI work might be copyrighted to the human author, while other parts (the AI-generated pieces) might be free for anyone to reuse.

Unsettled Areas and Future Outlook

The intersection of AI assistance and copyright law is full of unanswered questions. One grey area is threshold of creativity: how much human input into an AI’s output is needed before the human can be considered an “author”? Only where the human’s creative process truly shapes the expression (rather than the AI determining it autonomously) will the result likely qualify for copyright. For instance, using an image-generating AI and then painting over it or collaging multiple AI outputs might yield a human-authored work, whereas simply choosing the best AI-generated paragraph out of five options probably would not.

Another unsettled issue is joint authorship in collaborations involving AI. With two human co-authors, each owns a share of the copyright. But an AI cannot hold copyright, so if an AI “collaborates,” the human user might be deemed the sole author of whatever portions are sufficiently human-driven. Alternatively, the AI-produced portions might be viewed as uncopyrightable fragments inserted into a larger work. This complicates scenarios like an author using AI to ghostwrite portions of a novel or article. The human may claim full ownership only if they can show they creatively arranged or revised the AI content in a way that meets the originality standard. If not, the AI-written chapters could be unprotected text intermingled with the author’s protected chapters – a nightmare for determining rights. We’re essentially seeing a new species of derivative work questions: is an AI’s output a kind of raw material that an author can transform and claim, or an independent creation that no one can own?

Practical Guidance: Keep a copy of all prompts and AI outputs. This will allow a delta (or redline or blackline, basically a comparison between the AI output and the final product) to be generated, making the copyright office’s job of determining which parts of the output are human much easier. This could be the difference between winning and losing copyright infringement claims in court.

For now, the U.S. Copyright Office’s position is firm: no human authorship, no copyright. Even complex, creative prompts aren’t enough to create copyright in the output. Legislative or judicial developments could clarify matters down the road. Courts might eventually tackle scenarios like an AI-assisted screenplay or a machine-generated painting curated by a human, and refine the tests for authorship. Until then, creators using AI as a ghostwriter face a trade-off: they can leverage the technology’s prowess, but they must be prepared to do enough creative work themselves to secure copyright – or risk parts of their creation wandering unprotected in the public domain. In short, if your ghostwriter is a ghost in the machine, you may end up being the only legally-recognized author in the room (for better or worse).

Appendix A: Gemini Pro’s Detailed Instructions to Reproduce “The Scream”: UPDATE: I had to delete this because it was making Google think that I had AI write the article, and caused it to not be indexed on Google. I’m sorry. It isn’t my preference to be unable to give examples because Google misunderstands what they are, but it is out of my control. ☹️