Copyrightability

Can AI-Generated Content Be Copyrighted?

By Nora Whitfield · August 8, 2026

Can AI-generated content be copyrighted? The short answer is: parts of it, sometimes, if a human did enough of the creative work. The long answer is what this guide is for, because "sometimes" is doing a lot of work in that sentence and the difference between a registrable work and a rejected application usually comes down to details nobody explains clearly.

Here's the one rule that decides almost everything else: US copyright law protects human authorship. Not effort, not cost, not how clever your prompt was. Authorship. The Copyright Office has said this outright, in guidance, in registration decisions, and in the position it's taken in court. If a human didn't make the creative choices in a piece of expression, that piece isn't eligible for copyright, no matter who typed the prompt that produced it.

That rule sounds simple until you apply it to a 400-page novel drafted with ChatGPT, a Midjourney illustration you iterated on for six hours, or a Suno track built from lyrics you wrote yourself. Real work doesn't split cleanly into "human" and "AI" buckets. It's tangled. This guide walks through how the Office actually draws that line across art, writing, music, and code, and what you can do to make sure your human contribution survives the cut.

The foundational case: Thaler v. Perlmutter

Start here, because it's the case that settled the baseline question. Stephen Thaler built an AI system called the Creativity Machine and tried to register a visual artwork it produced entirely on its own, listing the machine as author and himself as owner by way of the work-for-hire doctrine. The Copyright Office refused. Thaler sued. He lost, and lost again on appeal.

The courts didn't rule that AI-assisted work can never be copyrighted. They ruled something narrower and more important: a work with zero human authorship can't be copyrighted by anyone, because copyright law has always required a human author. No human, no copyright. Full stop.

That's the floor everything else sits on. The interesting question, the one that actually affects most people using AI tools, isn't "did a human touch this at all" but "how much of the actual expression did a human control." That's where things get genuinely case-by-case.

Zarya of the Dawn: the case that drew the real line

If Thaler tells you the floor, Zarya of the Dawn tells you where the ceiling starts to bend. Kristina Kashtanova wrote a graphic novel and used Midjourney to generate the illustrations, then registered the whole thing as her own work. The Copyright Office initially approved it, then reversed course on the images once it understood how they were made.

The Office's reasoning is worth sitting with, because it's the clearest statement of the actual test. Kashtanova's text, the selection of images, and the arrangement of the whole book were hers, and stayed protected. The individual illustrations were not, because she didn't control the specific pixels Midjourney chose to render. She wrote prompts and picked results from what the model gave her. That's curation, not authorship of the image itself, in the Office's view.

This is the pattern that repeats across every medium: prompting and selecting is not the same as authoring, but arranging, editing, and adding your own expression on top of AI output usually is.

What "enough" human authorship actually looks like

There's no percentage threshold. Nobody at the Copyright Office is running your file through a checker that spits out "62% human, denied." The test is qualitative: did a human exercise creative control over the specific expressive choices in the work, not just the general idea or the starting point.

Things that count as creative control, based on actual Office guidance and decisions:

Writing original text yourself, even if AI helped you brainstorm or edit
Substantially editing, revising, or rearranging AI output so the final expression reflects your choices
Selecting and arranging AI-generated elements into a larger composition (a photo book, an album, a story built from generated scenes)
Making specific creative decisions about color, structure, composition, or narrative that a model executed under your direction, where you can show the direction was specific rather than a general prompt

Things that generally don't count, on their own:

Writing a detailed prompt, even a very detailed one. The Office has said explicitly that prompts function more like instructions to a commissioned artist than like the artist's own hand. Detail alone doesn't convert an instruction into authorship.
Picking your favorite output from a batch of AI generations without modifying it
Running an AI tool with default settings and accepting the first result

The pattern underneath all of this: the Office is asking whether you controlled the expression or just the conditions the expression came from. A photographer controls where the camera points, the lighting, the framing, the moment of the shutter. That's authorship. Someone who orders a portrait and describes what they want isn't the photographer, even if the description was excellent.

Visual art and images

AI-generated images sit at the sharpest edge of this rule because the model does the most visually obvious work. If you type a prompt into Midjourney, DALL-E, or Stable Diffusion and take the output as-is, that image is very likely not copyrightable on its own, per the Zarya precedent. If you generate elements and then substantially edit them, compose a layout, paint over portions, or combine multiple generations into something you actively shaped, the resulting composite has a much stronger claim to human authorship, though the exact boundary of what got registered is what you're on the hook to prove.

Written work: novels, articles, and scripts

Text is the medium where AI assistance is easiest to keep on the right side of the line, because writing naturally involves the kind of iterative human control the Office looks for: drafting, revising, restructuring, cutting, adding your own voice. A writer who uses AI to generate a rough first pass and then substantially rewrites it in their own words has a much stronger authorship claim than one who publishes AI output unedited. The safest posture, and the one that holds up if anyone ever challenges your registration, is to be able to show your own authored text and your own editorial decisions, distinct from whatever the model handed you first.

Music

This one's newer territory and still shaking out through litigation (the Suno and Udio lawsuits are worth reading if you're a musician using AI tools), but the underlying principle is the same as everywhere else. Lyrics you wrote are yours. A melody or arrangement you composed and merely used AI to help produce, orchestrate, or mix is yours. A track generated wholesale from a text prompt, where the model chose the melody, the chord progression, and the vocal performance, is much harder to register as your own composition. The human parts of a song, the words, the structure you dictated, the arrangement choices, are what the Office will look at.

Code

Software gets its own answer, because code's copyrightability was never about "creativity" in the artistic sense to begin with, it protects the specific expression of a solution, not the underlying idea or function. AI-generated code that a model wrote wholesale from a prompt sits in the same bucket as AI-generated prose: not independently copyrightable as written by you. But the architecture you designed, the specific logic you wrote or substantially edited, and the way you assembled a system from parts, some human-written, some AI-assisted, some copied from libraries, is where your authorship claim lives. A repo built with heavy AI assistance can still be a copyrightable work if you can show which parts carry your creative choices and which parts are AI output you selected but didn't materially transform.

The public domain trap

Here's the part people miss until it costs them something. A purely AI-generated work isn't just "harder to copyright." Under current guidance, it may not be protected at all, which means it can be used, copied, and resold by anyone, including your competitors, without recourse. There's no gray middle ground where an unregistrable AI work is still quietly yours by default. If the human authorship isn't there, the work can effectively be public domain, free for the taking, the moment it exists.

That's the commercial stake underneath all of this. It's not a bureaucratic hoop. It's the difference between owning an asset and giving one away.

What this means for your workflow

None of this means avoid AI tools. It means know which parts of your process are yours, and be able to show it. The Copyright Office increasingly asks applicants to disclose AI involvement and describe what was human-authored versus what was AI-generated (the "material excluded" field on an application exists specifically for this). Applicants who can point to a clear, contemporaneous account of their own creative decisions have an easier registration process and a stronger position if that registration is ever challenged.

That's the whole idea behind keeping a running record of your actual process, not reconstructing it after the fact from memory once a dispute or an application deadline forces the question.

What happens if you register without disclosing AI use

You might be tempted to just not mention the AI tools and register the whole thing as fully human-authored. Don't. The Copyright Office has been explicit that applicants have a duty to disclose AI-generated content, and a registration obtained by omitting that disclosure can be canceled later if it comes to light, which tends to happen at exactly the wrong moment: during litigation, during a licensing deal, during due diligence on an acquisition. A registration that falls apart under scrutiny is worse than no registration, because you built a deal or a lawsuit on top of it.

The honest path costs you a little precision up front (you have to actually think about which parts were AI-generated and describe them) and buys you a registration that holds up. That trade is almost always worth it.

Common questions people actually ask

Does using ChatGPT or Claude automatically void my copyright? No. Using an AI tool doesn't strip copyright from everything you touch that day. It affects which specific parts of the output are protectable. If you wrote the actual prose yourself and used the model for brainstorming, editing suggestions, or research, your writing is still yours. If the model generated finished paragraphs you kept largely as-is, those specific paragraphs are the part in question, not your whole manuscript.

What if I can't remember exactly how much I edited something six months later? This is the real problem behind most of this guide. Human memory reconstructs process badly, especially under the pressure of filling out a government form or responding to a legal challenge. A record made at the time you did the work, showing your prompts, your edits, your revisions, and the order you made them in, is worth far more as evidence than your best recollection months later. Courts and examiners weight contemporaneous records heavily over after-the-fact reconstruction, for the same reason a receipt beats a guess about what you paid.

Is there a percentage of AI content that's automatically safe? No, and be suspicious of anyone who tells you there is. The Office has never published a threshold like "under 20% AI content is fine." The test is about creative control over specific expression, not a ratio. A short story that's 90% your own prose with one AI-generated paragraph you kept unedited has that one paragraph as excludable material. A story where you wrote every word but used AI heavily for editing suggestions you then implemented yourself is fully yours, word count aside.

Do I need a lawyer to register AI-assisted work? Not for a straightforward case. The application process is designed for individual applicants, and the disclosure requirements, while they add a step, aren't legally complex to satisfy honestly. Where it gets harder is disputes: if someone challenges your registration or you're trying to enforce it against an infringer, that's when legal help earns its cost.

The bottom line

Copyright has never protected ideas, effort, or the fact that you spent hours iterating. It protects the specific expression a human author actually shaped. AI tools change how that expression gets produced, but they haven't changed what the law is looking for. The work that survives scrutiny, gets registered cleanly, and holds up if it's ever challenged is the work where a human can point to the actual creative choices and show, with real evidence, that they made them.

Start keeping that evidence now, while you're doing the work, not later when you need it.