Home → Blog → Can I Copyright My AI Character? What the USCO Actually Says in 2026

Playbook · Copyright

Can I Copyright My AI Character? What the USCO Actually Says in 2026

The USCO has refused most AI character registrations. One creator broke through in January 2025. Here's the line between registrable and not, and the test most coverage misses.

Ashton Levy-Park
Ashton Levy-ParkFounder · CharacterClaim
·14 min read

On March 2, 2026, the Supreme Court declined to hear Thaler v. Perlmutter. That single line of docket text settled a question creators have been asking for three years: can a work generated entirely by AI be copyrighted? At the federal appellate level, the answer is now no. The D.C. Circuit's 2025 ruling stands, and because every challenge to a USCO registration refusal is heard in that circuit, the rule is effectively national.

If you made a character with Midjourney, Stable Diffusion, DALL·E, or any other generative AI tool and you want to know whether you can register it with the U.S. Copyright Office, the question you're really asking has two parts, not one. Most coverage treats AI character copyright as a single question about authorship. It isn't. It's two filters stacked on top of each other, and a creator can clear the first and still fail the second.

This piece walks through both filters using the actual public record: the USCO's own decisions, the relevant court rulings, and the one creator who has so far broken through. I'm not a lawyer and this isn't legal advice. For decisions about registration or enforcement, talk to one. But the public record is clear enough to give you a useful map of where the line is right now.

The two filters every AI character has to clear

For an AI-generated character to be independently copyrightable in the United States, it has to clear two distinct legal tests. They are not the same test phrased two ways. They are different doctrines, decided by different bodies, with different evidentiary requirements.

Filter 1: Human authorship. Does the work itself qualify as the product of a human author under Thaler v. Perlmutterand the USCO's January 2025 Copyrightability Report? This is the question every AI copyright article focuses on. It applies to any AI-generated output: images, text, music, code, characters.

Filter 2: Character delineation. Even if a work qualifies for copyright, does the character within it qualify for independent copyright protection separate from the work? This comes from Nichols v. Universal Pictures (2d Cir. 1930) and DC Comics v. Towle (9th Cir. 2015). It applies to any character, AI-generated or not.

Most “can I copyright my AI character” coverage stops at filter 1. That's the mistake. A character that gets through filter 1 because the surrounding work has sufficient human authorship can still fail filter 2 because the character itself is not sufficiently distinctive. And so far in the public record, no AI-generated character has cleared both.

Filter 1: What Thaler actually settled

Stephen Thaler is a computer scientist who created an AI system called the Creativity Machine. In 2018 the system produced an image titled A Recent Entrance to Paradise. Thaler applied to register the copyright in 2018 listing the Creativity Machine as the author and himself as the owner under work-for-hire doctrine. The USCO refused. Thaler sued.

The path the case took matters because it explains why the rule applies to everyone now:

  • February 2022: USCO Review Board affirmed the refusal.
  • August 2023: U.S. District Court for D.C. affirmed the USCO. Judge Beryl Howell wrote that “human authorship is a bedrock requirement of copyright.”
  • March 18, 2025: D.C. Circuit unanimously affirmed. Judge Patricia Millett wrote that the Copyright Act “requires all eligible work to be authored in the first instance by a human being.”
  • May 2025: Full D.C. Circuit denied en banc rehearing.
  • March 2, 2026: Supreme Court denied certiorari.

Why this matters beyond Thaler's specific case: under 17 U.S.C. § 411, every challenge to a USCO registration refusal is heard in the U.S. District Court for the District of Columbia, with appeals going to the D.C. Circuit. The D.C. Circuit's ruling on human authorship is therefore binding on every future USCO refusal challenge, regardless of where the applicant lives or where the work was created. The rule is effectively national even without Supreme Court review.

The USCO's own position, laid out in Part 2 of its Copyright and Artificial Intelligence Report (January 29, 2025), goes further than the bare “AI can't be the author” rule. The Report concludes that prompts alone, no matter how detailed, do not make the user an author of AI output. The Office's reasoning: “prompts essentially function as instructions that convey unprotectable ideas,” and current AI systems do not give the user enough control over how those ideas get expressed. The Office analogized prompting to commissioning a work from another artist by describing what you want. The commissioner is not the author.

This is the position that's now the live battleground. A pending case, Allen v. Perlmutter in the District of Colorado, is testing exactly this question. Jason Allen used Midjourney to create Théâtre D'opéra Spatial, won a state fair art prize with it, and applied to register it with the USCO. The Office refused. Allen sued. According to the government's January 2026 cross-motion for summary judgment, Allen entered text prompts at least 624 times, used Midjourney's variation and upscaling tools, and used Photoshop to clean up cosmetic details. Even with that level of iteration, the USCO maintains that the work's expressive elements were generated by Midjourney, not authored by Allen.

Summary judgment briefing was completed in early 2026. As of this writing, no ruling has issued. Until it does, the USCO's position on iterative prompting holds.

Now stack the second filter on top.

Even when a work clears filter 1, a character within that work isn't automatically copyrightable as a character. Characters are protected as derivative works under 17 U.S.C. § 106, but only if they meet a separate test that has nothing to do with AI. The leading frameworks:

Nichols v. Universal Pictures (2d Cir. 1930).Judge Learned Hand's “distinctly delineated” test. “The less developed the characters, the less they can be copyrighted; that is the penalty an author must bear for making them too indistinctly.” A character must be specifically described and fully developed to qualify. Stock characters and prototypes do not qualify.

DC Comics v. Towle (9th Cir. 2015). The Towle test, applied to the Batmobile. A character is copyrightable if (1) it has physical as well as conceptual qualities, (2) it is sufficiently delineated to be recognizable as the same character whenever it appears (consistent traits, look, personality, context), and (3) it is especially distinctive and contains some unique elements of expression.

Warner Bros. v. CBS (9th Cir. 1954).The “story being told” test. A character is copyrightable if it constitutes the story being told, rather than being a “mere chessman in the game of telling the story.” This test is now applied less often than the delineation test, but courts in the Ninth Circuit sometimes apply both.

For a traditional character (a novelist's protagonist, a comic book hero, an animated mascot), these tests are passable with effort. The author writes the character into multiple appearances, gives them consistent traits, develops their personality, and the delineation accumulates over time.

For an AI-generated character, here's the problem. Generative AI tools do not produce identical outputs from identical prompts. The same prompt run twice gives you two different versions of “the character.” Across multiple appearances, the character's appearance can drift in ways the human user did not control. Under Towle's second prong (consistent and identifiable traits across appearances), this drift erodes the case for delineation. Under Nichols, the lack of consistent expression suggests the character is more “type” than “individual.”

This is on top of the human authorship problem from filter 1. Even if a creator successfully argues that their selection, arrangement, and editing of AI outputs makes them the author of a specific work, the character within that work has to clear delineation on top.

The American Cheese registration: what got through, and what didn't

On January 30, 2025, the USCO registered A Single Piece of American Cheese, a digital image created by Kent Keirsey, founder and CEO of Invoke AI. It was the first registration of a single image created entirely with the assistance of AI. The application had been initially refused in September 2024 and granted on reconsideration after Keirsey submitted additional evidence.

What the registration actually covers, per the USCO's own letter and Keirsey's whitepaper:

  • The work was registered on the basis of “the selection, coordination, and arrangement of material generated by artificial intelligence.”
  • Keirsey's process used Invoke AI's “inpainting” feature: he started with a base image generated from a text prompt, then used selection tools to mark specific regions of the image and regenerate those regions with new prompts, iteratively.
  • Over 35 rounds of edits were documented in a time-lapse video submitted to the Office.
  • Counsel from Cooley LLP handled the application and reconsideration.
  • The registration was credited to Invoke AI as the employer under work-for-hire doctrine, with Keirsey as the human author.

What the registration does NOT cover, and this is the part most coverage glossed over:

  • The individual AI-generated image fragments themselves are not protected. The protection runs to Keirsey's selection, coordination, and arrangement of those fragments.
  • The USCO's letter explicitly stated the registration was “based on the particular facts presented here.” It is not a blanket rule.
  • American Cheese is a single image, not a character. The work depicts a stylized portrait of a woman with three eyes, melted cheese on her hair, and stained-glass-style patterning. It has no character continuity, no narrative role, no second appearance. The delineation question never came up because there was no character to delineate.

So American Cheese is precedent that an AI-assisted single image can be registered if the human contribution is documented in detail and rises above mere prompting. It is not precedent that an AI-generated character can be registered.

Zarya of the Dawn: the closest the USCO has come to ruling on an AI character

Zarya of the Dawn is the case that comes closest. Kris Kashtanova published a graphic novel in 2022 with images created using Midjourney. The protagonist, Zarya, appears across multiple panels. Kashtanova applied for and received copyright registration in September 2022 without disclosing AI use. After the Office learned of the AI use from public statements, it canceled the original registration and issued a more limited one on February 21, 2023.

The USCO's reasoning is the cleanest character-adjacent ruling we have:

  • Kashtanova was credited as the author of “the Work's text as well as the selection, coordination, and arrangement of the Work's written and visual elements.”
  • The individual AI-generated images, including the images of the character Zarya, were excluded from the registration.
  • The Office's analysis of Kashtanova's edits to a specific image of Zarya (modifications to the lips and mouth) found that the edits were “too minor and imperceptible” to qualify the modified image for copyright.
  • The Office distinguished Midjourney from tools like Photoshop on the ground that Midjourney users cannot predict the specific output, while Photoshop users select what to modify with specific intent.

Translated: even though Kashtanova put hundreds of iterations into shaping how the character Zarya appeared across the comic, the USCO refused to extend protection to the AI-generated character imagery itself. Other creators are free to copy Zarya's AI-generated visual appearance, because that appearance has no copyright protection at the image level.

This is what makes the Zarya decision so important for AI character copyright specifically. The USCO addressed a character (not a single one-off image), it addressed a creator who put substantial effort into iterative prompting, and it still refused to protect the character's visual representation. The protection ran to the human-written story and the arrangement, not to Zarya as a character.

No public registration has yet extended copyright to an AI-generated character independently. That's the current ceiling.

So what about your character? The practical answer

If you have an AI-generated character and you're trying to figure out whether to register, here's the framework I'd walk through. None of this is legal advice. It's a structured way to think about which filter is your weak point so you know what to do next.

Step 1: Clear filter 1, or accept that you can't.

  • If the character was generated entirely by prompting (text-to-image, run-and-pick), you cannot currently register the AI-generated imagery. Period. Thaler foreclosed it. Allenmay or may not move that line, but it hasn't yet.
  • If the character was generated through a process with substantial documented human input beyond prompting (extensive inpainting like Keirsey's process, or human-drawn elements that the AI assisted with, or significant post-generation editing), you have a case for filter 1. Build the documentation now: time-lapse video, version history, written description of every creative decision. The American Cheese application is the template.
  • If the character is described primarily in text you wrote (a written description, character bible, backstory) and the AI imagery is illustrative of that human-authored description, the text is registrable. The imagery probably isn't, but the text plus arrangement is.

Step 2: Clear filter 2, or accept that you can't.

  • Has the character appeared in multiple works with consistent visual and personality traits? If the character looks meaningfully different every time it's generated, delineation is weak.
  • Is the character distinctive (a specific look, name, mannerism, role) or is it a type (“warrior princess,” “evil clown”)? Types fail Nichols.
  • Has the character been written into a story where it does specific things, says specific things, behaves consistently? The more developed the character, the stronger the delineation case.

Step 3: Match strategy to weak point.

  • If filter 1 is your weak point, focus on documenting the human contribution and structuring your creative process to maximize human creative control. Inpainting with detailed records. Hand-drawn elements. Extensive post-generation editing. The Cooley/Invoke playbook.
  • If filter 2 is your weak point, build delineation. Multiple consistent appearances. Written character bible. Story development. The character has to feel like a specific individual, not a recipe.
  • If both are weak, register what you can (text, arrangement) and treat the character imagery as unprotected. Plan accordingly.

Step 4: Consider jurisdiction.

This article is U.S.-only. China's courts have granted copyright to some AI-generated content under standards that look at iterative prompting more favorably (the BIC case, November 2023). The EU position is closer to the U.S. position, requiring an “author's own intellectual creation” that has so far been read to require a human. The UK has a separate provision for “computer-generated works” with a 50-year term and no human author requirement, though its application to current generative AI is contested. If your character has international commercial value, multi-jurisdictional registration strategy is its own project.

Trademark is a different path

One thing this article has not covered: trademark. Trademark protects marks that identify the source of goods or services in commerce. It has no human authorship requirement. The Lanham Act doesn't care whether your character was drawn by hand or generated by Midjourney. It cares whether the mark is distinctive and used in commerce.

For some AI characters, particularly ones being commercialized as merchandise or content franchises, trademark may be a more accessible protection than copyright right now. The USPTO has registered trademarks on AI-generated characters: registration #99171194 for Tung Tung Tung Sahur in Class 28 (toys) is one example we've covered in a separate piece. A future article will walk through the trademark pathway for AI characters specifically. For this piece, the takeaway is: copyright and trademark answer different questions, and a character that fails copyright registration may still be protectable as a trademark if it functions as a source identifier in commerce.

Where this is going

The current rule, as of May 2026: AI characters generated through prompting alone are not registrable as copyright in the United States. Characters created with substantial documented human creative control beyond prompting may be registrable as compilations or arrangements, but no AI-generated character has yet been independently registered. The character delineation requirement adds a second hurdle on top of the human authorship hurdle.

Allen v. Perlmutter could change part of this picture. If the District of Colorado rules that 624 prompts plus Photoshop edits do amount to human authorship, the line on filter 1 moves. Even then, filter 2 remains independent. Character delineation doctrine is over a century old and has no special AI exception.

Until the courts move, the practical answer for most creators searching “can I copyright my AI character” is: probably not the AI imagery itself, but possibly the surrounding written work, and possibly the trademark side of the same character. Documentation matters. Process matters. The creators who are clearing the bar right now are doing it with extensive evidence of human creative control, not with detailed prompts.

Sources

I am not a lawyer and this article is not legal advice. CharacterClaim.ai is an automated assessment tool, not a law firm. For decisions involving copyright registration, enforcement, or litigation, consult a licensed attorney. The article references active litigation (Allen v. Perlmutter) and should not be read as an opinion on the merits of pending litigation. Statements about the parties reflect public filings and reported statements as of the publication date. The methodology behind CharacterClaim's assessment is described at /how-it-works.