Two different questions routinely get conflated when researchers ask whether using AI writing tools “affects copyright” on a paper: who can be credited as an author (a question journals answer through policies like ICMJE’s criteria or CRediT), and who legally owns the copyright in the text (a question governed by copyright law, not journal policy). This guide is about the second question. It covers the U.S. Copyright Office’s human-authorship requirement, how it applies to AI-assisted manuscripts, and what it means in practice when you sign a publisher’s copyright transfer agreement (CTA).
For the authorship-credit side of AI use — whether AI tools can be listed as authors, what publishers require you to disclose, and how CRediT statements handle AI involvement — see CASRAI’s guide to journal and publisher policies on generative AI in manuscripts and the AI-assisted writing dictionary entry. This page assumes you already know AI cannot be listed as a byline author under any major publisher or funder policy; the question here is narrower and purely legal: when AI helps write text, who — if anyone — owns the copyright in that text?
The U.S. Copyright Office’s human-authorship requirement
U.S. copyright law protects “original works of authorship,” and both the courts and the Copyright Office have long read “authorship” to require a human author. The Office’s Compendium of U.S. Copyright Office Practices has stated for years that copyright protects “the fruits of intellectual labor” that “are founded in the creative powers of the mind,” and that works produced without any human involvement — the often-cited example is a photograph a monkey took of itself, at issue in the 2018 Ninth Circuit case Naruto v. Slater — are not registrable. Generative AI raised the same question in a new form: if a person types a prompt and an AI model produces the resulting text or image, has a human actually “authored” the output?
Zarya of the Dawn (2023)
The Copyright Office addressed this directly in its February 2023 decision on Zarya of the Dawn, a graphic novel by Kristina Kashtanova that combined AI-generated images (produced with Midjourney) with human-written text and human-selected page layout. The Office cancelled the original registration and reissued it covering only the human-authored elements — the text, and the selection, coordination, and arrangement of the images on the page — while explicitly excluding the individual AI-generated images themselves from the copyright claim, on the grounds that Kashtanova had not exercised sufficient creative control over how Midjourney rendered them.
March 2023 registration guidance
The Office followed with formal registration guidance (“Copyright Registration Guidance: Works Containing Material Generated by Artificial Intelligence,” March 2023) applying the same logic across the board: applicants must disclose the inclusion of AI-generated content in a work they are registering, and the AI-generated portions must be excluded from, or disclaimed within, the copyright claim, while any human-authored expression — text a person wrote, or a human’s original selection, coordination, and arrangement of AI outputs — remains claimable.
Part 2 of the Office’s AI report (February 2025) and Thaler v. Perlmutter
In its February 2025 report, Copyright and Artificial Intelligence, Part 2: Copyrightability, the Office went further into the mechanics of prompting. Its conclusion: prompting an AI system, by itself, does not supply the kind of human creative control the law requires, because a generative model “fills in” the specifics of expression rather than mechanically executing a human’s own creative choices — even a highly detailed prompt still leaves the system’s exercise of these choices largely up to the AI. The Office reaffirmed, however, that copyright protects original human expression layered onto or built from AI output: a human’s own text, a human’s substantive edits to AI-generated material, and a human’s original selection, coordination, or arrangement of AI outputs can all be independently protectable, assessed case by case. This tracks the same result reached in the litigation over Thaler v. Perlmutter, in which the courts upheld the Office’s refusal to register an image its creator, Stephen Thaler, had submitted as generated entirely by an AI system with no human authorship claimed at all.
Put together, the throughline across Zarya of the Dawn, the March 2023 guidance, and the 2025 report is consistent: copyright protects human-authored expression, not AI output, and the more an AI system supplies the actual wording or content rather than executing a human’s specific expressive choices, the less likely that content is to be independently copyrightable.
How this maps onto AI-assisted academic writing
Applying that standard to a manuscript means the outcome depends heavily on what kind of AI assistance was used — the Office’s framework draws roughly the same line researchers already navigate for disclosure purposes, between assistive and generative use.
AI-assisted editing and polishing
When a researcher writes their own text and uses an AI tool to check grammar, suggest phrasing, tighten sentences, or translate a passage they drafted, the underlying expression still originates with the human author. This is functionally close to the kind of assistance long provided by spell-checkers, style guides, professional copyeditors, or grammar-checking software, and it does not put the human’s authorship of that text in question under the Office’s framework — the creative choices about what to say remain the author’s. Most publisher generative-AI policies treat this category as low-risk and, in many cases, don’t require disclosure at all (see the publisher-by-publisher breakdown in CASRAI’s AI manuscript policy guide), precisely because the human is still the source of the expression.
AI-generated substantive content
The picture changes when an AI tool is asked to draft original passages — a paragraph, a section, a discussion of results — from a prompt, and that output is used largely as generated, with only light human editing. Under the Copyright Office’s current position, the resulting text is generated output, not the researcher’s own authored expression, and by itself is not copyrightable material the researcher can claim — regardless of how the manuscript’s authorship credit is ultimately decided under ICMJE, COPE, or a specific journal’s policy, which is a separate question governed by editorial policy rather than copyright law. A researcher who substantially rewrites, restructures, or builds on that AI draft with their own original judgment and wording is on firmer ground: their contribution is what the Office’s 2025 report frames as protectable human authorship layered onto AI output, even where the AI-generated starting material is not itself protectable.
There is no bright-line percentage or word count in the Office’s guidance that separates “assistive” from “generative” use — the analysis is inherently case-by-case, turning on how much of the actual expression traces back to human creative choices versus the model filling in the specifics. Researchers should not treat this guide as a substitute for institutional or legal advice on a specific manuscript.
What it means for the copyright transfer agreement you sign
Most subscription and many hybrid journals still require authors to sign a copyright transfer agreement (CTA) upon acceptance, under which the author(s) warrant that they are the sole owner(s) of copyright in the work and have full authority to transfer it to the publisher (open-access journals typically use an author-retains-copyright license instead, which raises a related but distinct question of what, exactly, the author is licensing). That warranty is where the human-authorship question becomes practically relevant to a working researcher, rather than a purely academic one: if a meaningful portion of a submitted manuscript is AI-generated text that was never substantially reworked by a human, the author may not actually hold copyright in that portion to transfer or license in the first place, because — per the Copyright Office’s position — no one does. Nothing in a CTA can transfer a right that doesn’t exist.
In practice, this is a reason (alongside the integrity and disclosure rationale that already drives most journal AI policies) to treat “AI drafted this and I lightly reviewed it” passages as needing real substantive revision before submission, not just as an authorship-disclosure formality. It also means that when a journal’s generative-AI policy asks you to disclose which tools were used and how, that disclosure is doing double duty: it supports the editorial integrity review, and it creates a record of which parts of the manuscript are the copyrightable output the CTA’s warranty clause is actually about. Publisher CTA and license language on this point is still evolving and not yet standardized across the industry — check the specific agreement text your target journal uses rather than assuming it addresses AI-generated content explicitly.
Practical guidance for researchers using AI writing tools
- Keep a record of how AI tools were used — which sections, what kind of assistance (editing vs. drafting), and what tool/version — both for journal disclosure requirements and as your own evidence of which text reflects your original authorship.
- Substantively revise any AI-drafted passages before submission. Beyond the integrity case for doing so, unedited AI output is the material least likely to be protectable copyright you can rely on owning.
- Follow your target journal’s generative-AI policy on disclosure and permitted use — these vary by publisher; CASRAI’s guides to Taylor & Francis, SAGE, IEEE, MDPI, and AMA/JAMA Network policies cover the specifics for each.
- Don’t conflate “AI can’t be an author” with “AI content isn’t copyrightable.” The first is an editorial-policy consensus (ICMJE, COPE, and every major publisher agree AI tools cannot meet authorship criteria or bear accountability). The second is a distinct, separate legal conclusion from the Copyright Office, though both point toward the same practical advice: a human needs to substantively author the work.
- Read your CTA’s warranty language before signing rather than treating it as boilerplate, particularly for any manuscript with substantial AI-assisted sections.
Frequently asked questions
Can I copyright a paper that ChatGPT or another AI tool wrote for me?
Not the AI-generated portions themselves, under the U.S. Copyright Office’s current position — copyright protects human-authored expression, and content an AI system generated from a prompt, without substantial human creative modification, isn’t independently copyrightable. Your own original writing, edits, and the way you select and arrange material remain protectable.
Does using an AI grammar or editing tool affect my copyright?
Generally no. Using AI to check grammar, suggest phrasing, or polish text you wrote doesn’t change who authored the underlying expression — you did. This is treated differently from asking AI to generate new substantive content.
Does my copyright transfer agreement need to mention AI use?
Not necessarily as a matter of current standard practice — CTA language on this point isn’t yet standardized across publishers. But because a CTA’s warranty typically requires you to certify you own the copyright you’re transferring, substantial undisclosed AI-generated content in a manuscript can create a real gap between what the agreement assumes and what’s actually protectable. Check your specific journal’s current generative-AI and submission policies.
Can an AI tool hold copyright or be named a co-owner?
No. Under the Copyright Office’s position (and consistent with rulings like Thaler v. Perlmutter), only a human can be an author or copyright holder; an AI system cannot hold copyright regardless of how the output was produced or credited.
Is this the same rule that applies to AI and patent inventorship?
No — related question, different regime. The U.S. Patent and Trademark Office and courts have separately held that only a natural person can be a patent “inventor,” most notably in Thaler v. Vidal (Fed. Cir. 2022). See CASRAI’s guide on AI and patent inventorship for that separate line of cases.
