Under current U.S. copyright law, artificial intelligence itself cannot be an author, and any work generated entirely by an algorithm falls into the public domain. A person can still claim copyright in work made with AI assistance, but only over the parts a human actually authored, and only if the AI’s involvement is disclosed on the application. Separately, the biggest open question in AI copyright law right now is not who owns the output but whether training a model on copyrighted material without permission is infringement. That fight is still in the courts.
The Human Authorship Rule
The Copyright Office will not register a work unless a human being created it. The Copyright Act uses the word “author,” and courts read that to mean a natural person, not a machine.1Office of the Law Revision Counsel. 17 U.S.C. 101 – Definitions
The controlling case is Thaler v. Perlmutter. Stephen Thaler tried to register a visual artwork called “A Recent Entrance to Paradise” and listed his AI system, the Creativity Machine, as the author. The Office refused. A federal district court agreed, and in March 2025 the D.C. Circuit affirmed, holding that “the Copyright Act of 1976 requires all eligible work to be authored in the first instance by a human being.”2U.S. Court of Appeals for the D.C. Circuit. Thaler v. Perlmutter The appellate court also rejected the argument that an AI could count as an “employee” for work-made-for-hire purposes, because the human authorship requirement applies to every category of copyrightable work.
The practical result: if an algorithm produced it end to end, nobody owns it. Anyone can copy, sell, or modify that output without asking. For creators and businesses building around generative tools, this is the single most important fact to internalize.
When AI-Assisted Work Can Be Copyrighted
Using software as a tool does not disqualify a work from protection, any more than using Photoshop or a camera does. The question is whether a human made meaningful creative decisions about the expression in the final piece.
The Copyright Office’s 2025 copyrightability report set out this line. Copyright still protects original human expression when the same work also contains AI-generated material, but it does not extend to the AI-generated portions or to material where the human lacked control over the expressive choices.3U.S. Copyright Office. Copyright and Artificial Intelligence, Part 2: Copyrightability Every registration is evaluated on what the human actually contributed.
The Zarya of the Dawn Registration
The graphic novel Zarya of the Dawn shows how this works. Kris Kashtanova used Midjourney to generate the book’s images, wrote the text, and arranged everything into a narrative. The Copyright Office first granted a full registration, then reconsidered and reissued it with limits.4U.S. Copyright Office. Zarya of the Dawn Registration Decision The written text and the selection and arrangement of the images were protected; the individual AI-generated images were not, because the human could not control what visual output the software produced.
Prompts Alone Are Not Authorship
A common assumption is that writing a careful prompt makes you the author of what comes out. The Copyright Office does not agree. Its 2025 report treats prompts as instructions that convey ideas rather than as authorship of expression, because the same prompt can produce very different results, the model may ignore parts of the instruction, and the user ends up accepting what the machine chose rather than fixing the expression themselves.3U.S. Copyright Office. Copyright and Artificial Intelligence, Part 2: Copyrightability
Protection is available when the human does more than prompt. Selecting particular outputs and arranging them, editing or painting over AI-generated elements, and writing accompanying text can each count as protectable authorship. What the Office and the courts look for is a human creative decision that shows up in the finished work.
Registering a Work That Used AI
Since March 2023, the Copyright Office has required applicants to disclose AI-generated content in their submissions.5Federal Register. Copyright Registration Guidance: Works Containing Material Generated by Artificial Intelligence Skipping this step can cost you the registration, and possibly the ability to enforce it.
Use the Standard Application rather than the Single Application, which lacks the necessary fields. Describe what the human contributed in the “Author Created” field. Disclaim the AI-generated portions in the “Limitation of the Claim” section under “Material Excluded,” and add detail in the “Note to CO” field if needed. If the AI-generated content is trivial enough to be de minimis, disclosure is not required, though the Office has not drawn a bright line for what counts as trivial. If you already filed without disclosing, contact the Public Information Office for a pending application, or file a supplementary registration if it has already processed.
The stakes on disclosure are real. Under Section 411(b) of the Copyright Act, a court can invalidate your registration entirely if it finds you knowingly left out information that would have caused the Office to refuse it.6U.S. Copyright Office. Copyright Registration Guidance: Works Containing Material Generated by Artificial Intelligence
What Platform Terms of Service Actually Give You
Reading OpenAI’s terms, it looks like users own everything the tool produces. OpenAI assigns “all our right, title, and interest, if any” in outputs to the user.7OpenAI. Terms of Use The words “if any” and “to the extent permitted by applicable law” carry the whole point: a contract cannot create copyright where the law does not recognize one.
The clause gives you a right against the platform. It does not give you a right against the world. If the output lacks the human authorship copyright requires, anyone else can copy it freely, and the terms of service will not stop them. Businesses building around AI content should not treat a platform’s ownership language as a substitute for registration.
The Fight Over Training Data
The larger question in AI copyright law is whether feeding copyrighted works into a training set without permission is itself infringement. Model developers scrape enormous quantities of text and images from the open web. Multiple lawsuits argue this is unauthorized copying under the Copyright Act.
Developers rely on fair use, which allows unlicensed use of copyrighted material when the four statutory factors line up: the purpose and character of the use, the nature of the work, how much was used, and the effect on the market for the original.8Office of the Law Revision Counsel. 17 U.S.C. 107 – Limitations on Exclusive Rights: Fair Use Their argument is that training is transformative because the model learns statistical patterns rather than storing the works. Plaintiffs answer that the purpose of the whole exercise is to build a commercial product that competes with what it ingested.
The Copyright Office’s 2025 training report accepted that different stages of AI development raise distinct fair use questions and said the analysis should look at the overall commercial use, not just the copying of data into a dataset.9U.S. Copyright Office. Copyright and Artificial Intelligence, Part 3: Generative AI Training
Thomson Reuters v. Ross Intelligence
The clearest ruling so far came out of Delaware federal court. Ross Intelligence used thousands of Westlaw headnotes to train a competing legal research tool. The court rejected the fair use defense on summary judgment, finding the use was not transformative because it served the same purpose as the originals and that Ross intended to build a market substitute for Westlaw.10U.S. District Court for the District of Delaware. Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc. The court also held that the effect on a potential licensing market for AI training data was enough to weigh the fourth factor against Ross, even though Thomson Reuters had not yet licensed its data for that use. If that reasoning spreads, it narrows the fair use argument for training substantially.
The Other Big Cases
Visual artists sued Stability AI and other image generators in Andersen v. Stability AI. The court let the direct infringement claims proceed past the motion to dismiss but dismissed the DMCA Section 1202 claims with prejudice.11Justia. Andersen et al v. Stability AI Ltd. et al A third amended complaint was filed in February 2026, and the case has not gone to trial.
The New York Times sued OpenAI and Microsoft in late 2023, alleging ChatGPT was trained on Times articles without permission and can reproduce substantial portions verbatim. In March 2025 the Southern District of New York denied OpenAI’s motion to dismiss the core infringement claims. Music publishers have also sued Anthropic over Claude’s ability to reproduce copyrighted song lyrics. These cases involve different types of works and different model architectures, and no single ruling will settle the training question across the industry.
One related theory, that data cleaning strips out copyright management information in violation of Section 1202, has struggled so far.12Office of the Law Revision Counsel. 17 U.S.C. 1202 – Integrity of Copyright Management Information The Andersen court dismissed those claims with prejudice, though the Copyright Office has flagged the issue for future analysis.9U.S. Copyright Office. Copyright and Artificial Intelligence, Part 3: Generative AI Training
Liability for Infringing AI Output
Separate from training, copyright liability can attach when a model’s output is too similar to an existing protected work. Courts use a substantial similarity test that asks whether an ordinary observer would recognize the new work as taken from the original, with both an objective comparison of expressive elements and a subjective evaluation of total concept and feel.13Ninth Circuit District and Bankruptcy Courts. Manual of Model Civil Jury Instructions – 17.19 Substantial Similarity
Prompting an AI to produce something “in the style of” an artist is not automatically infringement. Copyright does not protect style, technique, or method; it protects specific expression. The risk shows up when the output replicates identifiable expressive elements from particular protected works, which can happen without any conscious intent on the user’s part because the training material is inside the model.
Statutory Damages
Owners who registered their works before the infringement can elect statutory damages instead of proving actual harm. The standard range is $750 to $30,000 per work. If the infringement is proven willful, the ceiling rises to $150,000 per work.14Office of the Law Revision Counsel. 17 U.S.C. 504 – Remedies for Infringement: Damages and Profits Across a model trained on thousands of works, the aggregate exposure in the pending class actions is enormous.
Users, Developers, or Both
A user who publishes or sells infringing AI output is directly liable regardless of intent. Not realizing the model reproduced someone else’s work is not a defense, though it may lower the damages a court awards. Businesses that skip a human review step before releasing AI-generated content carry particular risk.
Developers face potential secondary liability. Contributory infringement requires knowledge of infringing activity and a material contribution to it. Vicarious liability applies when a party has the right to supervise the infringing conduct and gets a direct financial benefit from it, even without knowledge of specific acts. Whether AI platforms meet these standards is unresolved, but plaintiffs have plausible arguments given that companies profit from the tools and retain the ability to add guardrails.
Keeping Your Work Out of Training Sets
Creators who do not want their work ingested have limited options in the United States. The main tool is robots.txt, the long-standing protocol that tells automated crawlers where they may go. Major AI companies generally honor it as an industry practice, but robots.txt carries no legal force here. It is a practical barrier, not a legal one. The European Union’s Digital Single Market Directive gives machine-readable rights reservations actual legal effect; U.S. law has no equivalent, and until Congress or a court changes that, American creators are relying on voluntary compliance.
Working Under an Unsettled Rule
The Copyright Office’s 2025 report concluded that existing law can handle the questions AI raises and that no legislative change is needed at this time.3U.S. Copyright Office. Copyright and Artificial Intelligence, Part 2: Copyrightability Bills touching AI have been introduced in Congress, but none addressing copyright ownership of AI-generated works had passed as of early 2026. No appellate court has ruled on whether AI training is fair use. Thomson Reuters is a Delaware district court opinion, and the major cases in California and New York remain pre-trial.
Given that, treat the uncertainty as a live risk factor. Build human authorship into your creative process, disclose AI use on your copyright applications, keep records of the creative choices you made, and check AI output against known works before you publish or sell it.