Under current U.S. law, generative AI and copyright fit together in a narrow way: a work produced entirely by an AI system cannot be copyrighted by anyone, a work in which a person used AI as a creative tool can be registered for the human-authored portions only, and publishing AI output can expose you to infringement liability even when you had no idea the output resembled someone else’s protected work. The rules on registration are settled enough to follow today. The rules on whether AI companies can lawfully train on copyrighted material are not, and federal judges are actively splitting on the question.
Why Pure AI Output Belongs to No One
Copyright in the United States requires a human author. The Copyright Office refuses to register works without one, and in 2025 the D.C. Circuit affirmed that position in Thaler v. Perlmutter, holding that an image produced entirely by an AI system called the “Creativity Machine” could not be copyrighted because the Copyright Act “requires all eligible work to be authored in the first instance by a human being.”1United States Courts. Thaler v Perlmutter The applicant had listed the AI as sole author. Both the office and the courts said no.
The practical result: if you type a prompt into an image generator or chatbot and the software produces the output with no meaningful creative input from you beyond the prompt, that output belongs to nobody. It is in the public domain from the moment it exists. Anyone can copy it, sell it, or build on it, and you have no legal ground to stop them.
What Parts of AI-Assisted Work You Can Actually Own
The Copyright Office’s 2025 report on copyrightability concluded that “the use of AI tools to assist rather than stand in for human creativity does not affect the availability of copyright protection for the output,” while making clear that prompts alone are not enough: “Based on the functioning of current generally available technology, prompts do not alone provide sufficient control.”2United States Copyright Office. Copyright and Artificial Intelligence, Part 2 Copyrightability Report
The Zarya of the Dawn decision shows the line. Kris Kashtanova created a comic book using her own written text combined with images generated by Midjourney. The Copyright Office granted protection for the text and for the “selection, coordination, and arrangement” of written and visual elements, but excluded the AI-generated artwork itself, requiring a new registration that disclaimed “artwork generated by artificial intelligence.”3United States Copyright Office. Zarya of the Dawn The images themselves belonged to no one.
Three kinds of human contribution can qualify for protection in an AI-assisted work:
- Substantially reworking AI-generated content, so your changes add original creative expression.
- Choosing and organizing AI-generated elements in a way that reflects your own creative judgment.
- Embedding AI-generated pieces inside a work that is predominantly human-authored, the way Kashtanova’s written text carried her comic.
There is no numerical threshold. The office has not said you need to change some fixed percentage. Each application gets evaluated on whether the human controlled the expressive elements that make the work original.
Registering an AI-Assisted Work
If your work combines human creativity with AI-generated elements and you want to register it, the Copyright Office requires transparency about which is which. The March 2023 registration guidance requires applicants to disclose any AI-generated content that goes beyond a trivial amount and to describe the human author’s actual contribution.4Federal Register. Copyright Registration Guidance Works Containing Material Generated by Artificial Intelligence
The Two Fields That Matter
When you file through the Electronic Copyright Office (eCO) system, two fields carry the weight.5U.S. Copyright Office. Register Your Work Registration Portal “Author Created” is where you describe your human contributions: the text you wrote, the arrangement you designed, the modifications you made to AI-generated material. Be concrete. “Text” or “selection, coordination, and arrangement of text and images” are the kind of specific descriptions the office is looking for. “Limitation of Claim” is where you exclude the AI-generated portions from the registration, disclaiming the elements the AI produced without your direct creative control.
Getting the disclosure wrong creates real problems. If you fail to identify the AI-generated portions, the examiner will likely refuse registration. If the office discovers the omission after the fact, it can cancel the certificate. And because a valid registration is a prerequisite for filing an infringement lawsuit over a U.S. work, an inaccurate registration can leave you unable to sue when you need to.6Office of the Law Revision Counsel. 17 USC 411 – Registration and Civil Infringement Actions
Fees and Timing
The filing fee for the most common scenario, a single work by a single author who is also the claimant, is $45 through eCO.7U.S. Copyright Office. Fees Other application types cost more. You will need to upload a digital deposit copy of the work. Electronic applications that require no back-and-forth with the examiner currently average about two months to process; applications that require correspondence average about four months, and individual cases can run to eight.8U.S. Copyright Office. Registration Processing Times
Your Liability When You Publish AI Output
Even if the ownership question does not matter to you, a separate risk shows up every time you publish AI-generated content. If the output closely resembles an existing copyrighted work, the person who published it, meaning you, can be liable for infringement. A copyright holder needs to show that the AI had access to their work (usually easy to prove when models train on the open internet) and that the output is substantially similar to the protected original.
Statutory damages range from $750 to $30,000 per work infringed, and a court can raise the ceiling to $150,000 per work for willful infringement.9Office of the Law Revision Counsel. 17 US Code 504 – Remedies for Infringement Damages and Profits Not knowing you infringed offers limited comfort. Innocent infringement can reduce the minimum statutory award to $200 per work, but it does not eliminate liability. You still infringed. You just did it unknowingly.
Who bears the risk depends heavily on the platform. Several major AI companies now offer some form of copyright indemnification for paying customers. OpenAI’s “Copyright Shield” covers ChatGPT Enterprise and API users. Microsoft, Google, Adobe, and GitHub offer similar protections for their commercial AI products. These programs share the same limits: they apply only to paid tiers, they often require you to use built-in safety features like duplication-detection filters, and they come with liability caps buried in the terms of service. If you use a free tool or a smaller platform without indemnification, you carry the full risk yourself.
One more registration point matters here. To sue for infringement of your own work, you generally need a registration (or at least an application that has been filed and refused) in hand first.6Office of the Law Revision Counsel. 17 USC 411 – Registration and Civil Infringement Actions Registering promptly also unlocks statutory damages and attorney’s fees, which are unavailable if you wait until after the infringement to file.
The Training-Data Fight Is Unsettled
Building a generative AI model requires enormous amounts of data, and developers scrape text, images, code, and music from across the internet to train their systems. Much of that data is copyrighted. The central legal question is whether ingestion qualifies as fair use under federal copyright law, which allows limited use of copyrighted material without permission when the use is sufficiently transformative.10Office of the Law Revision Counsel. 17 US Code 107 – Limitations on Exclusive Rights Fair Use Courts weigh four factors: the purpose and character of the use, the nature of the copyrighted work, the amount taken relative to the whole, and the effect on the market for the original.
Federal judges have gone opposite ways, and no appellate court has issued a definitive ruling on AI training specifically. In Thomson Reuters v. ROSS Intelligence, a Delaware federal court granted summary judgment against the AI developer, finding that ROSS’s use of copyrighted legal headnotes to train a competing legal research tool was not fair use. The court emphasized the fourth factor, concluding that the “effect on a potential market for AI training data” weighed against the defendant even though Thomson Reuters had not yet licensed its data for that purpose.11United States District Court for the District of Delaware. Thomson Reuters Enterprise Centre GmbH v ROSS Intelligence Inc
Other courts have gone the other way. In mid-2025, federal judges in Bartz v. Anthropic and Kadrey v. Meta Platforms both found AI model training to be “highly transformative,” a factor favoring fair use. Those rulings treated ingesting copyrighted text as fundamentally different from copying it, because the models produce new expression rather than regurgitating originals. The higher-profile New York Times v. OpenAI case remains in active litigation, with the court narrowing several claims but leaving the core fair use dispute intact. In Andersen v. Stability AI, a California federal court allowed direct copyright infringement claims against Stability AI, Midjourney, and DeviantArt to move past the motion-to-dismiss stage, while dismissing related DMCA claims.12Justia Law. Andersen et al v Stability AI Ltd et al None of these cases has reached a final judgment on the merits. Until they do, the legal obligations of AI companies toward the owners of training data are genuinely unresolved.
Keeping Your Own Work Out of Training Sets
If you publish creative work online and want to keep it out of AI training datasets, technical opt-out tools exist but have real limits. The most common approach is the robots.txt file, a decades-old protocol that tells web crawlers which parts of your site they may access. Major AI companies have published specific crawler names (GPTBot for OpenAI, ClaudeBot for Anthropic, Google-Extended for Google) that you can block individually.
Newer proposals go further. Microsoft has proposed a “DisallowAITraining” directive that would let site owners block all AI training crawlers with a single rule. Google has proposed “Content-Usage,” and Cloudflare has introduced “Content-Signal,” which lets publishers set granular preferences for search indexing, real-time AI use, and model training separately. None of these are finalized standards, and all share the same weakness: compliance is voluntary. A crawler that ignores your robots.txt file faces no automatic technical barrier. The protocol is an honor system, not a locked door. Whether ignoring an opt-out signal creates legal liability for the scraper is an open question no court has definitively answered.
Documenting Your Creative Process
Well before you file a registration, build a paper trail. The Copyright Office’s copyrightability report stresses that proof of human input will be essential for asserting copyright protection, and the burden falls entirely on you. If your registration is ever challenged, or if you need to enforce your copyright in court, documentation of your creative process will matter enormously.
Save everything. Drafts showing how you modified AI-generated material. Screenshots of your prompt history and the iterative changes you made. Notes explaining your selection and arrangement choices. Earlier versions of the work that show progression from raw AI output to finished piece. The documentation is proof that the creative choices were yours.
What to Expect Next
The legal framework around AI and copyright is being shaped in courtrooms faster than in Congress. No comprehensive federal AI copyright legislation has been enacted as of early 2026. Proposals such as the CLEAR Act (S. 3813), which would require AI developers to report copyrighted works used in training datasets to the Copyright Office, have been introduced but not passed. Executive Order 14110, which had directed the Copyright Office and Patent Office to study AI-related intellectual property issues, was revoked in January 2025.13The White House. Removing Barriers to American Leadership in Artificial Intelligence The Copyright Office’s multi-part study continues; Part 2 addressed copyrightability, and a later part on training data, licensing, and liability allocation has not yet been published.2United States Copyright Office. Copyright and Artificial Intelligence, Part 2 Copyrightability Report
Until that report arrives and the major pending lawsuits reach appellate decisions, the safest approach for anyone creating with AI tools is the same one that already works at the registration counter: maximize and document your own creative contributions, disclose the AI-generated components honestly, and treat every output as potentially containing echoes of someone else’s protected work.