Artificial Intelligence and Copyright

By Hank FasthoffJuly 4, 2026

People are using artificial intelligence to produce text, images, music, code, and video at an unprecedented scale, and copyright law hasn't caught up. Two foundational questions relating to use of AI to produce copyrightable materials are in active litigation across the federal courts. First, can a work generated by AI be copyrighted, and if so, who is the author? Second, does training an AI model on copyrighted works constitute infringement? As of the publication date of this article, neither question has been definitively resolved, and the answers to those questions will determine how businesses that use AI tools manage both the rights in what they produce and the liability for what their models consume.

AI Cannot Be an Author

Under the Copyright Act, copyright protection subsists in original works of authorship fixed in a tangible medium of expression. 17 U.S.C. § 102(a). Neither the Constitution nor the Copyright Act defines "author," but courts, the Copyright Office, and legal tradition have consistently interpreted the term to require a human being.

In Thaler v. Perlmutter, 130 F.4th 1039 (D.C. Cir. 2025), the D.C. Circuit affirmed the Copyright Office's refusal to register a visual artwork titled "A Recent Entrance to Paradise," which was created entirely by an AI system called the Creativity Machine without any creative contribution from a human. Dr. Stephen Thaler, the computer scientist who built the system, listed the Creativity Machine as the sole author and himself as the copyright claimant.

Judge Millett's opinion held that the Copyright Act requires all work to be authored in the first instance by a human being. Multiple provisions of the Act, including the ownership, duration, and work-for-hire provisions, presuppose human authorship. A machine isn't a person and can't hold property, enter into contracts, or respond to economic incentives. Because Dr. Thaler represented that the work was created "autonomously by artificial intelligence without any creative contribution from a human actor," the court affirmed the denial without reaching the question of whether Dr. Thaler could claim authorship by virtue of building and operating the system. On March 2, 2026, the Supreme Court denied certiorari, leaving the D.C. Circuit's holding intact.

Works Made with AI May Be Copyrightable

While a work generated entirely by AI isn't copyrightable, a work created by a human using AI as a tool may be. In March 2023, the Copyright Office published registration guidance addressing works containing material generated by artificial intelligence. 88 Fed. Reg. 16,190 (March 16, 2023). Under that guidance, whether a work made with AI is registrable depends on the circumstances, particularly how the AI tool was used and how it contributed to the final work.

Applicants must disclose the inclusion of AI-generated content and describe the human author's contributions. If a work contains both human-authored and AI-generated elements, the Copyright Office may register the human-authored portions while excluding the AI-generated material from the registration.

In the case of Zarya of the Dawn, a graphic novel created by Kris Kashtanova using Midjourney to generate images and human authorship for the text and selection/arrangement of images, the Copyright Office granted partial registration. Registration covered Kashtanova's original text and the selection, coordination, and arrangement of the text and images, but didn't cover the individual AI-generated images themselves, because the author didn't exercise sufficient creative control over how Midjourney rendered each image.

In January 2025, the Copyright Office published a report concluding that prompts alone are insufficient to establish authorship. Prompts "essentially function as instructions that convey unprotectible ideas," and current generative AI technologies don't offer enough control and predictability over outputs for a prompt to constitute the kind of creative direction that establishes authorship. A pending case, Allen v. Perlmutter, challenges the Copyright Office's refusal to register a work generated through more than 600 AI prompts and may provide further guidance on where the line falls.

What This Means for Businesses Using AI Tools

A company that uses a generative AI tool to produce marketing copy, product descriptions, images, code, or other content faces an ownership question. If the AI tool generated the content with minimal human creative input, the output may not be copyrightable, meaning the company can't register it, can't enforce it against competitors who copy it, and can't rely on copyright to prevent others from using identical or similar content.

If a human employee selected, arranged, revised, or substantially directed the AI output, the resulting work may contain sufficient human authorship to support registration. But the human contribution must go beyond prompting. Under the Copyright Office's current guidance, the human must exercise creative control over the expressive elements of the work, not just instruct the AI on what subject to address.

For code generated by AI coding assistants, the same analysis applies. Code that a developer prompts an AI tool to write and then uses without modification may not be copyrightable. Code that a developer writes with AI assistance, modifying, selecting, and integrating AI-generated suggestions into a larger human-authored program, is more likely to contain sufficient human authorship to support registration.

AI Training and Copyright Infringement

Training a large language model or an image generation model involves ingesting and computationally processing vast quantities of text, images, code, and other copyrighted works. AI companies have scraped websites, digitized books, downloaded image databases, and incorporated copyrighted content from across the Internet into training datasets, often without the knowledge or consent of the copyright owners.

Whether this training process infringes copyright is the central question in more than 50 active lawsuits filed against AI companies in federal courts across the country. As of late 2025, three judges have ruled on fair use in the AI training context, with two ruling for the AI companies and one ruling against, and no further summary judgment decisions were expected until mid-2026.

In New York Times Co. v. Microsoft Corp. (S.D.N.Y., filed December 2023), the Times alleges that OpenAI trained its models on millions of Times articles without authorization, and that OpenAI's products can reproduce near-verbatim portions of copyrighted articles, effectively allowing users to circumvent the Times's paywall. In March 2025, the court denied most of OpenAI's motions to dismiss, allowing the copyright infringement claims to proceed to discovery. The case remains pending.

In Authors Guild v. OpenAI (S.D.N.Y.), a consolidated class action brought by novelists, nonfiction authors, and the Authors Guild alleges that OpenAI infringed copyrights by ingesting copyrighted books into its training data. In Concord Music Group v. Anthropic (N.D. Cal., transferred from the Middle District of Tennessee), music publishers allege that Anthropic's AI model can reproduce copyrighted song lyrics.

In Andersen v. Stability AI (N.D. Cal.), visual artists allege that Stability AI, Midjourney, and DeviantArt trained image-generation models on billions of copyrighted images scraped from the Internet. In Thomson Reuters v. Ross Intelligence (D. Del.), Thomson Reuters alleges that Ross Intelligence copied content from Westlaw to train a legal research AI tool.

AI companies have raised fair use as their primary defense. OpenAI has argued that AI training is a transformative, non-expressive analytical use that doesn't substitute for the original works. Copyright owners have argued that ingesting copyrighted works to build a commercial product that competes with those works isn't transformative, and that AI outputs that reproduce or closely paraphrase copyrighted content directly substitute for the originals.

As of mid-2026, no appellate court has ruled on whether AI training constitutes fair use. District courts have reached conflicting results. In Kadrey v. Meta Platforms (N.D. Cal.) and Bartz v. Anthropic (N.D. Cal.), judges found that AI training was highly transformative, which weighs in favor of fair use. The Bartz ruling was split, however, as Judge Alsup held that training on books Anthropic had lawfully acquired was fair use, but that downloading and retaining pirated copies from the LibGen and PiLiMi shadow libraries was not. It was that second holding, on acquisition rather than training, that exposed Anthropic to class-wide liability and led the company to agree to a $1.5 billion settlement. As of this writing, the court has held a fairness hearing but hasn't entered final approval, and objectors remain, so appeals haven't been foreclosed.

Anthropic denied wrongdoing in agreeing to settle, and no court adjudicated the infringement question, so the settlement establishes no precedent that binds any future court. The release addresses only past conduct, it doesn't grant Anthropic a license to train on copyrighted works going forward, and it has no effect on any other AI company. In Thomson Reuters v. Ross Intelligence (D. Del.), the court found that Ross's copying wasn't fair use where the AI product directly competed with the copyrighted legal research product it was trained on, and the Third Circuit heard argument on that ruling in June 2026, the first appellate test of the question.

The Input Problem and the Output Problem

AI copyright disputes involve two distinct infringement theories. Input infringement concerns whether the act of copying copyrighted works into a training dataset is itself an infringing reproduction. Output infringement concerns whether the AI model's outputs are substantially similar to copyrighted works in the training data, either because the model "memorized" portions of the training data or because its outputs are derivative of copyrighted source material.

Input and output infringement raise different fair use questions. An AI company might argue that copying works into a training dataset is transformative because the model learns statistical patterns rather than storing or reproducing the works. But if the model then generates outputs that reproduce or closely paraphrase copyrighted text, the output may infringe regardless of whether the training process was fair use.

For businesses that use AI-generated content, the output problem creates direct liability exposure. If an AI tool generates marketing copy that reproduces copyrighted text from its training data, the business that publishes that copy may be liable for infringement, even if the business didn't know the content was derived from a copyrighted source.

What Businesses Should Do Now

Register copyrightable works promptly, including works that AI tools might ingest. Registration preserves the right to statutory damages and attorney's fees, and an owner who discovers that an AI company used copyrighted content for training will need a registration to bring suit.

Disclose AI-generated content in copyright registration applications. Under the Copyright Office's current guidance, failure to disclose AI-generated material in a registration application may jeopardize the validity of the registration.

Evaluate AI-generated output for potential infringement before publishing it. AI tools can reproduce copyrighted text, code, or visual elements from their training data, and the user who publishes the output bears infringement risk. Review AI-generated content the same way you would review content from any other source whose originality you can't independently verify.

Monitor AI companies' use of your copyrighted works. Several services now track whether copyrighted content appears in AI training datasets. Copyright owners who discover unauthorized use should evaluate whether the use constitutes infringement and whether enforcement is economically justified given the available remedies.

Understand that copyright law in this area is unsettled. No appellate court has ruled on whether AI training is fair use. No bright-line rule exists for how much human contribution is required for an AI-assisted work to qualify for copyright protection. Businesses that rely on AI tools for content creation and businesses whose content AI tools consume are both operating in a legal environment where the rules are being written through litigation, and the outcomes will depend on facts, circuits, and cases that haven't yet been decided.