AI-generated art carries copyright infringement risk on two fronts at once: the training that feeds copyrighted images into the model, and the outputs that can look uncomfortably close to specific protected works. Federal courts began sorting both questions in earnest in 2025, with a landmark class action trial set for September 2026. And there’s a twist that catches many users off guard — an image you generate purely by prompting an AI usually gets no copyright protection at all, which means anyone can copy it.
The Training Data Problem
Every image generator is built by feeding the software vast quantities of existing images paired with descriptive text. One widely used dataset, LAION-5B, contains roughly 5.85 billion image-text pairs scraped from the open internet.1arXiv. LAION-5B: An Open Large-Scale Dataset for Training Next Generation Image-Text Models Those images include copyrighted photographs, illustrations, and fine art posted by their creators on personal websites, portfolio platforms, and social media. In most cases, no one asked for permission.
Copyright holders have exclusive rights to reproduce their work and to prepare derivative works from it.2Office of the Law Revision Counsel. 17 USC 106 – Exclusive Rights in Copyrighted Works Copying billions of images into a training pipeline plainly implicates the reproduction right. The question that decides everything is whether that copying is fair use.
The Fair Use Fight
Fair use allows limited copying without permission when the use serves a favored purpose. Courts weigh four factors: the purpose and character of the use, the nature of the copyrighted work, how much was taken, and the effect on the market for the original.3Office of the Law Revision Counsel. 17 US Code 107 – Limitations on Exclusive Rights: Fair Use AI developers argue training is transformative because the model doesn’t store actual copies; it learns statistical relationships between pixels and words, then generates something new. Artists respond that using their work to build a tool designed to compete with them is the opposite of transformative.
The clearest signal so far arrived in 2025, in Thomson Reuters v. Ross Intelligence. Ross built a legal research AI by training on Thomson Reuters’ copyrighted headnotes. The court rejected Ross’s fair use defense, finding that using copyrighted material to develop a competing product was not transformative and that the market-harm factor weighed heavily against it.4U.S. District Court for the District of Delaware. Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc. The case involved text, but the reasoning maps directly onto image generators trained on artists’ work to build competing creative tools.
If training is ultimately found not to be fair use, the financial exposure is enormous. Statutory damages run from $750 to $30,000 per work, and courts can push that ceiling to $150,000 per work for willful infringement.5Office of the Law Revision Counsel. 17 USC 504 – Remedies for Infringement: Damages and Profits Multiplied by the millions of works in a single training dataset, the potential liability becomes staggering.
When an AI Output Infringes
Even if training is eventually blessed as fair use, an individual AI-generated image can still infringe if it copies too much from a specific work. Infringement means violating any of the copyright holder’s exclusive rights.6Office of the Law Revision Counsel. 17 US Code 501 – Infringement of Copyright To prove it, an artist typically has to show that the AI had access to the original and that the output is substantially similar to it.
Access is usually easy to establish. Any model trained on open-internet data can be presumed to have ingested popular images. Substantial similarity is harder. Courts in many circuits use a two-step comparison: an objective look at specific expressive elements like composition, arrangement, and distinctive details, followed by a subjective judgment about whether an ordinary observer would feel the works share the same concept and feel.7Ninth Circuit District and Bankruptcy Courts. 17.19 Substantial Similarity – Extrinsic Test; Intrinsic Test
One distinction matters more than any other: copyright does not protect style. Anyone can paint in a Cubist style or make moody watercolor landscapes without infringing on a particular artist. What’s protected is the specific expression of an idea — the exact composition and detail of a specific painting. An output that captures the general feel of an artist without reproducing identifiable elements from a specific work probably doesn’t cross the line. An output that recreates a character design or a recognizable scene almost certainly does.
Memorization and Near-Copies
Sometimes models don’t just approximate a style. They memorize specific training images and reproduce near-identical copies when prompted a certain way. Researchers have shown that diffusion models occasionally regurgitate training data, particularly for images that appeared in the dataset many times. That scenario is the clearest infringement case, because a near-copy easily satisfies substantial similarity. Whether memorized outputs infringe is a question that exists independently from the fair use debate over training, and no court has fully resolved it yet.
Remedies
When an output does infringe, courts can order injunctions stopping distribution of the image.8Office of the Law Revision Counsel. 17 USC 502 – Remedies for Infringement: Injunctions The copyright holder can also recover actual damages, including lost profits and any profits the infringer earned from the image, or elect statutory damages instead.5Office of the Law Revision Counsel. 17 USC 504 – Remedies for Infringement: Damages and Profits
Who Actually Gets Sued
The current wave of lawsuits targets AI developers, but individual users are not automatically safe. If you use an AI tool to generate an image that substantially copies a copyrighted work, then publish or sell it, you can face an infringement claim whether or not the developer is also on the hook. Copyright infringement does not require intent; it requires reproduction or distribution of protected expression without authorization.
Most AI platform terms of service shift risk toward the user through indemnification clauses that require users to cover claims arising from how they use the output. Some providers have started offering limited copyright indemnity for enterprise customers, promising to cover legal costs if the AI produces something that triggers a lawsuit, but the scope and dollar limits vary widely. Read the terms before assuming the platform has your back.
Standard commercial general liability insurance often will not cover an AI-related copyright claim either. CGL coverage for copyright infringement usually applies only when the infringement occurs in an advertisement, which leaves broader use of AI-generated images exposed. Businesses relying heavily on AI visuals should look at specialized intellectual property insurance or errors-and-omissions coverage to close that gap.
Why Purely AI-Generated Images Can’t Be Copyrighted
Here is the flip side: if you generate an image entirely through AI, you almost certainly cannot register a copyright on it. Copyright protection covers original works of authorship fixed in a tangible medium,9Office of the Law Revision Counsel. 17 US Code 102 – Subject Matter of Copyright: In General and the Copyright Office reads “authorship” to mean creative work by a human being. The Office’s Compendium states plainly that it will not register works produced by a machine operating without creative human input.10U.S. Copyright Office. Compendium of US Copyright Office Practices – Chapter 300 Copyrightable Authorship: What Can Be Registered
Typing a text prompt is not enough. A photographer controls lighting, framing, timing, and composition. Someone typing “a sunset over mountains in the style of Monet” hands the compositional decisions to the algorithm. The Copyright Office treats the resulting image as the product of the machine, not the creative expression of the user. In practice, that pushes purely AI-generated images into the public domain, where any competitor can copy them freely.
Mixed Human and AI Works
Works that combine human creativity with AI-generated elements can get partial protection. The Copyright Office requires applicants to disclose AI-generated material and to exclude those portions from the copyright claim.11Federal Register. Copyright Registration Guidance: Works Containing Material Generated by Artificial Intelligence Failing to disclose AI involvement can lead to cancellation of the registration.12eCFR. 37 CFR 201.7 – Cancellation of Completed Registrations
The comic book Zarya of the Dawn shows how this plays out. Kristina Kashtanova used Midjourney to generate the comic’s images while writing the story text and arranging the panels herself. The Office granted copyright for the text and for the selection and arrangement of the visual and written elements, but denied protection for the individual AI-generated images.13United States Copyright Office. Zarya of the Dawn (Registration # VAu001480196) The original registration was cancelled and reissued to exclude the Midjourney-generated content.
An AI Cannot Be Named as Author
Stephen Thaler tried to register an image created entirely by his AI system, the “Creativity Machine,” naming the AI as sole author. The Copyright Office refused, and in March 2025 the D.C. Circuit affirmed, holding that the Copyright Act requires all eligible works to be authored by a human being.14United States Court of Appeals for the District of Columbia Circuit. Thaler v. Perlmutter That is currently the highest-level judicial confirmation that AI cannot be a recognized author under U.S. law.
The Case That Could Reshape the Field
The class action Andersen v. Stability AI pits a group of visual artists against Stability AI, Midjourney, and DeviantArt. The artists allege that scraping their copyrighted images to train image generators was mass copyright infringement. In August 2024, a federal judge denied motions to dismiss the direct and induced infringement claims against Stability AI and Midjourney, allowing the case to move into discovery. The court found the artists’ theories plausible on two fronts: that the model itself constitutes an infringing copy, and that distributing the model amounts to distributing the copyrighted works embedded in it. Trial is scheduled for September 2026. The outcome could define the legal footing for every major image generator on the market.
What Artists Can Do Now
The courts are moving slowly. Artists don’t have to wait for a verdict to take protective steps.
DMCA Takedowns
DMCA takedown notices still apply to AI-generated content that infringes your work. Platforms have to remove infringing content once they receive a valid notice identifying the specific material and the copyrighted work it infringes. The limitation is that each notice covers one piece of content at one location, and platforms are not required to monitor for repeat infringement. It’s a game of whack-a-mole, but it’s the fastest tool available for pulling infringing images off major platforms.
Opt-Out and Technical Protections
Website owners can use robots.txt to block specific AI crawlers from scraping their pages, and the protocol has been widely used for this purpose since the rise of generative AI. No law requires AI companies to honor it. Spawning AI’s “Have I Been Trained” tool lets artists search training datasets for their images and submit opt-out requests. Some platforms, including DeviantArt, have added HTML-level signals telling scraping bots not to download images for training. These are requests, not enforceable barriers.
For stronger protection, tools like Glaze and Nightshade from the University of Chicago apply invisible perturbations to images before upload. Glaze disrupts style mimicry by subtly altering how a model perceives the image; Nightshade goes further by “poisoning” training data so that models learn unpredictable behaviors from the altered images.15University of Chicago. Protecting Copyright – Nightshade The perturbations survive cropping, compression, screenshots, and even photos of a screen. Unlike opt-out requests, they don’t depend on the AI company’s cooperation.
The Copyright Claims Board
For individual artists who can’t take on federal litigation, the Copyright Claims Board offers a streamlined alternative. The CCB is a three-member tribunal inside the Copyright Office that handles disputes involving claims up to $30,000.16Copyright Claims Board. Copyright Claims Board Filing costs significantly less than a federal lawsuit, and the process is designed to work without an attorney. The CCB can hear infringement claims where a specific AI-generated image copies a registered work, which makes it a realistic option for creators whose damages don’t justify hiring a litigation team.
Content Provenance
The Coalition for Content Provenance and Authenticity (C2PA) has developed an open standard that attaches verifiable metadata to digital content, functioning like a nutrition label recording where an image came from and how it was edited.17C2PA. C2PA – Verifying Media Content Sources Artists who embed Content Credentials in their work create a tamper-evident chain of provenance that separates human-created originals from AI-generated copies. Major platforms and camera manufacturers are adopting the standard, and it may eventually help artists prove authorship in disputes.
Where Congress Stands
As of 2026, Congress has not enacted legislation specifically addressing AI and copyright. The Copyright Office has recommended against new legislation on the copyrightability of AI outputs, taking the position that the existing human authorship requirement is sufficient. On training data, the Office has encouraged voluntary licensing between rights holders and AI companies rather than a government-mandated framework, recommending the licensing market be allowed to develop without intervention for now.18Congress.gov. Generative Artificial Intelligence and Copyright Law Several bills addressing related issues like AI-generated digital replicas of real people have been introduced, including the NO FAKES Act and the No AI FRAUD Act, but none have been enacted.19Congress.gov. Artificial Intelligence Prompts Renewed Consideration of a Federal Right of Publicity For now the courts, not Congress, are shaping the rules, and the cases already in the pipeline will do more to set them than any pending bill.