When AI Trains on Your Art: Copyright, Style Imitation, and Legal Options for Visual Artists

The Problem: Your Art Is Already in the Training Data

If you are a visual artist who has ever posted work online — on ArtStation, DeviantArt, Instagram, or a personal portfolio site — there is a strong chance your images have been scraped into the training datasets behind generative AI image models like Stable Diffusion, Midjourney, and DALL-E. These models ingested billions of images from the web, often without the knowledge, consent, or compensation of the artists who created them. The models learned to generate images in the styles of specific working artists, sometimes producing outputs that closely mimic a named creator aesthetic.

This is not a hypothetical scenario. In early 2023, visual artists Sarah Andersen, Kelly McKernan, and Karla Ortiz filed a class-action lawsuit against Stability AI, Midjourney, and DeviantArt, alleging that the companies used the plaintiffs copyrighted works — without permission — to train their AI image generators. Many similar lawsuits have followed. As we write this in mid-2026, multiple artist-led cases are in active discovery, and the first major court rulings are expected to shape the legal landscape for years to come.

In this guide, we break down what legal rights exist today for visual artists whose work has been used to train AI image models — what the law protects, what it does not, where the key litigation stands, what the Copyright Office has said, and what practical steps you can take to protect your work going forward.

What Copyright Protects (and What It Does Not): Expression vs. Style

U.S. copyright law protects original works of authorship fixed in any tangible medium of expression, including paintings, illustrations, photographs, and digital art. 17 U.S.C. § 102(a). If you created an original image and fixed it in a tangible form — whether that is a canvas, a digital file, or a posted JPEG — you almost certainly hold a copyright in that specific work.

But here is the critical distinction for artists dealing with AI: copyright protects specific expressive works, not styles, ideas, or techniques. This principle, known as the idea-expression dichotomy, means that no one can copyright a loose brushwork style or a high-saturation sci-fi aesthetic in the abstract. What is protectable is the particular way you have executed those concepts in a specific image.

This matters enormously for AI art disputes. When someone prompts Midjourney to produce an image in the style of [Artist Name], the model is not necessarily reproducing any single copyrighted image. It is synthesizing patterns learned across thousands of examples. Under current U.S. law, a style — even a distinctive one — is generally not independently protectable. What could be protectable is if an AI model outputs an image that is substantially similar to one of your specific copyrighted works, which would raise a direct infringement claim.

The D.C. Circuit recently reaffirmed the human-authorship requirement in Thaler v. Perlmutter, upholding that only human-created works are eligible for copyright. This matters for artists in two ways: first, your human-authored works are and remain protectable; second, purely AI-generated images based on your style are not independently copyrightable by the AI company.

The Fair Use Battleground: Andersen v. Stability AI and Getty v. Stability AI

The central legal question in every AI training copyright case is whether training a generative model on copyrighted images without a license constitutes fair use under 17 U.S.C. § 107. Fair use is determined by weighing four factors: the purpose and character of the use, the nature of the copyrighted work, the amount and substantiality of the portion used, and the effect on the potential market for the original work. AI companies argue that training is transformative and does not substitute for the originals. Artists argue that the models were built on their labor, compete with them in the marketplace, and were created without consent.

Andersen v. Stability AI: Artists Claims Survive — So Far

In Andersen v. Stability AI, No. 3:23-cv-00201 (N.D. Cal.), Judge William Orrick issued a significant order on August 12, 2024, that allowed key artist claims to move forward while dismissing others. The Copyright Alliance analysis highlights three important takeaways:

First, the court rejected the argument that generative AI is analogous to past technologies like the VCR. Judge Orrick distinguished the case from the Sony Betamax precedent, explaining that unlike a VCR — a device capable of copying but not pre-loaded with copyrighted content — Stable Diffusion was built to a significant extent on copyrighted works and necessarily invokes copies or protected elements of those works. This is a critical distinction that could shape how courts evaluate fair use in AI training cases.

Second, the court allowed direct copyright infringement claims to proceed, including the theory that the Stable Diffusion model itself is an infringing work because the artists works are contained, in some manner within it. Judge Orrick noted that the works may be contained in Stable Diffusion as algorithmic or mathematical representations — and are therefore fixed in a different medium than they may have originally been produced in — which is not an impediment to the claim.

Third, the court dismissed DMCA claims for removal of copyright management information (CMI) with prejudice, finding no evidence that any Stable Diffusion output was identical to the plaintiffs works. This identicality requirement has been applied in the Ninth Circuit and presents a significant hurdle for artists seeking CMI-based damages.

Getty Images v. Stability AI: The UK High Court Weighs In

In November 2025, the UK High Court issued the first major ruling in Getty Images v. Stability AI, and the outcome was mixed. According to Mayer Brown analysis, the court rejected Getty core copyright infringement claim, holding that Stable Diffusion model weights are not infringing copies under UK law because they contain statistically trained parameters, not stored reproductions of photographs. The court found that an AI model contains statistically trained parameters, not stored copies or reconstructions of photographs.

However, the court did find limited and historic instances of trade mark infringement, because earlier versions of Stable Diffusion could generate images containing Getty or iStock watermarks under realistic prompting. This suggests that while the training process itself may not directly infringe under UK law, the outputs that replicate protected marks or watermarks can create liability.

For U.S. artists, the UK ruling is instructive but not binding. The parallel U.S. case, Getty Images v. Stability AI (D. Del.), is still progressing, and U.S. courts may reach different conclusions under American fair use doctrine. The key takeaway: the legal landscape is still in flux, and no court has yet issued a final ruling on whether AI training on copyrighted images constitutes fair use in the United States. We expect key rulings in 2026.

The Copyright Office Weighs In: Part 2 Report (January 2025)

On January 29, 2025, the U.S. Copyright Office released Part 2 of its Report, Copyright and Artificial Intelligence: Copyrightability. As the Library of Congress blog explains, the report addresses the copyrightability of outputs created using generative AI and analyzes the type and level of human contribution sufficient for outputs created using generative AI to be eligible for copyright protection.

The Office key conclusions are directly relevant to visual artists:

• AI-generated outputs are copyrightable only where a human author has determined sufficient expressive elements. Merely providing prompts to an AI system is not enough. This means that someone who types a prompt into Midjourney cannot claim copyright in the resulting image.
• The use of AI to assist in creation does not bar copyrightability. If you incorporate AI-generated elements into a larger human-authored work — say, using AI as one tool among many in a creative process — the portions you authored are still protectable.
• Existing law is adequate. The Office concluded that the case has not been made for legislative changes to provide additional protection for AI-generated outputs. This means the current framework — human authorship required, no protection for purely AI-generated content — remains in force.

Crucially, the Office noted that its forthcoming Part 3 of the report will address the legal implications of training AI models on copyright-protected works, licensing considerations, and the allocation of any potential liability. That section — which directly addresses the core question for artists — has not yet been released as of this writing. For a deeper dive on the training-data side of this analysis, see our earlier coverage of AI Training Data and Copyright: Fair Use, Licensing, and Governance for Model Developers.

Technical Opt-Out Countermeasures: Glaze, Nightshade, and Beyond

While the legal system works through these questions at a pace that can feel glacial to working artists, several technical tools have emerged that let you take protection into your own hands. We have covered the broader opt-out landscape in our Visual Artist AI Opt-Out Guide, but here we will focus on the two most prominent tools and what they actually do.

Glaze: Defensive Protection Against Style Mimicry

Glaze, developed by the SAND Lab at the University of Chicago, is a defensive tool that applies imperceptible perturbations to your images before you post them online. These perturbations are designed to disrupt AI models ability to learn and replicate your artistic style. To the human eye, a Glazed image looks essentially identical to the original. To an AI model, the image features are scrambled in ways that make style mimicry much harder. Glaze is currently on version 2.2 and is available as a free download for Windows and macOS.

Nightshade: Offensive Disruption of Unauthorized Training

Nightshade, also from the SAND Lab, takes a different approach. Rather than defending individual images, Nightshade is an offensive tool designed to be used collectively. It transforms images into poison samples that, when ingested by AI models during training, cause the models to learn unpredictable and incorrect behaviors. As the Nightshade team explains: a prompt that asks for an image of a cow flying in space might instead get an image of a handbag floating in space.

The goal of Nightshade is not to break AI models — it is to increase the cost of training on unlicensed data, making licensing images directly from artists a more economically rational alternative for AI developers. Nightshade effects are robust to common image modifications: cropping, resampling, compression, and even screenshots of displayed images retain the poison effect.

Limitations and Practical Reality

These tools are not silver bullets. Both Glaze and Nightshade acknowledge that their protections may not remain effective indefinitely as AI companies develop countermeasures. Nightshade team notes that as with any security attack or defense, Nightshade is unlikely to stay future proof over long periods of time. Additionally, images with flat colors and smooth backgrounds may show more visible artifacts when processed. Artists should use these tools as part of a broader protection strategy — not as a sole line of defense.

Contract Clauses: Protecting Your Work in Commissions and Licensing

Beyond technical tools and litigation, one of the most effective things you can do right now is to update your contracts. Whether you are taking on a commission, licensing an image for commercial use, or selling a piece, you can add provisions that specifically address AI training. Here are the key clauses to consider:

AI Training Prohibition

Include an explicit prohibition on using the commissioned or licensed work — or any derivative of it — to train, develop, or improve any machine learning model or AI system. This clause should be broad enough to cover not just the client but any downstream users or sublicensees. A sample formulation: Licensee shall not, and shall not permit any third party to, use the Licensed Work or any derivative thereof to train, fine-tune, develop, or improve any artificial intelligence system, machine learning model, or generative AI tool.

Output Restriction

Specify that the client or licensee may not use any AI tool to generate images in the style of you or that substantially replicate your work. This is harder to enforce than a training prohibition, but it puts the restriction on the record and gives you a breach-of-contract claim if the client violates it.

Retention of Training Rights

Explicitly state that you retain all rights not granted in the agreement, including the right to control whether your work is used in AI training datasets. This prevents any implied-license argument that by providing a digital file, you have implicitly consented to broader use.

Representation and Warranty

If you are on the commissioning side, require the artist to warrant that the work is original and does not itself incorporate AI-generated content — or at minimum, disclose any AI tools used in the creative process. The Copyright Office Part 2 report makes clear that the level of human authorship matters for copyrightability, and this affects the value and enforceability of the commissioned work copyright.

For broader context on how creators can protect their voice and likeness from AI cloning — which raises analogous legal issues — see our guide on AI Voice Clones and the NO FAKES Act.

Actionable Next Steps

If you are a visual artist concerned about AI training on your work, here is what we recommend you do — in order of urgency:

1. Register your copyrights. U.S. copyright registration is a prerequisite for filing an infringement lawsuit and unlocks statutory damages ($750 to $30,000 per work, or up to $150,000 for willful infringement). If your work has been scraped, having registrations in place strengthens your position immeasurably. Registration is inexpensive ($65 per work through the Copyright Office electronic system) and can be done at copyright.gov/registration.
2. Apply Glaze to images before posting online. This is your first line of technical defense against style mimicry. Download Glaze 2.2 from the University of Chicago SAND Lab, process your images, and post the Glazed versions.
3. Consider Nightshade for additional deterrence. Nightshade is optional and should be used thoughtfully, but applying it to images you post publicly increases the cost of unauthorized scraping for AI developers.
4. Update your commission and licensing agreements. Add the AI training prohibition, output restriction, and training-rights retention clauses described above. If you work with clients who use template agreements, propose an addendum.
5. Document everything. Keep records of where and when your work was posted online, whether you applied technical protections, any instances where you believe your style or specific works have been replicated by AI outputs, and any communications with platforms or AI companies. This documentation will be essential if you decide to join a class action or file an individual claim.
6. Monitor the litigation. The outcomes of Andersen v. Stability AI and Getty v. Stability AI will set precedent that affects every artist. Key rulings are expected in 2026. If you believe your work has been used without consent, consider consulting with an attorney about your options — including joining an existing class action or filing an individual claim.
7. Watch for the Copyright Office Part 3 report. The forthcoming section of the Office AI report will address training-data liability directly. It could recommend legislative changes, new licensing frameworks, or clarified enforcement mechanisms. Artists should be prepared to comment and advocate.

The legal landscape is evolving rapidly. What is clear today is that your copyright in your individual works is valid and enforceable, that you have tools to make unauthorized training more difficult and more costly, and that contract provisions can create additional protections that work alongside — not instead of — your statutory rights. The courts will ultimately determine whether AI training on copyrighted images constitutes fair use, but you do not have to wait for that ruling to start protecting your work.