AI Art Training Lawsuits and Artist Protections: What Visual Artists Actually Own and How to Opt Out

If you are an illustrator, painter, or digital artist, your work is almost certainly in an AI training dataset already. Billions of images scraped from the internet — from DeviantArt portfolios to personal websites to ArtStation galleries — have been ingested into generative AI models like Stable Diffusion, Midjourney, and DALL-E without the consent of the artists who created them. You were not asked. You were not credited. You were not paid. And the legal system is still catching up to what that means for your rights.

The good news is that the legal landscape is shifting. The first major class-action lawsuit by visual artists against AI companies — Andersen v. Stability AI — is actively litigated in the Northern District of California, and key claims have survived motions to dismiss. The U.S. Copyright Office released the third part of its AI report in May 2025, addressing the use of copyrighted materials in AI training. Platforms are rolling out opt-out mechanisms. And technical tools like Glaze and Nightshade give artists a way to fight back at the pixel level.

This guide covers what you can actually do — as a working artist, today — to protect your work from AI training. We walk through the copyright protections that exist, what the Andersen litigation means for you, how copyright registration strengthens your legal position, which platform opt-outs work (and which are performative), what Glaze and Nightshade do, and how the Visual Artists Rights Act intersects with the AI scraping era. This is not a policy essay. It is a practical legal guide for visual artists who want to understand their rights and take action.

What Copyright Protection Exists for Scraped Art

Under U.S. copyright law, you hold copyright in your original visual works from the moment they are fixed in a tangible medium — which includes saving a digital file. You do not need to register your copyright to own it. The exclusive rights granted under 17 U.S.C. § 106 include the right to reproduce the work, prepare derivative works, distribute copies, and publicly display the work. When an AI company scrapes your image from the internet and uses it to train a model, that act implicates at least two of these exclusive rights: reproduction (copying the image into a training dataset) and derivative works (the model's ability to generate outputs influenced by your artistic style).

But owning a copyright and enforcing it against an AI company are very different things. AI companies have primarily raised two defenses: first, that training constitutes fair use under 17 U.S.C. § 107, and second, that what their models store are not copies of your work but statistical representations — mathematical weights and parameters that do not contain protectable expression. The fair use defense is untested in the context of generative AI training at trial, and the "statistical representation" argument has been contested in ongoing litigation. The U.S. Copyright Office addressed these questions directly in its AI Initiative Report, Part 3, released in pre-publication form on May 9, 2025, which analyzes the copyright issues raised by training generative AI models on copyrighted works.

The practical reality for artists: your copyright exists, but the legal framework for enforcing it against AI training is still being built through litigation and regulatory guidance. That is why proactive steps — registration, opt-outs, and technical protection — matter more now than ever.

Andersen v. Stability AI: What the Case Means for Artists

In January 2023, visual artists Sarah Andersen, Kelly McKernan, and Karla Ortiz filed a class-action lawsuit against Stability AI, Midjourney, and DeviantArt in the Northern District of California, alleging that the companies scraped their copyrighted works into training datasets for AI image generators without consent. It was the first lawsuit of its kind — creators banding together to challenge the unauthorized ingestion of their works by a generative AI company. According to the Copyright Alliance's analysis, over a dozen similar class action lawsuits have followed, brought by authors, visual artists, and other creators against various AI companies.

The case has produced several significant rulings. In October 2023, Judge William Orrick largely granted the defendants' motions to dismiss but allowed the direct infringement claims to move forward and gave plaintiffs leave to amend. Then, on August 12, 2024, the court issued an order on the first amended complaint that revealed how courts may approach generative AI copyright claims going forward. Three takeaways matter for artists:

First, the "VCR analogy" does not work for AI companies. Judge Orrick distinguished generative AI from past technologies like the VCR, which were not pre-loaded with copyrighted works. He found that "Stable Diffusion is built to a significant extent on copyrighted works and that the way the product operates necessarily invokes copies or protected elements of those works." This is a critical distinction: unlike a VCR (a tool capable of copying but with substantial non-infringing uses), generative AI models are built by ingesting copyrighted works before they ever reach the consumer.

Second, DMCA Section 1202 claims for removal of copyright management information are struggling. The court dismissed both 1202(a) and 1202(b) claims with prejudice, adopting an "identicality" requirement — meaning artists must show that AI outputs are identical to their works to prove CMI removal. This is a high bar that most artists cannot meet, since generative AI outputs are rarely exact reproductions.

Third, the argument that AI models only store "unprotectable data" is not holding up. Judge Orrick found that works may be "contained in Stable Diffusion as algorithmic or mathematical representations" and that this being "fixed in a different medium than they may have originally been produced in — is not an impediment to the claim." If courts ultimately accept that statistical representations of copyrighted works in a model are simply a different medium in which expressive elements are fixed, it would undermine one of the AI industry's core defenses.

The case is still in active litigation. Class certification issues and discovery are proceeding. For artists whose work has been scraped, the Andersen case represents the leading edge of a legal frontier that could determine whether unauthorized scraping constitutes infringement — but it has not yet produced a final judgment, and no settlement has been reached.

How Copyright Registration Strengthens Your Position

Copyright registration is the single most important step you can take to protect your visual work — and it matters even more in the AI era. While copyright exists automatically upon creation, 17 U.S.C. § 411 requires that you register your work (or at least have an application pending) before you can file an infringement lawsuit. More critically, 17 U.S.C. § 412 provides that if you register your work before infringement begins — or within three months of first publication — you are eligible to seek statutory damages (up to $150,000 per work for willful infringement) and attorneys' fees. If you register after infringement begins, you are generally limited to actual damages and infringer's profits, which are far harder to prove and often amount to less.

In the AI training context, this matters enormously. If your work was scraped into a training dataset before you registered, and you later sue for infringement, your ability to recover statutory damages may depend on when you registered relative to when the scraping occurred. Artists who register their work promptly after publication — within the three-month grace period — preserve their right to seek the full range of remedies. Those who never register may find that even a strong infringement claim yields only nominal damages.

Registration also creates a prima facie presumption of validity if made within five years of publication, which shifts the burden to the alleged infringer to prove the copyright is invalid. In litigation against well-funded AI companies, this presumption is a meaningful advantage. You can register visual works through the U.S. Copyright Office's online registration system, and the visual arts category — which covers paintings, drawings, photographs, sculptures, and digital art — has a straightforward application process. The registration fee is modest (typically $45-$65 per application as of 2025), and you can register multiple works under certain circumstances.

For artists concerned about AI scraping specifically, we recommend registering each significant body of work within three months of publication. If you post regularly on platforms like Instagram, DeviantArt, or ArtStation, batch registration of your recent works at regular intervals is a practical approach. As we discuss in our coverage of AI training data exposure and IP risk, the line between protected expression and statistical representation is being litigated right now — and registered works give you the strongest possible position.

Platform-Specific Opt-Outs: What Works and What Is Performative

DeviantArt: The NoAI Toggle

DeviantArt — one of the defendants in the Andersen litigation — implemented a "NoAI" setting that lets artists flag their work as not authorized for inclusion in third-party AI training datasets. The setting can be applied individually to each deviation, in bulk from the Studio or gallery, or across an entire gallery at once. According to DeviantArt's support documentation, the NoAI label "unambiguously communicates that artwork with this flag is not authorized for inclusion in third-party datasets for training AI models."

DeviantArt also announced that all deviations are opted out of AI datasets by default — a significant reversal from its initial position. Artists should still verify that the NoAI label is active on their work, especially for older uploads. The practical limitation: the NoAI label is a contractual signal, not a technical barrier. It communicates your intent not to be scraped, but it does not prevent a determined scraper from ignoring the tag. Its legal value is in establishing that your work was not voluntarily contributed to a training dataset — which matters for infringement claims.

ArtStation: The NoAI Meta Tag

ArtStation, owned by Epic Games, released a similar mechanism. According to ArtStation's official guidance, projects tagged with "NoAI" are automatically assigned an HTML "NoAI" meta tag that "will explicitly disallow the use of the content by AI systems and mark the project so that they know they are not allowed to use it." ArtStation also allows users to filter out AI-generated artwork from homepage galleries, search, and marketplaces.

Notably, ArtStation does not enable NoAI by default. The platform's guidance states: "If you do not use the tag, you are neither allowing nor disallowing the use of your project for AI. Choosing not to use the tag leaves copyright law to govern whether or not the artwork was fairly used." This is a weaker default than DeviantArt's, and artists on ArtStation should manually apply the NoAI tag to all their work.

Adobe Firefly: The Licensed-Only Model

Adobe Firefly takes a fundamentally different approach. According to Adobe's official documentation, Firefly is "trained on content where we have permission or rights" — specifically, Adobe Stock content, openly licensed works, and public domain content where copyright has expired. Adobe does not train on user data and offers IP indemnification to qualifying enterprise customers for content generated by Firefly. For artists, this means that Adobe's generative AI tools are not built on scraped artwork — making Firefly the most artist-friendly generative AI platform currently available. If you are looking for AI tools that do not exploit your work or your peers' work, Firefly's licensed-only training model is the current industry standard.

Glaze and Nightshade: Technical Protection Tools

While opt-out tags are contractual signals, Glaze and Nightshade — developed by the SAND Lab at the University of Chicago — are technical tools that alter your images at the pixel level to interfere with AI model training. Glaze applies imperceptible perturbations to your artwork that confuse AI style-learning models, making it harder for them to replicate your artistic style. Nightshade goes further: it is a "data poisoning" tool that causes AI models that ingest Nightshade-treated images to produce distorted or nonsensical outputs. Both tools are free and available for download.

The legal status of these tools is straightforward: using Glaze or Nightshade on your own artwork is not illegal. You are modifying your own intellectual property. There is no law that requires you to make your work machine-readable or training-friendly. However, these tools are not a substitute for legal protection. They are a technical countermeasure — effective against current generation models but potentially circumventable as AI training techniques evolve. Think of them as a seatbelt, not a force field. Use them alongside copyright registration and platform opt-outs, not instead of them.

VARA and Moral Rights in the AI Era

The Visual Artists Rights Act (VARA), codified at 17 U.S.C. § 106A, grants authors of visual arts works certain "moral rights" that are independent of the economic rights in § 106. These include the right to claim authorship of the work, the right to prevent the use of the artist's name as author of a work they did not create, the right to prevent the use of the artist's name in connection with a distorted or modified work that would be prejudicial to the artist's honor or reputation, and the right to prevent intentional distortion, mutilation, or modification of the work that would be prejudicial to the artist's reputation.

VARA's application to the AI scraping context is limited but worth understanding. The rights under § 106A apply specifically to "works of visual art" — a defined term that covers paintings, drawings, prints, sculptures, and photographs produced for exhibition, but excludes works made for hire, promotional material, and many digital works. The rights also do not apply to reproductions or depictions of the work in other media. This means that when an AI model ingests a digital image of your painting, VARA likely does not extend to the digital copy used in training — the statute's exceptions for reproductions limit its reach.

However, VARA's attribution right could be relevant if an AI system generates output that is explicitly attributed to you without your consent, or if your name is used in connection with a distorted or modified version of your work in a way that harms your reputation. These are narrow applications, and VARA has not yet been tested in the AI training context. The rights under VARA last for the life of the author and cannot be transferred — though they can be waived in a written instrument signed by the artist. As we explore in our analysis of AI governance frameworks and their intersection with IP rights, the gap between existing IP law and AI-era realities is one that legislatures and courts are still working to close.

Actionable Next Steps

1. Register your copyright within three months of publication. This is the most important step you can take. Registration within the three-month grace period preserves your right to statutory damages and attorneys' fees. The U.S. Copyright Office's online registration system makes this straightforward for visual works. Register in batches if you produce work frequently.
2. Enable NoAI tags on every platform that offers them. On DeviantArt, verify that the NoAI label is active on all your deviations — it is enabled by default, but check older uploads. On ArtStation, manually apply the NoAI tag to all your projects, since it is not enabled by default. These tags create a record of your objection to AI training use.
3. Apply Glaze to your portfolio images before posting. Download Glaze from the University of Chicago SAND Lab website and apply it to your work before uploading. If you are particularly concerned about style mimicry, also apply Nightshade to images that are most likely to be targeted by AI scrapers.
4. Choose AI tools that respect artist rights. If you use generative AI tools in your own practice, choose platforms like Adobe Firefly that train only on licensed content. Avoid tools built on scraped datasets — using them perpetuates the system that exploits your peers.
5. Document your work's publication history. Keep records of when and where each work was first published, when you registered it, and when you enabled NoAI tags. If your work appears in an AI output and you need to prove infringement, this documentation is your evidence.
6. Watch the Andersen case and Copyright Office developments. The Andersen v. Stability AI litigation and the Copyright Office's Part 3 report are the two most significant developments for visual artists' rights in the AI era. The legal standards that emerge from these proceedings will shape your enforcement options.
7. Talk to an attorney if your work has been scraped. If you discover your work in an AI training dataset or recognize your style in AI-generated outputs, consult with an attorney who understands copyright law and AI training issues. The Andersen case may eventually produce a settlement or judgment that affects artists whose work was scraped, and you need to understand your options before deadlines expire.

The legal infrastructure protecting visual artists from unauthorized AI training is still being built — through litigation, regulatory guidance, and platform policy. But you are not powerless while it is under construction. Copyright registration, platform opt-outs, and technical tools like Glaze and Nightshade give you layers of protection that, used together, significantly strengthen your legal position and reduce the practical value of your work to AI scrapers. The artists who take these steps now will be in the strongest position when the legal landscape finally settles — and the ones who do not will have given away their leverage before the rules are even written.