What Changed in AI Copyright Law and Why Creators Should Pay Attention

Fact-checked by the YoureNewsSource editorial team

A federal judge just approved a $1.5 billion settlement in the Bartz v. Anthropic class action, one of the largest copyright payouts in U.S. history, and it all turned on a single, deceptively simple question: where did the training data come from? The ruling is the clearest signal yet that AI copyright law is no longer a theoretical debate confined to law reviews. It’s a fast-moving, cash-on-the-table reality that directly affects anyone who creates, publishes, or licenses expressive work.

More than 70 copyright infringement lawsuits are now active against AI companies, more than double the roughly 30 cases pending at the end of 2024. The plaintiffs range from individual photographers and authors to major record labels and stock photo agencies. The money involved isn’t hypothetical. The Anthropic settlement alone works out to roughly $3,000 per book for the 482,460 titles the company downloaded from pirate libraries, a figure that establishes a tangible, per-work benchmark for damages in future cases.

If you create visual art, write books, compose music, or produce any kind of content that could end up in a training dataset, the legal ground has shifted underneath you. By the time you finish this article, you’ll understand exactly what the U.S. Copyright Office now requires for registration, which uses of AI output are legally defensible, and how to make practical decisions about protecting your work without waiting for Congress to act.

What Just Happened in AI Copyright Law

The U.S. Copyright Office dropped a series of reports between 2023 and May 2025 that amount to the most consequential AI copyright law guidance in decades. The Part 2 report, released in January 2025, confirmed that copyright protection still requires human authorship, no exceptions. The Part 3 report, published in pre-publication form in May 2025, tackled fair use in AI training head-on. Together, they form the foundation that courts and creators are now operating on.

Meanwhile, the lawsuits piled up. The Copyright Alliance tracked over 70 active cases by early 2026, but the inflection point came in mid-2025 when the Bartz v. Anthropic ruling handed creators a concrete win. The court held that training on lawfully purchased books could qualify as fair use, but downloading nearly half a million titles from pirate libraries crossed a clear legal line. That distinction matters enormously because it gives creators a practical benchmark: if a company can show it paid for the data, the fair-use argument gets stronger; if it can’t, the damages clock starts ticking.

Not all the news cut in creators’ favor. The D.C. Circuit’s March 2025 decision in Thaler v. Perlmutter, with a rehearing denied in May 2025, unanimously reinforced the human-authorship requirement. The court pointed to statutory language referencing an author’s death, widow, or signature as proof that Congress never intended to protect machine-generated work. For anyone hoping that U.S. courts might recognize AI as an author, that door is now firmly shut.

Why the Timing Matters for Individual Creators

Most of the early lawsuits were filed by large publishers and record labels with deep pockets. But the per-work payout established in the Anthropic case creates a template that individual authors and artists can use in future claims. If you can show your work was scraped from a pirate source and used for training, you now have a dollar figure to anchor your demand. That’s a big shift from the uncertainty of two years ago, when damages for AI training were anyone’s guess.

Human Authorship Is Still the Gatekeeper

The Copyright Office’s Part 2 report could not be clearer: “copyright protection requires human authorship.” If you type a prompt into Midjourney and hit enter, the output, no matter how impressive, is not copyrightable on its own. The Office has consistently rejected applications where the human contribution stopped at the prompt. Here’s the nuance: when a human selects, arranges, or modifies AI-generated material in a way that reflects creative choices, those human-authored elements can be protected.

As of early 2024, the Copyright Office had registered well over 100 AI-assisted works, according to RAND Corporation data. That number has almost certainly climbed since. The key differentiator in every successful registration wasn’t the sophistication of the AI tool, it was the level of human control over the expressive output. A graphic novel that used AI-generated backgrounds but featured hand-drawn characters and a human-written script got through. A raw AI image with no post-generation editing did not.

What the Rejections Teach Us

The Office’s denials are just as instructive as the registrations. Applications that listed an AI system as the author were rejected outright under the Thaler precedent. Works where the human contribution was limited to curating AI outputs, picking the best image from a batch, also failed. The Office wants evidence that a human shaped the expressive content, not just that a human set the parameters and approved the result.

The Fair-Use Ruling That Changed Everything

In June 2025, the Bartz v. Anthropic court drew a line that every creator should understand: training on books you bought is fair use; training on books you stole is not. The ruling didn’t declare all AI training fair use, it made the acquisition source the decisive factor. That’s a narrower, more creator-friendly outcome than many expected.

For individual creators, the immediate takeaway is practical. If you discover your work in a dataset compiled from pirate sources, you have a stronger legal position than if it was scraped from publicly available web pages. The ruling also opens the door for future litigation targeting companies that can’t prove lawful acquisition.

Registering AI-Assisted Work: What You Need to Know

The registration process hasn’t changed at a structural level, but the expectations have sharpened. You’ll need to identify which parts of the work were AI-generated and which were human-authored. The Copyright Office now expects a disclaimer that specifically carves out the non-human elements. If you used AI to generate a background but painted the foreground yourself, you claim copyright only in the painted portions. It’s that granular.

This is where a lot of creators get tripped up. They assume that because they “made” the image, by writing prompts and iterating, they own the whole thing. Legally, they don’t. The official policy guidance is clear: the prompt alone is an idea, not expression. Copyright protects expression, not ideas. Until you add your own expressive layer, you’re standing on thin ice.

The $1.5 Billion Lesson for Registrants

The Anthropic settlement didn’t directly involve registration, but it established a market value for copyrighted works used in training. If you do register your work and later discover it was used without permission, your statutory damages claim is stronger. Registration within three months of publication, or before the infringement occurs, unlocks statutory damages and attorney’s fees under U.S. copyright doctrine. That’s a leverage point most creators overlook.

Licensing and Transparency: The Missing Pieces

Right now, no federal law requires AI companies to disclose what data they used for training. Several bills have been introduced in Congress, but none have passed. This leaves creators in a maddening position: you can’t prove your work was used unless the company admits it or a third-party audit uncovers it. Voluntary licensing markets are emerging, but they’re uneven and largely benefit creators with large catalogs or collective bargaining power.

The transparency gap is especially painful for independent creators. Without mandatory disclosure, you’re left relying on tools like Spawning’s “Have I Been Trained?” to check whether your images appear in popular datasets. These tools are useful but incomplete, they can only search datasets that are publicly accessible, which excludes many proprietary training corpora.

Collective management organizations are starting to fill the void. Groups like the Authors Guild are exploring blanket licensing models similar to those used in the music industry. These efforts are still in early stages, and the per-creator payouts, if they materialize, remain unknown.

Protecting Your Work in Practice

Practical protection right now comes down to three things: registration, documentation, and strategic enforcement. You register your most valuable works with the Copyright Office, ideally before any AI company gets its hands on them. You document your creative process so you can prove what’s yours. Then you decide whether to join a class action, negotiate an individual license, or use watermarking and opt-out tools as a first line of defense.

Watermarking has improved, but it’s not foolproof. Adversarial techniques like Glaze and Nightshade can make images harder for AI models to learn from, though neither is impenetrable. Consider them a deterrent, not a guarantee. The same goes for opt-out signals: they’re a statement of intent, but without legal backing, they’re a request.

Class Actions vs. Going Solo

For most individual creators, joining a class action is the only realistic path to compensation. Litigating against a well-funded AI company can easily run into six figures. The Bartz settlement showed that class actions can yield real money, $3,000 per book isn’t life-changing for a single title, but for authors with dozens of pirated books, it adds up. The tradeoff is that you surrender control over strategy and settlement terms. It’s messy, but it’s better than nothing.

Small Creators vs. Big AI: Is the Playing Field Level?

Let’s be blunt: the fair-use arguments that worked for Anthropic and Meta in 2025 are built on resources most individual creators will never have. These companies can afford to buy massive datasets, hire top-tier legal teams, and litigate for years. A solo photographer or indie author trying to assert fair use against a scraping claim faces a vastly different calculus. The doctrine isn’t the problem, access to it is.

The Copyright Office’s Part 3 report acknowledged this asymmetry indirectly. It noted that outcomes depend on case-specific factors, including the nature of the copyrighted work and the effect on its market. For a small creator whose work is niche, proving market harm is harder than for a major publisher whose sales data is extensive. That’s a real disadvantage, and no legislative fix is on the immediate horizon.

What Smaller Creators Can Realistically Do

Your best move is to band together. Collective action, whether through a professional association like the Graphic Artists Guild, a class action, or a licensing collective, amplifies your leverage. Solo litigation is expensive and uncertain; group action distributes the cost and pools the evidence. It’s not a perfect solution, but it’s the most practical one available right now. You might also explore tools that track where your work appears online, though the digital rights space is still maturing.

The EU Factor: How Overseas Rules Hit U.S. Creators

The EU AI Act, which entered into force in 2024, includes training-data transparency requirements that are far more stringent than anything in U.S. law. Companies deploying AI systems in the EU must disclose summaries of the copyrighted data used for training. This doesn’t directly apply to purely domestic U.S. training, but here’s the catch: if you sell your creative work into European markets, your work falls under that disclosure umbrella. Suddenly, you have a right to know that you don’t get under U.S. law.

This creates a strange two-tier system. U.S. creators who distribute internationally may have more transparency into AI training than those who stay domestic. European creators also have a structural advantage in negotiating licenses because they can point to a legal requirement that U.S. counterparts lack. For American creators eyeing global audiences, it’s worth understanding how AI productivity tools are adapting to these cross-border rules.

Compliance Pressures That Reach Across the Atlantic

Large AI companies can’t easily segment their training data by geography. If they want to operate in the EU, they need to clean up their data practices globally, or at least document where the data came from. This has a spillover effect: the transparency that European law demands may end up benefiting U.S. creators even without domestic legislation, simply because multinational companies find it cheaper to apply one standard everywhere.

What Comes Next: Legislation and Court Battles

Don’t hold your breath for a sweeping federal AI copyright law. Multiple bills have been introduced, covering disclosure, digital replicas, and licensing, but none have passed. The more likely near-term developments are at the state level, where laws targeting unauthorized digital replicas are gaining traction. These don’t directly address training, but they do create liability for AI-generated outputs that mimic a specific person’s likeness or voice.

Court battles will continue to shape the law faster than legislation does. Appeals in the Bartz and Thaler cases could refine fair-use boundaries further. More class actions are likely, especially now that the per-work damages model has been tested. For creators, the smart play is to stay informed, register your work, and be ready to act when an opportunity for compensation arises, because the law is moving, but not fast enough to protect you on its own.

Your Action Plan

1. Register your most valuable works now

   File copyright registrations for books, images, music, or other creative works that have commercial value. Do it within three months of publication to preserve statutory damages. If you’ve already published, register before any known infringement occurs. This is the single most powerful legal tool you have.

2. Document your creative process in detail

   Keep records of drafts, edits, and decisions. If you use AI tools, save screenshots of your prompts and, more importantly, your post-generation edits. The Copyright Office cares about what you did after the AI output, not what you typed into the box.

3. Check whether your work is in known training datasets

   Use tools like “Have I Been Trained?” or Spawning’s search to see if your images or text appear in publicly accessible datasets. If you find your work, document it. This evidence strengthens any future claim.

4. Join a creator organization or collective

   Groups like the Authors Guild, the Graphic Artists Guild, or discipline-specific associations are actively negotiating licensing deals and organizing class actions. Membership gives you access to legal resources you can’t afford on your own.

5. Apply watermarking and opt-out signals

   Even though they’re not legally binding in the U.S., they signal your intent and may deter casual scraping. Use platform-specific opt-outs where available, and consider adversarial tools like Glaze for visual work.

6. Monitor EU market opportunities

   If you distribute your work in Europe, familiarize yourself with the EU AI Act’s disclosure requirements. You may be able to request information about training data usage that U.S. law doesn’t require companies to provide domestically.

7. Evaluate class action opportunities carefully

   When a new lawsuit is announced, check whether your work might be included. Joining early can be advantageous, but read the terms. Understand what rights you’re giving up and what the expected per-work payout might look like based on recent settlements.

8. Stay current on court rulings and Copyright Office guidance

   This area of law is evolving monthly. Bookmark the Copyright Office’s AI page and the Copyright Alliance’s litigation tracker. The legal standard that applies today may be refined, or upended, by the next appellate ruling.

Frequently Asked Questions

Can I copyright an image I created with Midjourney?

Not by itself. If you only wrote a prompt and selected an output, the Copyright Office considers that insufficient human authorship. However, if you substantially edit the image afterward, changing composition, adding elements, or incorporating it into a larger work, those human-authored contributions may be copyrightable.

Does the $1.5 billion Anthropic settlement mean all AI training on books is illegal?

No. The court specifically distinguished between lawfully purchased books (which could be fair use) and pirated books (which are not). The settlement covers only the pirated titles. If an AI company paid for the books, the fair-use analysis is different and remains unresolved in many circuits.

What is the Copyright Office’s position on AI-generated music?

The same as for visual art: human authorship is required. If an AI generates a melody and you layer vocals and instrumentation on top, you can claim copyright in the human-added elements. The raw AI-generated melody would not be protected.

Are there any laws requiring AI companies to disclose training data?

In the U.S., no. No federal law mandates training data disclosure, though several bills have been proposed. The EU does have such requirements under the AI Act, which can indirectly benefit U.S. creators who distribute their work in Europe.

How many AI-assisted works has the Copyright Office actually registered?

Well over 100, according to the RAND Corporation citing Copyright Office data. That number has grown since, but the Office has not released updated aggregate figures.

Can I sue an AI company if my work was used without permission?

Possibly, but individual litigation is expensive. Joining a class action is more practical for most creators. If you registered your work before the infringement, you have stronger remedies including statutory damages and attorney’s fees.

Does opting out of AI training via robots.txt protect me legally?

Not in the U.S. right now. Opt-out signals are a request, not a legal barrier. They may discourage some companies, but they don’t create a cause of action if ignored. The legal status could change if Congress passes a disclosure or opt-out law.

What’s the difference between the U.S. and EU approaches to AI copyright?

The U.S. relies on case-by-case fair-use analysis with no mandatory transparency. The EU requires companies to disclose summaries of copyrighted training data. U.S. creators selling into Europe can benefit from that transparency even though domestic rules don’t require it.

Is my prompt alone enough to claim copyright in an AI output?

No. The Copyright Office treats prompts as ideas, not expression. Copyright protects the expression, which requires human creative choices in the final work, not just the instructions that generated it.

Sources