If 2023 was the year generative artificial intelligence made its splashy mainstream debut, then 2025 is shaping up to be the year the courts brought it back down to Earth with a gavel.

In a one-two punch delivered on June 23 and 25, two federal judges from the U.S. District Court for the Northern District of California weighed in on a pressing, polarizing issue: Is it fair use for generative AI models to train on copyrighted books without permission? Both said yes, but with caveats.

Welcome to Bartz v. Anthropic PBC and Kadrey v. Meta, the first U.S. rulings addressing this issue head-on. The decisions weren't exactly soulmates, but they sure made headlines.

Bartz v. Anthropic: "Spectacularly Transformative," but Hold the Piracy

On June 23, in Bartz, U.S. District Judge William Alsup ruled that Anthropic's use of legitimately purchased books to train its Claude model was fair use. He was apparently swayed by Anthropic's arguments over how "exceedingly" transformative it was, with his opinion likening large language model training to teaching kids to write.

But then came the record scratch. Anthropic had also scarfed up millions of pirated books from sketchy online repositories to build a "central research library." That use, Judge Alsup ruled, was not fair use and would proceed to trial.

The net takeaway is that transformative use wins the day, but only if your platform is trained on lawfully acquired materials.

Kadrey v. Meta: Fair Use Redux, with a Market Harm Twist

Just two days later, U.S. District Judge Vince Chhabria blessed Meta's Llama model with a similar fair use halo, but in a more nuanced tone. He agreed the training use was transformative. The books were written to be read and enjoyed; Llama was being trained to churn out generic outputs at scale.

But Judge Chhabria also honed in on market harm with surgical precision. He broke it down into three flavors:

• Direct substitution: Could Meta's AI spit out a facsimile of your novel? Nope. Not even 50 words. Meta built in filters to prevent this. So, no dice.
• Lost licensing revenue: Could authors have cashed in by licensing their books for AI training? Maybe. But the court found that argument circular: If the use is transformative, no license is required. If no license is required, there's no lost licensing revenue. Catch-22, copyright edition.
• Indirect substitution: Could AI-generated books saturate the market and drown out the originals? Chhabria said maybe, and even seemed sympathetic. But plaintiffs didn't raise that theory in their complaint or support it with real evidence. So, again, Meta wins (at least this round).

Let's Talk About the Split

Both judges reached the same fair use outcome, but their methods were anything but aligned.

Judge Alsup analogized AI training to teaching students to write better. Judge Chhabria clapped back: Teaching kids is not like building an algorithm that can generate a thousand romance novels before breakfast. Humans have filters. Llama just needs a prompt.

So, who won? If you're Meta or Anthropic, you're high-fiving your lawyers. If you're a rightsholder, you're probably revising your legal strategy and praying the next judge isn't nestled in Silicon Valley.

Both rulings underscore that fair use is not a get-out-of-litigation-free card for the AI platforms. It's a fact-specific defense against a finding of infringement, and if the facts change, possibly with stronger proof of market harm or more brazen content scraping, so could the outcome.

And let's not forget that Judge Alsup recently took a pivotal step by granting class certification for authors whose works were scraped from illegal sources to train Anthropic's Claude model. He approved a class of authors whose books were pirated and used without permission, allowing them to collectively pursue claims for copyright infringement.

This is particularly significant because it shifts the focus from individual lawsuits to a mass class action, potentially exposing Anthropic to statutory damages of up to $150,000 per work, a legal liability that could cascade into the hundreds of millions depending on the volume of pirated content identified.

By greenlighting class treatment, the court signaled that unauthorized mass ingestion of copyrighted works is not just a procedural anomaly, but a systemic issue warranting group accountability, a development that dramatically escalates the stakes for AI companies relying on unsanctioned content.

Enter: Disney Enterprises v. Midjourney

Just when you thought the courtroom drama couldn't get more cinematic, along comes Disney Enterprises v. Midjourney, a June lawsuit in the U.S. District Court for the Central District of California that promises to bring the full force of Hollywood intellectual property muscle into the AI copyright cage match.

Unlike Bartz and Kadrey, which centered on textual works, this case dives headfirst into the world of visual art.

The media giants plaintiffs, including Marvel Characters Inc., Lucasfilm Ltd. LLC and Twentieth Century Fox Film Corp., allege that Midjourney, the image-generating darling of the generative AI scene, trained its models on copyrighted film stills, concept art and other proprietary visuals, without permission. And they're not amused that the platform can now spit out "new" artworks suspiciously reminiscent of Elsa, Darth Vader or Jurassic Park.

Why does this matter in the post-Bartz/Kadrey world? Because it tests the limits of how courts apply the fair use framework to a different kind of creative content. Textual works and images aren't judged by the same metrics when it comes to market harm or transformative use. Courts have traditionally given less leeway when the outputs are stylistically or thematically similar to highly branded intellectual property, especially in visual media.

And let's be real: Judge Chhabria's take that Llama doesn't output more than 50 words from a novel isn't going to fly if Midjourney coughs up a picture that looks like Spider-Man in everything but name. If Bartz and Kadrey were about whether AI training can be transformative, the Disney/Midjourney brawl is about whether an AI-generated image looks too much like the original to get a pass.

This case could also test the limits of indirect market substitution, the very theory that Judge Chhabria flagged but didn't rule on. If Midjourney enables users to create near-identical versions of Pixar-style art for free when such images could be sold by Disney or Universal, that could be a slam-dunk example of crowding out an established market with AI-created knockoffs.

Translation: We're entering the "show me the output" phase of AI copyright law. Bartz and Kadrey dealt mostly with inputs. Disney v. Midjourney is all about what comes out the other side, and whether it's just "inspired" by Disney, or practically interchangeable with it.

And Now: Congress Enters the Chat

Hot off the presses, Sens. Josh Hawley, R-Mo. and Richard Blumenthal, D-Conn., are putting the AI industry on notice with the AI Accountability and Personal Data Protection Act, a newly introduced bipartisan bill that would ban AI companies from training their models on copyrighted material or personal data without clear, affirmative consent.

The legislation gives both individuals and companies the right to sue if their creative work or personal information is used without permission, finally offering artists, writers and ordinary citizens a way to push back against the data-grabbing free-for-all. It also targets shady consent practices, outlawing data collection when consent is coerced, deceptive or wildly disproportionate to the service offered.

Among its provisions are greater transparency around training data sources, mandatory disclosures when AI outputs are substantially similar to copyrighted works, and a new right for authors and artists to opt out of future model training.

If passed, this could upend the current patchwork of judge-by-judge decisions and create a national standard, one that may make life harder for AI companies trying to play fast and loose with intellectual property. But for now, the act is still in committee.

What Now?

Expect more lawsuits. Expect appeals. And expect courts to scrutinize not just what was copied, but how, why and with what effect. Also, note that these rulings only apply in California for now. Judges elsewhere may not be as AI-friendly as those in the Northern District of California.

And hey, just because something is legally permissible doesn't mean it won't blow up in your face in the court of public opinion. If your AI model has a diet of stolen books or princesses, expect blowback, no matter how transformative your defense sounds in court.

The law is still in flux? You bet.

Grab your popcorn. The fair use fight is just getting started — and the sequel might be even juicier.