You don’t have to be OpenAI, Anthropic, or Midjourney to have a stake in these lawsuits. If your marketing team runs copy through ChatGPT, your design team generates images with Midjourney, or your product embeds Claude or Gemini under the hood, the outcome of cases you’ve probably never heard of is quietly deciding how much legal exposure you’re carrying right now. A settlement in San Francisco can change what your vendor’s indemnification actually covers. A ruling in London can determine whether “the model was trained overseas” is a real legal shield or just a talking point. A Supreme Court decision about an internet provider from 2018 can rewrite the risk calculus for every AI company you rely on, including the one whose logo is on your dashboard.
This isn’t a list of court cases for lawyers to argue about at conferences. It’s the part of running a business in 2026 that most executives have quietly outsourced to “someone in legal will handle it,” and in a lot of companies, nobody actually has.
The Quick Answer
Courts in 2026 are converging on a specific line: training an AI model on copyrighted material you or your vendor acquired legally has a real shot at being fair use, but training on material pulled from pirate sites is a different animal entirely, and it is losing in court and in settlements. Anthropic’s $1.5 billion settlement with authors, finalized this past July, is now the reference point every plaintiff’s lawyer uses to value these claims. Meanwhile, a March 2026 Supreme Court ruling made it structurally harder to sue an AI platform just because its tool could be misused, which is good news for AI vendors and, by extension, their business customers. The part that should worry you most isn’t the training data fight. It’s that your vendor’s copyright indemnification, the thing your legal team probably assumes covers you, has conditions attached that a live lawsuit is currently testing in court.
Key Takeaways
- The central legal question in 2026 isn’t “is AI training illegal.” It’s “how was the training data acquired.” Legally licensed or purchased data is trending toward fair use. Pirated data is trending toward liability.
- Anthropic’s $1.5 billion settlement with authors, given final court approval on July 20, 2026, set a rough benchmark of roughly $3,000 per infringed work, a number plaintiffs elsewhere are already citing.
- A March 2026 Supreme Court ruling in a case involving an internet provider, not an AI company, narrowed contributory copyright liability industry-wide, making “you knew this could be misused” a much weaker legal theory on its own.
- Vendor indemnification from Microsoft, OpenAI, Anthropic, and Google comes with real conditions, and a lawsuit filed in August 2026 alleges those conditions can be broken by something as simple as re-prompting past a safety filter.
- Purely AI-generated content still cannot be copyrighted in the US under current law, which means you may not be able to stop a competitor from copying material your business produced entirely through AI with no human editing involved.
- If your business operates in the EU, the fair use protections your AI vendor relies on in US courts do not automatically travel with you. The EU AI Act has its own training data transparency and copyright compliance rules, with enforcement underway since August 2026.
Why a Lawsuit You’ve Never Heard Of Could Still Cost You
Here’s the mental model worth adopting. Every generative AI tool your business uses sits on top of a training process that is currently being litigated somewhere. You didn’t choose that data, you can’t audit it, and in most cases you’ll never even see it. What you inherit instead is whatever legal exposure your vendor ends up carrying, filtered through whatever contract you signed with them. When a court decides that a particular kind of training data acquisition was illegal, that ruling doesn’t just affect the AI company being sued. It reshapes the indemnification terms, the pricing, and sometimes the actual product features available to every downstream business using that tool.
That’s why a business that has never generated an image of Bart Simpson still has a reason to understand the Disney v. Midjourney case, and why a company that has never touched a line of Claude-generated song lyrics should still care that Sony Music and Warner Chappell just sued Anthropic’s CEO personally. The cases below aren’t abstract. They’re the terms and conditions nobody reads, playing out in real time, in federal court.
The Case That Set the Template: Bartz v. Anthropic
If you only understand one case, make it this one, because every other AI copyright dispute in 2026 is being measured against it.
Three authors, Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson, sued Anthropic in 2024 alleging the company used their books to train Claude without permission. In June 2025, Judge William Alsup issued a ruling that split the case in two, and that split is now the framework nearly every other court is using. Training an AI model on books Anthropic had acquired legally, he found, was transformative enough to qualify as fair use. But millions of the books in Anthropic’s training library hadn’t been acquired legally. They’d been downloaded from Library Genesis and a mirror site called Pirate Library Mirror, essentially pirate book repositories, to build what the company internally treated as a permanent research collection. That part, the judge ruled, had nothing to do with fair use. It was just downloading pirated books, and courts don’t extend fair use protection to how you acquired your source material, only to what you did with it afterward.
A class was certified covering rightsholders of books pulled from those two pirate sites, and a trial to determine damages was scheduled for December 2025. Instead, the parties settled for $1.5 billion, the largest copyright class action settlement in history. After a wave of procedural steps, including a change of judge when Alsup retired and Judge Araceli Martínez-Olguín took over, the settlement received final court approval on July 20, 2026. The claims rate came in at 92.77 percent, an unusually high figure for a class action, and Anthropic is required to destroy all the original pirated files it downloaded along with any copies derived from them.
Do the math on that settlement and you land on roughly $3,000 per book, out of a theoretical maximum of $150,000 per work under statutory damages for willful infringement. That’s the number every plaintiff’s attorney filing a similar case now has in their back pocket as a credible settlement floor.
The lesson for any business isn’t really about book publishing. It’s that courts are drawing a hard line between “we trained on data we licensed or bought” and “we trained on data we pirated,” and that line is turning out to matter enormously in practice, even when the training methodology itself is identical.
The Appeal That Could Decide Everything Else
Before Bartz, there was Thomson Reuters v. Ross Intelligence, and it’s still working its way through the appeals process in a way that could reshape how every other pending AI case gets decided.
Ross Intelligence wasn’t a generative AI company. It built a legal research tool trained in part on Thomson Reuters’ Westlaw headnotes, the editorial summaries Westlaw writes to accompany case law. A federal court found the headnotes were original enough to be protected by copyright, and that training an AI tool on them wasn’t fair use. Ross lost, and the ruling became the first real fair use loss for an AI company anywhere in the country. The case is now on appeal before the Third Circuit, which heard oral argument on June 11, 2026. Whatever that panel decides will be the first appellate-level word on how fair use applies to AI training, and even though Ross wasn’t a generative model, legal analysts across the industry expect the reasoning to ripple into every pending generative AI case, Bartz’s fair use holding included.
If you’re a business waiting to see whether the ground under generative AI licensing is stable, this appeal, not any single generative AI lawsuit, is arguably the one to watch most closely over the next year.
The Image Wars: Getty v. Stability AI and Disney v. Midjourney
Text and books aren’t the only battlefield. Two major image-generation cases show how differently this can play out depending on jurisdiction and what a studio is actually alleging.
Getty Images v. Stability AI (UK)
Getty sued Stability AI in the UK in 2023, alleging Stable Diffusion was trained on millions of Getty’s copyrighted photographs. By the time the case reached trial, Getty had dropped its central claims about training and AI outputs, because it couldn’t establish that the training had actually happened inside the UK, and Stability’s training infrastructure was based in the US. What remained was a narrower “secondary infringement” theory: that the trained model itself was an illegal copy being imported into the UK. The High Court rejected that argument in a November 2025 ruling, holding that model weights don’t constitute a “copy” under UK law because they don’t store a recognizable reproduction of the original images. Getty did win a limited, historic trademark infringement claim over Getty watermarks that occasionally appeared in Stable Diffusion outputs, and the company has been granted permission to appeal the secondary infringement finding.
The takeaway that matters for business readers: where training physically happens can determine which country’s courts even have jurisdiction to rule on the core question. Getty’s parallel US case is still alive precisely because Stability concedes the training happened there, which keeps the primary infringement question open in American courts even after the UK loss.
Disney, Universal, and Warner Bros. Discovery v. Midjourney
Disney and Universal sued Midjourney in June 2025, later joined by Warner Bros. Discovery in September, and the cases were consolidated into a single action that November. The studios allege Midjourney trained on their copyrighted characters, everything from Darth Vader and Bart Simpson to Batman and Scooby-Doo, and that its image generator will produce near-identical renditions of those characters on request. They’re seeking an injunction and statutory damages of up to $150,000 per infringed character. Midjourney’s defense is a straightforward fair use argument: training on publicly available images, even copyrighted ones, doesn’t require permission.
The case is currently deep in a genuinely interesting discovery fight. A court initially ruled the studios only had to disclose their own internal AI usage as it relates to consumer-facing products, and Midjourney is now pushing to expand that, seeking the studios’ internal AI strategy documents, training data, and even employee prompts entered into Midjourney’s own platform. Expert discovery is scheduled to run through November 2026, with no trial date yet set. Unlike Bartz, this case is squarely about outputs, not just training, which makes it a genuinely different legal theory worth tracking separately.
The News Industry’s Fight, and the DOJ’s Surprise Move
The New York Times sued OpenAI and Microsoft back in December 2023, and the case has turned into one of the most consequential discovery battles in tech litigation history. Courts have ordered OpenAI to produce a staggering volume of output logs, an initial 20 million in January 2026, followed by an additional 78 million and 10 million more in March, well over 100 million logs in total. In August 2026, the Times tried to add a new legal theory accusing Microsoft of actively encouraging OpenAI’s alleged infringement, and the judge rejected the attempt as too late, noting the Times could have raised it earlier and chose not to.
Then, on September 2, 2026, the US Department of Justice filed a brief in the case supporting OpenAI and Microsoft, arguing that a thriving domestic AI industry is a genuine national security interest. It’s the first time the federal government has weighed in on the underlying copyright question in any AI training case, and regardless of how much weight a court ultimately gives it, it’s a real signal about which direction federal policy currently leans.
The Times case isn’t the only front here anymore, either. In June 2026, a coalition of roughly 400 local newspaper outlets across 33 states, led by Richner Communications, filed their own suit against Microsoft and OpenAI in the Southern District of New York. That expansion matters because it shows this fight is no longer limited to a handful of the biggest media names. It’s becoming a much broader industry-wide claim, with a lot more plaintiffs and a lot more potential exposure for the companies on the other side of the table.
The Ruling Nobody in AI Expected to Matter
Here’s a case that has nothing to do with AI on its face and quietly changed the legal odds for every AI company facing a lawsuit.
Cox Communications, an internet provider, was hit with a billion dollar jury verdict for failing to crack down on subscribers who used its network to pirate music, on the theory that Cox knew about the infringement and kept providing service anyway. On March 25, 2026, the Supreme Court unanimously reversed that verdict in Cox Communications v. Sony Music Entertainment, holding that contributory copyright liability requires actual intent to induce infringement, or a service specifically built to facilitate it, not merely a general awareness that some users might misuse an otherwise lawful product.
The implications for AI companies are direct. A plaintiff can no longer argue that an AI company is liable simply because it knew its tool could be used to reproduce copyrighted material. They now have to show the company actively encouraged that outcome, or built the product specifically to enable it. That’s a meaningfully higher bar, and it’s already showing up in how courts are handling AI cases: it’s part of why the New York Times’ attempt to add an inducement theory against Microsoft got so much scrutiny over timing. For AI vendors, this ruling is a genuine tailwind. For businesses relying on those vendors’ legal stability, it’s one less doomsday scenario to worry about, though it does nothing to protect you if your own use of a tool crosses into active infringement.
The Music Industry’s Escalating War With Anthropic
If you want to see how fast this kind of litigation can multiply, watch what’s happened to Anthropic over the last three years.
Universal Music Publishing Group, Concord Music Group, and ABKCO filed the first suit in October 2023, covering roughly 500 songs. Around January 2025, the parties reached an early stipulation under which Anthropic agreed to build guardrails preventing Claude from reproducing specific copyrighted lyrics. That should have been the end of it. Instead, the same publishers came back on January 28, 2026, with a second, far bigger suit covering more than 20,000 works and seeking over $3 billion, alleging Anthropic’s entire training pipeline was built on pirated material. BMG Rights Management filed its own case in March 2026 over roughly 493 compositions. Round Hill Music filed a fourth case in mid-August.
Then, on August 28, 2026, Sony Music Publishing and Warner Chappell Music filed a joint suit, on behalf of a coalition of 35 music publishing companies in total, that completed the set: Universal, Sony, and Warner, the three major music publishers, are now all suing Anthropic simultaneously. The complaint names Anthropic’s CEO Dario Amodei and co-founder Benjamin Mann personally as defendants, and describes the alleged conduct in blunt terms, calling it one of the largest and most blatant intellectual property thefts in history. It alleges Anthropic torrented millions of books from the same pirate sites at issue in Bartz, then separately scraped song lyrics directly from licensed lyric platforms such as MusixMatch and LyricFind.
The detail that should get every business’s attention isn’t the dollar figure, it’s the guardrail allegation. The publishers claim the lyric-blocking measures Anthropic put in place after the earlier stipulation can be defeated just by re-prompting the model until it produces the lyrics anyway. If that holds up, it’s a live demonstration that a safety filter your vendor tells you exists on paper may not hold up as a practical shield, and that matters directly for indemnification, which is the next section.
Can You Even Own What Your AI Creates?
Separate from all of the training-data litigation is a quieter but equally important question: can your business actually copyright content it generates with AI?
Stephen Thaler spent years trying to register a copyright for artwork generated entirely by an AI system, with no human author listed, and lost at every level of the federal court system. The Supreme Court denied his final appeal on March 2, 2026, closing the door for good under current US law. The underlying principle survives: US copyright law requires a human author, and content generated by a machine with no meaningful human creative input isn’t eligible for copyright protection at all.
This matters more for businesses than it might first appear. If your team generates marketing copy, product images, or code entirely through AI, with no human selection, arrangement, or editing involved, you may have no legal grounds to stop a competitor from copying it outright. The safer practice, and the one that actually preserves your ability to enforce your own rights, is treating AI output as a first draft that a human meaningfully shapes, edits, and arranges before it becomes a finished asset. That human contribution is what creates something you can actually own.
What This Actually Means for Your Business
Pull all of the above together and a few concrete, actionable points fall out.
Your Vendor’s Indemnification Has Conditions, Read Them
Every major AI vendor now offers some form of copyright indemnification to paying business customers. Microsoft’s Copilot Copyright Commitment is the most structured: it comes with a specific checklist of mitigations you’re required to have enabled, and it expects you to keep a red-team or measurement report on file to prove it if a claim is ever made. OpenAI’s Copyright Shield offers comparable protection, but only for ChatGPT Enterprise and API customers, not the free tier or ChatGPT Plus. Anthropic and Google both offer indemnification for enterprise customers, but the conditions are typically negotiated into the contract in prose rather than laid out as a concrete, auditable checklist.
That gap matters. The Sony and Warner Chappell suit against Anthropic specifically alleges that a safety guardrail can be defeated by simply re-prompting the model, and re-prompting past a safety filter is exactly the kind of “customer misuse” carve-out that voids an indemnity in most vendor agreements. If your business can’t point to a documented, repeatable process showing you kept your usage within the vendor’s stated conditions, you may discover the hard way that the indemnity you assumed you had doesn’t actually apply to your situation.
Document Human Authorship
Given the Thaler ruling, any AI-assisted content your business intends to protect, publish widely, or license to others should have a documented trail of human involvement: prompts refined by a person, output selected and edited by a person, final arrangement decided by a person. This isn’t bureaucratic overhead for its own sake. It’s the difference between owning an asset and owning nothing enforceable.
If You Operate Internationally, the EU AI Act Applies to You Too
US fair use rulings don’t travel. The EU AI Act requires providers of general-purpose AI models to publish detailed summaries of their training data and to respect copyright opt-outs under the EU’s text and data mining exception, with enforcement authority in place since August 2026. A non-binding European Parliament resolution from March 2026 has even floated a flat licensing fee on global AI revenue and argued EU copyright rules should apply regardless of where a model was actually trained. If your business serves EU customers or deploys AI-generated content into EU markets, don’t assume a US court’s fair use finding gives you any cover there.
Licensing Is Becoming the Safer Default
The volume of licensing deals signed over the past two years, OpenAI’s agreements with the Financial Times, Vox Media, The Atlantic, and a reported $70 million a year deal with Reddit, Google’s own Reddit licensing arrangement, Universal Music’s settlement with Udio, Warner Music’s settlement with Suno, is itself a signal. None of these companies would be paying for licenses if they were confident fair use was a guaranteed defense. If a vendor your business relies on has moved toward licensed data and transparent sourcing, that’s a real risk-reduction signal worth factoring into your vendor selection, not just a marketing line.
A Practical Risk Checklist
- Read your AI vendor’s actual indemnification terms, not just the marketing page. Note the specific conditions that could void it, especially around bypassing safety filters or guardrails.
- Keep a record of any red-teaming, content filtering, or safety measures you’ve enabled, particularly if your vendor is Microsoft, whose commitment explicitly asks for one.
- Add a human review and editing step before any AI-generated content goes out publicly, both for quality control and to preserve your ability to claim copyright in the finished product.
- If you operate in the EU or serve EU customers, confirm your AI vendor’s GPAI training data summary and copyright compliance policy, published under the AI Act, actually exists and covers the model you’re using.
- Favor vendors that have moved toward licensed training data and published data provenance information over those still fighting piracy allegations in active litigation.
- Revisit your contracts annually. This area of law is moving fast enough that terms that were adequate in 2025 may already be outdated.
AI Copyright Lawsuits at a Glance
| Case | Status (September 2026) | Core Issue | Why It Matters for Your Business |
|---|
| Bartz v. Anthropic | Settled, $1.5B, final approval July 2026 | Training on legal data can be fair use; training on pirated data is not | Sets the benchmark value and legal test for every other case |
| Thomson Reuters v. Ross Intelligence | On appeal, 3rd Circuit (argued June 2026) | Trial court rejected fair use for AI training on proprietary content | First appellate ruling on AI training fair use, expected to guide everything else |
| Getty Images v. Stability AI (UK) | Getty lost core claims, limited trademark win, appeal pending | Model weights are not a “copy” under UK law | Where training happens can decide which country’s courts even have jurisdiction |
| Disney/Universal/WBD v. Midjourney | Ongoing, discovery phase, no trial date | Alleges AI-generated outputs replicate copyrighted characters | Shows output-level infringement, not just training data, is a live legal theory |
| NYT v. OpenAI and Microsoft | Ongoing, DOJ backed OpenAI/Microsoft in Sept 2026 | Whether training on news content is fair use | Signals current federal government posture favors the AI industry here |
| Cox Communications v. Sony Music | Decided, SCOTUS, March 2026 | Contributory liability requires intent, not mere knowledge | Raises the bar for suing AI platforms just because a tool could be misused |
| Music publishers v. Anthropic | Ongoing, escalating, five suits filed | Alleges lyric safety guardrails can be defeated by re-prompting | Direct proof that vendor indemnification conditions can actually be tested and broken |
| Thaler v. Perlmutter | Final, SCOTUS denied review March 2026 | Purely AI-generated output isn’t copyrightable | You may not own enforceable rights in fully AI-generated business content |
The Bottom Line
None of this litigation is going to conclude with a single, tidy answer for the whole industry. What’s emerging instead is a workable pattern: courts are rewarding AI companies that can show their training data was legally acquired, punishing the ones that cut corners with pirated content, and raising the bar for holding a platform liable just because a tool theoretically could be misused. For your business, the practical task isn’t predicting how any one lawsuit ends. It’s making sure your own house is in order: know what your vendor’s indemnification actually requires of you, keep a human hand in anything you intend to own or defend, and treat your AI vendor selection as a real risk decision rather than a checkbox. The lawsuits will keep coming. The businesses that come out fine are the ones that read the terms before they needed to.
