AI Copyright Litigation in 2026: Inside the Global Lawsuits Reshaping AI Training Data

Direct answer: 2026 has produced the densest sequence of AI copyright milestones to date, and they point in different directions depending on where you look. The US Supreme Court declined to review AI-authorship questions in March 2026, Anthropic's $1.5 billion Bartz v. Anthropic book-training settlement received final approval in July 2026, and a separate NYT v. OpenAI summary-judgment fight remains unresolved. The UK High Court rejected Getty Images' secondary copyright claim against Stability AI, German courts ruled against OpenAI and Suno over song-lyric memorization, and both Australia and the UK publicly rejected broad AI-training copyright exemptions. No appellate court anywhere has yet settled whether training an AI model on copyrighted material is lawful, which means every business building or buying AI capability is currently operating in a legal environment courts are still actively defining.

The 2026 Copyright Reckoning: What's Actually Happening

Every major AI system built in the last several years was trained on enormous quantities of text, images, code, and audio, much of it copyrighted, most of it used without a specific license negotiated for that purpose. For years, that fact sat mostly in the background — acknowledged, occasionally criticized, rarely tested in a courtroom with a final ruling attached. 2026 is the year that changed. A dense sequence of rulings, settlements, and government policy decisions across multiple countries has moved AI copyright from a theoretical debate into a body of actual, sometimes contradictory, legal precedent, and the pace has not slowed as the year has progressed.

What makes this moment different from the years of quieter debate that preceded it is not any single ruling — it is the density of activity across so many independent legal systems at once, each reasoning from its own statutory text, its own procedural rules, and its own sense of what copyright is fundamentally meant to protect. A US court asking whether training on a book is "fair use" is applying a specific, four-factor statutory test with decades of prior case law behind it. A UK court asking whether a model's weights constitute a "copy" is applying a completely different definitional framework rooted in the Copyright, Designs and Patents Act. A German court asking whether a model can be made to reproduce specific lyrics is asking a third, more output-focused question entirely. These are not three courts converging on one answer through slightly different paths — they are three courts answering three different legal questions, shaped by three different statutory traditions, about what is functionally the same underlying technology. That is precisely why the results diverge as sharply as they do, and why no business should expect a tidy, unified global standard to emerge quickly, even once more appellate rulings land.

Start with the United States, where the most financially significant development so far is the Bartz v. Anthropic settlement. On 20 July 2026, a court gave final approval to Anthropic's $1.5 billion settlement covering roughly 482,000 works, which works out to an implied rate of approximately $3,113 per book, according to 2026 settlement reporting tracked by Axis Intelligence's AI Copyright Lawsuits Tracker. That is one of the largest copyright settlements in history in absolute dollar terms, and it resolved a huge share of the authors' claims against Anthropic in one stroke — but not all of them. Roughly 350 opt-out claimants chose not to participate in the class settlement and are instead pursuing individual statutory-damages claims that could reach up to $150,000 per work, a per-work ceiling that, multiplied across even a modest number of titles, dwarfs the settlement's implied per-book rate. That gap between what the class settlement paid per work and what an individual opt-out plaintiff could still win is significant enough that Axis Intelligence's tracker gives it its own name: the Opt-Out Exposure Delta.

The Bartz settlement did not exist in isolation. In the same general window, the US Supreme Court denied certiorari on 2 March 2026 in the Thaler case, declining to disturb the lower-court position that copyright requires human authorship — meaning purely AI-generated output, without a human author, still cannot be copyrighted in the US, a question that had been percolating through the courts for years and is now settled at least until Congress or a future case revisits it. Thomson Reuters v. Ross Intelligence produced a different kind of milestone: a federal court granted summary judgment for Thomson Reuters, and the case moved to the Third Circuit for oral argument on 11 June 2026, making it the first appellate-level review of fair use specifically in the context of AI training, per Norton Rose Fulbright's 2026 litigation update. Meanwhile, the New York Times' case against OpenAI and Microsoft remains very much alive, with summary-judgment cross-motions scheduled for April 2026 in the Southern District of New York — a case widely watched because of how directly it tests fair use against one of the most consequential AI companies in the market. And in January 2026, Universal Music, Concord, and ABKCO filed a combined $3.1 billion suit against Anthropic, a case distinct from the book-publishing settlement and aimed instead at the music industry's own claims about lyrics and composition rights, with a related case, Sony v. Suno, now scheduled for dispositive motions on 9 April 2027.

Outside the US, the picture diverges sharply rather than converging toward a single global standard. The UK High Court ruled on 4 November 2025 in Getty Images v. Stability AI that Stable Diffusion's model weights are not an infringing "copy" under the UK's Copyright, Designs and Patents Act, rejecting Getty's secondary copyright infringement theory outright — Getty won only a narrow, largely historic trademark finding related to watermarks appearing in some outputs, a much smaller victory than the broad infringement finding it had sought, per Ropes & Gray's analysis of the ruling. Germany moved in the opposite direction: Munich Regional Court I ruled against OpenAI on 11 November 2025 over the memorization and reproduction of nine German songs' lyrics, ordering the company to stop storing and outputting those lyrics, pay damages, and disclose usage data, according to the Library of Congress Global Legal Monitor's January 2026 report. A separate Munich ruling in GEMA v. Suno, decided 31 July 2026, applied similar reasoning against a different AI company, in a decision reportedly notable for addressing US copyright law as it applied to US-based training activity — a German court reasoning about American law, which is itself an unusual and consequential detail.

Government policy has moved in the same fractured direction as the courts. Both the UK and Australia spent significant parts of 2026 explicitly rejecting the idea of a broad AI-training copyright exemption — the kind of blanket carve-out that would have simplified life considerably for AI developers by removing the licensing question almost entirely. The UK government's 18 March 2026 Report on Copyright and Artificial Intelligence abandoned its previously preferred text-and-data-mining opt-out exception, leaving no near-term legislative reform on the table. Australia went further in publicly stating its position: Prime Minister Albanese announced on 15 July 2026 that a mandatory framework protecting creators under "consent, credit and compensation" principles is being developed, explicitly rejecting a training exemption, though without yet specifying a fixed compensation mechanism or a timeline for implementation. Put together, courts on two continents have reached opposite conclusions about secondary infringement and memorization, and governments on two continents have independently walked away from the idea of simply exempting AI training from copyright law altogether — which is about as clear a signal as you can get that there is no emerging global consensus, only an emerging global pattern of active, unresolved contestation.

A World of Diverging Rulings: The Global Regulatory Map

The lack of consensus described above is easiest to see clearly when the seven markets most relevant to AI copyright risk are examined side by side, because the differences are not subtle — they are structural disagreements about what copyright law is even supposed to protect against when the "infringer" is a statistical model rather than a person making a literal copy.

United States

The US remains the most litigated jurisdiction by a wide margin, and it is also the jurisdiction with the most unresolved fair-use questions. The Supreme Court's March 2026 cert denial in the Thaler case settled the authorship question — no human author, no copyright — but left the much bigger question, whether training on copyrighted material without a license is fair use, entirely open. Thomson Reuters v. Ross Intelligence is the case to watch on that front: with a Third Circuit oral argument held 11 June 2026, it is positioned to become the first appellate ruling squarely addressing fair use in AI training, and whichever way it goes will shape how every pending case, including NYT v. OpenAI, gets argued and evaluated afterward. The Bartz v. Anthropic settlement, at $1.5 billion and covering roughly 482,000 works at an implied $3,113 per book, resolved the class claims but left the door open for the roughly 350 opt-out claimants pursuing individual statutory-damages claims worth up to $150,000 per work — and it explicitly did not grant Anthropic a forward-looking license, since the release covers only past conduct through August 2025, meaning Anthropic's future training practices are not protected by anything in that settlement. Layer on top of that the Universal Music/Concord/ABKCO $3.1 billion suit against Anthropic and the Sony v. Suno case moving toward dispositive motions in April 2027, and the picture is of a jurisdiction with enormous financial exposure already realized, enormous financial exposure still pending, and no appellate court yet on record about whether the underlying training conduct is even lawful.

United Kingdom

The UK's defining 2026 moment was the High Court's ruling in Getty Images v. Stability AI, delivered 4 November 2025 but continuing to shape UK AI-copyright strategy well into 2026. The court's finding that Stable Diffusion's model weights do not constitute an infringing "copy" under UK law was, per Latham & Watkins' analysis, a first-of-its-kind rejection of the secondary-infringement theory that rights holders had hoped would let them reach AI training conducted anywhere in the world, so long as the resulting model was later imported into or used in the UK. That theory's rejection matters well beyond this single case, because it was one of the more creative legal strategies rights holders had developed specifically to work around the difficulty of proving what happened during training, which usually occurs outside any court's direct evidentiary reach. Getty's only win was a narrow, largely historic trademark finding tied to watermarks appearing in some Stable Diffusion outputs — a real result, but a small fraction of what the case originally sought. On the policy side, the UK government's 18 March 2026 Report on Copyright and Artificial Intelligence walked away from its previously favored text-and-data-mining opt-out exception, per Reed Smith's analysis, leaving UK copyright law unreformed and the Getty ruling's interpretation of existing law as the most concrete guidance currently available to businesses operating there.

UAE and Dubai

There is no distinct regional-specific reporting on AI copyright litigation or policy in the UAE or Dubai in the current research base. That is worth stating directly rather than working around it, because it means businesses operating in that market cannot currently point to a UAE-specific court ruling or government policy statement comparable to the UK's Getty decision or Australia's "consent, credit and compensation" framework. The absence of reporting is not the same as an absence of risk — international AI companies operating in the UAE remain subject to the same underlying training-data provenance questions being litigated everywhere else, even without a local case to reference yet.

Australia

Australia has taken one of the clearest public policy positions of any jurisdiction reviewed here, even without a headline court case to match the UK's Getty ruling or the US's Bartz settlement. The government explicitly rejected a broad AI-training copyright exemption, and Prime Minister Albanese announced on 15 July 2026 that a mandatory framework is being developed around three stated principles: consent, credit, and compensation for creators, according to reporting from the University of Melbourne's Pursuit publication. What has not yet been specified is exactly how compensation would be calculated or when the framework would actually take effect — Australia has stated its direction clearly while leaving the mechanics for later, which puts it in a middle position between the UK's now-abandoned reform effort and the pure case-law-driven approach playing out in the US courts. For AI companies training models on Australian creative content, the clearest current signal is that a licensing-based compensation obligation is coming in some form, even though its precise shape remains undefined.

Germany

Germany has produced the most direct, plaintiff-favorable rulings of any jurisdiction covered here, and both came out of the same court. Munich Regional Court I ruled against OpenAI on 11 November 2025 over the memorization and reproduction of nine German songs' lyrics — including tracks like "Atemlos" and "Männer" — ordering OpenAI to stop storing and outputting the lyrics, pay damages, and disclose usage data, per the Library of Congress Global Legal Monitor's January 2026 report. The court's focus on memorization — the model's ability to reproduce protected text essentially verbatim in response to a simple prompt — is a meaningfully different legal theory than the "was training itself infringing" question at the center of the US and UK cases; it sidesteps some of the harder abstract questions about what happens during training and focuses instead on a much more concrete, demonstrable harm: the model outputting copyrighted material to end users. A second Munich ruling, in GEMA v. Suno on 31 July 2026, applied similar reasoning against a different AI company, and was reportedly notable for engaging with US copyright law as applied to training activity that occurred in the United States — a German court reasoning through American legal standards to reach its conclusion, which raises real questions about how far a national court's reasoning in one country can practically extend into training conduct that physically occurred somewhere else entirely.

Europe and France

Beyond Germany, France has been the most active EU jurisdiction on AI copyright policy in 2026, though through legislative proposal rather than court ruling. A proposed law that would create a presumption of exploitation of cultural content by AI providers training their models was scheduled for discussion in the French Senate in April 2026, according to Bird & Bird's France AI tracker. A presumption of exploitation would function as a significant burden-shifting tool: rather than a rights holder having to prove an AI company specifically used their protected work in training, the law would presume that use occurred once certain conditions are met, shifting the burden to the AI company to prove otherwise. That is a structurally different, and for AI developers considerably more difficult, legal posture than the case-by-case evidentiary battles playing out in US and UK courts, and if it passes, France would have one of the more rights-holder-favorable frameworks in the world, at least on paper.

China

As with the UAE, there is no distinct regional-specific reporting on AI copyright litigation or policy found in the current research base for China. This stands in notable contrast to the extensive, binding regulatory activity China has undertaken in other AI policy areas during the same period, discussed at greater length in Scult's coverage of the parallel wave of AI companion-chatbot child-safety regulation. The lack of comparable copyright-specific reporting here should be read as a gap in currently available research on this specific question, not as evidence that China has no domestic position on AI training and copyright.

Why This Matters Beyond Publishing and Music

It is easy to read all of this as a dispute between a handful of very large AI labs and a handful of very large media companies, with limited relevance to an ordinary business building or buying AI capability. That reading misses the actual mechanism of risk. Every one of these cases turns, in one way or another, on a single underlying question: where did the data that trained this model actually come from, and was it used lawfully? That question does not stay contained to the original defendant. It travels downstream to every business that fine-tunes a foundation model on its own data, every business that builds a product on top of a third-party model's API, and every business that licenses a vendor's AI-powered feature without asking what that feature was trained on.

The most direct exposure sits with any company doing its own model training or fine-tuning using scraped or aggregated third-party content — text, images, code, or audio gathered without a clear license — because that is precisely the fact pattern at issue in Bartz v. Anthropic, the Munich GEMA rulings, and NYT v. OpenAI. A smaller but real exposure sits with companies further downstream: a business that builds a customer-facing product on top of a foundation model does not control how that model was trained, but if the underlying model is later found to have been trained unlawfully, or if it can be prompted to reproduce protected material — the exact mechanism at issue in the German memorization rulings — the downstream business inherits at least some of that risk through its own use and distribution of the model's output. This is not a hypothetical concern invented for this article; it is the specific legal theory, memorization and reproduction via a simple prompt, that a German court has already used to rule against an AI company, and it is a theory that does not require proving anything about what happened during training itself, only what the deployed system does when asked.

There is also a straightforward commercial dimension that has nothing to do with courtroom risk directly: licensing markets are forming in real time as a consequence of this litigation wave, and companies that wait to see how the case law fully settles before thinking about licensed training data may find themselves negotiating from a weaker position later, once rights holders have more leverage and more precedent behind them. The Bartz settlement's implied rate of roughly $3,113 per book is, whether intended this way or not, an early data point in exactly that kind of price discovery — a real, court-approved number that both future licensing negotiations and future litigation will likely reference. Businesses that treat training-data provenance as a genuine input cost, the same way they would treat a software license or a data-processing agreement, are positioning themselves ahead of a market that is still working out its own pricing, rather than behind one that has already settled into place.

The size of a company is not much protection here, and in some respects it cuts the other way. The businesses named as defendants throughout this piece — Anthropic, OpenAI, Stability AI, Suno, Microsoft — are among the best-resourced technology companies in the world, with substantial legal budgets and, in Anthropic's case, the ability to absorb a $1.5 billion settlement and continue operating. A smaller company fine-tuning an open-weight model on its own scraped dataset, or a mid-sized software vendor building a niche AI feature on top of a foundation model without asking hard questions about that model's training data, generally has far less capacity to absorb a comparable legal event, even a much smaller one in absolute dollar terms. That asymmetry is worth sitting with: the companies best equipped to survive this litigation wave financially are also, in most cases, the ones whose conduct is most directly at issue in it, while smaller downstream businesses building on top of their models carry real exposure with far less capacity to weather it.

Finally, there is a reputational and customer-trust dimension that matters even for businesses with limited direct legal exposure. Enterprise customers evaluating an AI vendor increasingly ask, as a matter of routine procurement diligence, what data trained the model they are being asked to build on, and whether that vendor has any exposure to the kind of litigation described throughout this piece. A vendor that cannot answer clearly is a harder sell in 2026 than the same vendor would have been in 2023, before this wave of rulings and settlements made the question concrete and financially quantified rather than abstract.

What Comes Next: Precedent, Appeals, and How Businesses Should Prepare

The single most important thing to understand about where this litigation stands in August 2026 is that nothing has been finally settled at the level that matters most. As Axis Intelligence's tracker puts it, no US appellate court has yet answered the core fair-use question. The Third Circuit's review of Thomson Reuters v. Ross Intelligence, following oral argument on 11 June 2026, is positioned to be the first word on that question from an appellate bench, and whichever way it lands will immediately become the reference point every other pending US case gets argued against, including NYT v. OpenAI's April 2026 summary-judgment fight and the still-developing Universal Music/Concord/ABKCO and Sony/Suno music-industry cases. Outside the US, the UK's Getty ruling and Germany's GEMA rulings already point in different directions on different legal theories — secondary infringement rejected in the UK, memorization-based liability upheld in Germany — and neither country's courts are bound by the other's reasoning, so it is entirely possible for this fragmentation to persist or even deepen rather than resolve toward a single global standard.

For a business trying to make practical decisions in the middle of that uncertainty, a few concrete steps hold up regardless of how any individual case eventually resolves. The first is provenance diligence on any AI vendor or model a company is building on: asking directly what data trained the model, whether that data was licensed, and whether the vendor has current or pending litigation exposure of the kind catalogued throughout this piece. This is now a standard, reasonable procurement question, not an unusual one, and a vendor's willingness or unwillingness to answer it clearly is itself useful signal. The second is treating output-side risk as distinct from training-side risk: even a company with no exposure to how a model was trained can face real risk if that model reproduces copyrighted material in its outputs, which is exactly the theory Germany's courts have already validated, so any AI-powered product that generates text, images, or audio for end users benefits from output-filtering and monitoring designed to catch unintentional reproduction of protected material before it reaches a customer.

The third is building any custom AI capability — whether that is a fine-tuned model, a retrieval-augmented system pulling from a licensed content library, or an AI-powered feature layered onto an existing product — with a clear, documented record of where the underlying data came from and under what terms. That kind of documentation is unglamorous, easy to defer, and exactly the kind of evidence that determines how a company fares if a provenance question is ever raised, whether by a regulator, an enterprise customer's procurement team, or opposing counsel in litigation. Businesses working with a development partner on custom AI systems are generally better served treating this as a core requirement scoped in from the start of a project — see Scult's custom software development services for how data-provenance and licensing considerations get built into a system's architecture rather than bolted on afterward, and Scult's compliance resources for tracking how obligations in this area continue to shift as more rulings land. Given how unsettled the underlying legal terms still are — a glance at Scult's AI and technology glossary is a reasonable place to get comfortable with terms like fair use and text-and-data-mining exceptions before they show up in a vendor contract — the businesses that come out ahead here are the ones treating provenance and licensing as an ordinary cost of doing business now, rather than a problem to solve only once a court forces the issue.

What Businesses Want to Know About the AI Copyright Litigation Wave

What happens to authors who opted out of the Anthropic settlement?

Authors who opted out of the Bartz v. Anthropic class settlement are pursuing individual statutory-damages claims rather than accepting the settlement's implied rate of roughly $3,113 per book, according to Axis Intelligence's 19 August 2026 tracker. Individual claims can seek statutory damages of up to $150,000 per work, a ceiling far above the class settlement's per-book rate, which is why roughly 350 claimants made that choice despite it meaning a longer, less certain path to any recovery. This is precisely the dynamic behind what Axis Intelligence calls the Opt-Out Exposure Delta — the gap between the settled class rate and the much larger potential exposure Anthropic still faces from claimants who chose to litigate individually rather than accept the collective settlement.

Has the Bartz v. Anthropic settlement been finally approved?

Yes. The $1.5 billion settlement received final court approval on 20 July 2026, per Axis Intelligence's tracker. It covers approximately 482,000 works at an implied rate of roughly $3,113 per book, making it one of the largest copyright settlements on record in dollar terms. Final approval closed the class-wide claims covered by the settlement, but it did not close the broader litigation landscape around Anthropic or the AI industry generally — the roughly 350 opt-out claimants pursuing individual damages, the separate $3.1 billion Universal Music/Concord/ABKCO suit against Anthropic, and the unresolved question of whether AI training constitutes fair use all remain live regardless of this specific settlement's approval.

What is Anthropic arguing in the music publishers' $3.1 billion case?

The $3.1 billion suit brought by Universal Music, Concord, and ABKCO against Anthropic, filed in January 2026, is a separate matter from the Bartz book-training settlement, centered on music publishing and lyric rights rather than books. Axis Intelligence's 19 August 2026 tracker notes this case as part of the broader wave of music-industry litigation against AI companies, alongside the Sony v. Suno case now scheduled for dispositive motions on 9 April 2027. The specific legal arguments Anthropic is advancing in its defense are still developing as the case proceeds, but the case sits within the same general fair-use battleground as Anthropic's book-related litigation, just applied to a different category of copyrighted content with its own licensing conventions and industry structure.

Is AI training on copyrighted works fair use in the US?

As of Axis Intelligence's 19 August 2026 tracker, no US appellate court has yet answered this question. Thomson Reuters v. Ross Intelligence produced a district-court summary judgment for Thomson Reuters and is now before the Third Circuit, with oral argument held 11 June 2026, positioning it to become the first appellate-level ruling on fair use in AI training specifically. Until that ruling — or another appellate decision — lands, the fair-use status of AI training in the US remains a live, actively litigated question rather than settled law, which is precisely why cases like NYT v. OpenAI, with summary-judgment motions pending in April 2026, carry so much weight: each one is contributing to case law that does not yet have appellate-level clarity behind it.

What happened in GEMA v. Suno, and why does it matter outside Germany?

A Munich court ruled against Suno on 31 July 2026 in a case brought by GEMA, Germany's music-rights organization, applying reasoning similar to the earlier Munich ruling against OpenAI over song-lyric memorization, per Axis Intelligence's tracker. What makes this ruling notable beyond Germany's borders is that it reportedly engaged with US copyright law as applied to training activity that occurred in the United States — meaning a German court reasoned through American legal standards to reach its conclusion about a US company's US-based conduct. That cross-border legal reasoning raises real questions about how far any single national court's rulings can practically reach into AI training conduct that occurred entirely outside that court's own jurisdiction, a question with no settled answer yet.

When will the Sony v. Suno fair use case be decided?

Dispositive motions in Sony v. Suno are scheduled for 9 April 2027, per Axis Intelligence's 19 August 2026 tracker, which means a resolution is not expected imminently. This timeline places it well behind several other major 2026 cases, including the Bartz v. Anthropic settlement, already finally approved, and the Thomson Reuters v. Ross Intelligence appeal, already argued before the Third Circuit as of 11 June 2026. Businesses tracking AI copyright risk in the music and audio-generation space should treat Sony v. Suno as a longer-horizon case to watch rather than an imminent data point, with the German GEMA v. Suno ruling likely to remain the more immediately relevant precedent for that company's litigation exposure in the nearer term.

Did the $1.5 billion settlement give Anthropic a licence to use those books?

No. Per Axis Intelligence's 19 August 2026 tracker, the settlement's release covers only past conduct through August 2025 — it resolves the specific claims covered by the class action but does not grant Anthropic any forward-looking license to continue using the same works in future training. This is an important distinction for anyone assuming a large settlement functions like a licensing deal: it does not. Anthropic's future training practices involving these or similar works remain subject to whatever the broader, still-unsettled legal landscape around AI training and fair use ultimately determines, including the outcome of the Third Circuit's review of Thomson Reuters v. Ross Intelligence and any future litigation.

What is the 'Opt-Out Exposure Delta'?

The Opt-Out Exposure Delta is a term used in Axis Intelligence's 19 August 2026 tracker to describe the gap between what the Bartz v. Anthropic class settlement paid per work — roughly $3,113 per book — and the far larger potential damages, up to $150,000 per work, that the roughly 350 opt-out claimants could still win by pursuing individual statutory-damages claims instead of accepting the class settlement. It functions as a proprietary way of quantifying just how much financial exposure a company can retain even after a large, seemingly comprehensive settlement, because a settlement only resolves the claims of the parties who accept it — anyone who opts out keeps their own claim, at their own preferred damages theory, fully alive.

What did the US Supreme Court decide about AI-generated work authorship in March 2026?

The US Supreme Court denied certiorari on 2 March 2026 in the Thaler case, declining to review the question of whether purely AI-generated work can receive copyright protection. That denial left standing the lower courts' position that copyright requires human authorship, meaning an output with no human author remains uncopyrightable in the US under current law. This resolved one specific, narrower question in the broader AI-copyright landscape — who, or what, can hold a copyright in AI-generated output — while leaving the much larger and more commercially consequential question, whether training on copyrighted material is itself lawful, completely untouched and still working through the lower and appellate courts.

What did the court rule in Thomson Reuters v. Ross Intelligence?

A federal court granted summary judgment for Thomson Reuters, according to Norton Rose Fulbright's 2026 litigation update, in a case centered on whether Ross Intelligence's use of Thomson Reuters' legal-research content to train a competing AI product was protected by fair use. The ruling favored Thomson Reuters at the district-court level, and the case has since moved to the Third Circuit, which heard oral argument on 11 June 2026. Because it is now before an appellate court, this case is widely regarded as the most significant near-term opportunity for a definitive, appellate-level answer to the fair-use-in-AI-training question that every other pending US case, including NYT v. OpenAI, is effectively also litigating in parallel.

When is the Third Circuit hearing oral argument in Thomson Reuters v. Ross Intelligence?

Oral argument was held on 11 June 2026, per Norton Rose Fulbright's 2026 litigation update. This makes Thomson Reuters v. Ross Intelligence the first case to reach appellate oral argument specifically on the question of fair use in AI training, ahead of other closely watched cases like NYT v. OpenAI, whose summary-judgment fight was still pending at the district-court level as of April 2026. Whatever the Third Circuit ultimately decides will likely be cited heavily in every other pending AI-copyright case in the US, simply because it will be the first appellate-level guidance available on this exact question.

What is at stake in NYT v. OpenAI and Microsoft?

NYT v. OpenAI and Microsoft tests whether training AI models on New York Times journalism, and generating outputs that can closely track or reproduce that journalism, is protected by fair use or constitutes infringement, with summary-judgment cross-motions scheduled for April 2026 in the Southern District of New York, per 2026 litigation-tracker reporting. The case is closely watched because of the scale and prominence of the parties involved — one of the most recognized news organizations in the world against one of the most consequential AI companies — and because a ruling either way will land squarely in the middle of the broader, unresolved fair-use debate that Thomson Reuters v. Ross Intelligence is simultaneously working through at the appellate level.

Why did Universal Music, Concord, and ABKCO sue Anthropic for $3.1 billion?

Universal Music, Concord, and ABKCO filed their combined $3.1 billion suit against Anthropic in January 2026, per 2026 news reporting tracked by Axis Intelligence, placing music-publishing and lyric-rights claims against Anthropic on a similar litigation track to the book-publishing claims resolved in the Bartz settlement. The suit reflects the music industry pursuing its own claims independently rather than relying on outcomes from book-industry litigation, recognizing that lyrics, compositions, and recordings raise their own distinct rights questions. It sits alongside the separate Sony v. Suno case, now scheduled for dispositive motions in April 2027, as one of two major music-industry fronts in the broader 2026 AI copyright litigation wave.

What did the Munich Regional Court rule in GEMA v. OpenAI over song lyrics?

Munich Regional Court I ruled against OpenAI on 11 November 2025, finding that the company's model had memorized and could reproduce the lyrics of nine German songs, and ordering OpenAI to stop storing and outputting those lyrics, pay damages, and disclose usage data, according to the Library of Congress Global Legal Monitor's January 2026 report. The ruling's focus on memorization and reproduction — rather than on the training process itself — gave the court a concrete, demonstrable harm to point to: the model outputting protected lyrics essentially verbatim in response to prompts, a much easier fact pattern to establish than abstract claims about what happened during training.

Which nine German songs were at issue in GEMA v. OpenAI?

Reporting on the case, including coverage cited by the Library of Congress Global Legal Monitor, names tracks including "Atemlos" and "Männer" among the nine German songs found to have been memorized and reproducible by OpenAI's model. These are well-known German-language songs, and the court's finding that their lyrics could be reproduced through simple prompts was central to establishing the memorization-based infringement theory that the ruling ultimately rested on, distinct from broader claims about the lawfulness of the training process that produced the model in the first place.

Did the UK High Court find Stability AI liable for copyright infringement in Getty Images v. Stability AI?

No, not for the core claim Getty was pursuing. The High Court ruled on 4 November 2025 that Stable Diffusion's model weights are not an infringing "copy" under the UK's Copyright, Designs and Patents Act, rejecting Getty's secondary copyright infringement theory, according to Ropes & Gray's analysis of the ruling. Getty did win a narrow, largely historic trademark-related finding connected to watermarks appearing in some outputs, but that result was a small fraction of the broad infringement finding Getty had sought, making this predominantly a loss for Getty on the central legal question the case was built around.

What did the UK court rule about AI model weights not being a 'copy' under UK copyright law?

The court held that Stable Diffusion's model weights — the trained parameters that make up the AI model itself — do not constitute a "copy" of the copyrighted images used in training, under the specific legal definition that applies within the UK's Copyright, Designs and Patents Act, according to Latham & Watkins' analysis of the ruling. This was significant because it rejected the legal theory that would have let rights holders reach AI training conducted anywhere in the world, simply by pointing to the resulting model being used or distributed within the UK afterward — a strategy designed specifically to work around the practical difficulty of gathering evidence about what actually happened during training.

Did Getty Images win anything against Stability AI?

Yes, but narrowly. Getty won a limited trademark-related finding tied to watermarks that appeared in some Stable Diffusion outputs, a largely historic result rather than the sweeping copyright-infringement victory the case had originally sought, according to reporting from both Mayer Brown and Ropes & Gray. The core secondary copyright infringement claim — the theory that model weights themselves constitute an infringing copy — was rejected outright by the High Court. For a case this closely watched, the practical outcome favored Stability AI far more than Getty, and the ruling has been treated across the industry as a significant setback for rights holders pursuing similar theories elsewhere.

Why did the UK government abandon its proposed text-and-data-mining opt-out exception?

The UK government's 18 March 2026 Report on Copyright and Artificial Intelligence abandoned its previously preferred text-and-data-mining opt-out exception, according to Reed Smith's analysis, though the report itself is the primary source for the decision rather than a single named reason. The proposed exception would have let AI developers train on copyrighted content by default unless a rights holder specifically opted out, an approach that drew significant pushback from creative industries who argued it inverted the normal presumption that a license is required before use. Abandoning the proposal left UK copyright law unreformed on this question, meaning the Getty Images v. Stability AI ruling's interpretation of existing law remains the most concrete available guidance for businesses operating in the UK.

What did the UK's 2026 Report on Copyright and Artificial Intelligence conclude?

Published 18 March 2026, the report's most consequential conclusion was to abandon the government's previously favored text-and-data-mining opt-out exception, leaving no near-term legislative reform of UK copyright law specific to AI training. In practical terms, that means the UK is not moving toward a bespoke statutory framework in the near term, and businesses and courts alike are left working with existing copyright law, as interpreted through rulings like Getty Images v. Stability AI, rather than a new AI-specific statute. That makes the UK's legal environment more court-driven and precedent-based for now, compared to jurisdictions like Australia that have committed to building a new regulatory framework instead.

Why does Australia refuse to create an AI copyright training exemption?

Australia's government rejected a broad AI-training copyright exemption on the basis that it would let AI companies use creative content without the consent, credit, or compensation the government wants to guarantee creators, according to the University of Melbourne's Pursuit publication. Prime Minister Albanese's 15 July 2026 announcement framed the alternative not as blocking AI development, but as ensuring creators are compensated as part of it — a mandatory framework rather than a voluntary licensing market left to negotiate itself. This puts Australia in a notably different position from jurisdictions relying primarily on case-by-case court rulings, since it is committing, at least in policy direction, to a structured regulatory answer rather than waiting for litigation to define the boundaries.

What does 'consent, credit and compensation' mean in Australia's AI copyright policy?

"Consent, credit and compensation" is the three-principle framework Prime Minister Albanese announced on 15 July 2026 as the basis for Australia's forthcoming mandatory AI-copyright framework, per Pursuit's July 2026 reporting. Consent implies creators would need to agree to their work being used in AI training, credit implies attribution requirements, and compensation implies some form of payment for that use — together describing a licensing-style relationship between AI developers and creators rather than either an outright ban or an exemption. As of that announcement, the specific compensation mechanism and implementation timeline had not yet been finalized, leaving the framework's practical operation still to be defined.

How many works are covered under the Bartz v. Anthropic settlement, and at what implied per-book rate?

The settlement covers approximately 482,000 works at an implied rate of roughly $3,113 per book, based on the $1.5 billion total settlement figure, according to 2026 settlement reporting. That per-book figure is one of the more concrete, citable data points to emerge from the entire 2026 AI copyright litigation wave, and it is already being referenced as an early benchmark in discussions about what fair licensing for AI training data might look like going forward, even though it emerged from a litigation settlement rather than a negotiated commercial licensing deal.

What statutory damages can opt-out authors seek individually in the Anthropic litigation?

Authors who opted out of the Bartz v. Anthropic class settlement can pursue individual statutory-damages claims of up to $150,000 per work, according to Axis Intelligence's 19 August 2026 tracker. That ceiling is dramatically higher than the settlement's implied per-book rate of roughly $3,113, which is exactly why around 350 claimants chose the opt-out route despite the added time, cost, and uncertainty of individual litigation rather than accepting the class settlement's guaranteed, if smaller, payout.

Is training an AI model on copyrighted books without a license illegal in the US?

There is no single, settled answer to this in the US as of August 2026 — that is precisely why it remains the subject of active litigation. No US appellate court has yet ruled definitively on whether training constitutes fair use, per Axis Intelligence's tracker, and the Bartz v. Anthropic settlement resolved specific claims through a negotiated payment rather than through a court ruling on the underlying legality of the conduct. The Third Circuit's pending review of Thomson Reuters v. Ross Intelligence, argued 11 June 2026, is the case most likely to move this from an open question toward a clearer appellate answer, though even that ruling would not necessarily resolve every fact pattern across every pending case.

What's the difference between 'input' copyright claims (training data) and 'output' copyright claims (generated content) in AI litigation?

Input claims focus on whether it was lawful to use copyrighted material during training in the first place — the theory at the heart of Bartz v. Anthropic and NYT v. OpenAI. Output claims focus on whether what a model actually generates infringes copyright, regardless of the training question — the theory at the heart of Germany's memorization-based GEMA rulings against OpenAI and Suno, which centered on the model's ability to reproduce specific song lyrics via simple prompts. A business can face output-side risk even without any exposure to how a given model was trained, which is why output-side monitoring and filtering matters as an independent risk-management practice, not merely a downstream concern once input questions are resolved.

Could the NYT v. OpenAI ruling set a precedent for all generative AI copyright cases?

It could carry significant persuasive weight, particularly given the scale and prominence of the parties, but its formal precedential reach depends on where and how it is ultimately decided. A district-court ruling on the pending summary-judgment motions would be persuasive but not formally binding on other courts, whereas an appellate ruling would carry more weight within its own circuit. Given that Thomson Reuters v. Ross Intelligence is already ahead of NYT v. OpenAI in reaching appellate review, with oral argument held 11 June 2026, that case may end up shaping the legal landscape NYT v. OpenAI gets decided within, rather than the reverse.

What is 'memorisation' in the context of AI copyright law, per the German GEMA rulings?

Memorisation, as used in the Munich court's reasoning against OpenAI and Suno, refers to a model's demonstrated ability to reproduce specific copyrighted content — in this case, song lyrics — essentially verbatim in response to a prompt, rather than generating genuinely new content that merely reflects patterns learned from training data. The German court's rulings focused specifically on this reproducibility as the basis for liability, ordering OpenAI to stop storing and outputting the lyrics at issue and disclose usage data, according to the Library of Congress Global Legal Monitor. This theory is notable because it does not require resolving the harder abstract question of whether training itself was lawful — it targets a concrete, observable output-side behavior instead.

Can a company be held liable if its AI model can reproduce copyrighted lyrics via simple prompts?

Germany's Munich Regional Court I has already answered this question affirmatively in its rulings against OpenAI and Suno, finding liability based on a model's ability to reproduce specific copyrighted song lyrics through straightforward prompts, and ordering remedies including stopping the storage and output of the lyrics, paying damages, and disclosing usage data. This memorization-based theory of liability is a meaningful risk for any company deploying a generative AI product, regardless of where that company sits in the training-versus-deployment chain, because it attaches to demonstrated output behavior rather than requiring proof of what happened during the training process itself.

Does opting out of the Anthropic class settlement help or hurt an author financially?

It depends entirely on the specific facts of an individual author's claim, which is exactly why the concept of an Opt-Out Exposure Delta exists as a way of thinking about the decision. Accepting the class settlement guarantees a payout at the implied rate of roughly $3,113 per book, with certainty and without further litigation. Opting out forgoes that certainty in exchange for the possibility of statutory damages up to $150,000 per work — a far larger potential number, but one that requires successfully litigating an individual claim rather than simply accepting a negotiated group outcome. Roughly 350 claimants judged that trade-off worth making, but it is a genuinely individual calculation, not a strategy that is automatically better for everyone covered by the settlement.

What is the legal status of AI-generated content copyright registration in the US after the Thaler case?

Following the Supreme Court's denial of certiorari on 2 March 2026, the position that copyright requires human authorship remains the governing US standard, meaning purely AI-generated content, without meaningful human creative input, cannot be registered for copyright protection. This does not mean AI-assisted works are categorically excluded — works involving substantial human creative contribution alongside AI tools have generally been treated differently by the Copyright Office than works where an AI system is effectively the sole author — but the cert denial closed off, at least for now, any near-term US Supreme Court reconsideration of the human-authorship requirement itself.

What compensation mechanism is Australia developing for copyright holders affected by AI training?

As of Prime Minister Albanese's 15 July 2026 announcement, Australia has committed to the principle of compensating creators as part of a mandatory "consent, credit and compensation" framework, but the specific mechanism — how compensation would actually be calculated, collected, and distributed — had not yet been specified, per the University of Melbourne's Pursuit publication. This puts Australia at an earlier implementation stage than jurisdictions like Germany, where courts have already ordered specific damages in individual cases, or the US, where the Bartz settlement has already produced a concrete implied per-book rate through litigation rather than policy design.

Will France's proposed 'presumption of exploitation' law pass the Senate?

As of the most recent reporting reviewed, the proposed law was scheduled for discussion in the French Senate in April 2026, according to Bird & Bird's France AI tracker, but its ultimate passage was not yet determined at that point. Legislative proposals of this kind typically face amendment and negotiation before any final vote, and a presumption-of-exploitation mechanism — which would shift the burden of proof onto AI companies rather than requiring rights holders to prove specific unauthorized use — is a significant enough departure from typical copyright litigation procedure that it would likely draw substantial debate before any final version is adopted.

What would France's presumption-of-exploitation law mean for AI companies training on French cultural content?

If adopted, a presumption of exploitation would mean AI companies training models using French cultural content would be presumed, by default, to have exploited that content once certain conditions are met, shifting the burden onto the AI company to prove otherwise rather than requiring the rights holder to prove unauthorized use occurred. That is a materially more rights-holder-favorable starting position than the evidentiary battles playing out in US and UK courts, where plaintiffs generally bear the burden of establishing what happened during training. For AI companies operating in France, this would meaningfully raise the practical cost and difficulty of defending against a copyright claim, even where their training practices were, in fact, properly licensed.

Is there a global consensus on whether AI training constitutes fair use?

No. The evidence points in the opposite direction: the UK rejected a secondary-infringement theory in Getty Images v. Stability AI while Germany upheld a memorization-based theory against OpenAI and Suno; the US has settled some claims through negotiated payment (Bartz v. Anthropic) while leaving the core fair-use question for appellate courts to resolve; and both the UK and Australia have separately walked away from creating a broad training exemption, opting instead for continued case-by-case legal development or a still-undefined compensation framework. Rather than converging, these jurisdictions are actively diverging on both legal theory and policy approach, and there is no indication as of August 2026 that a single global standard is imminent.

What's the significance of a German court applying US copyright law in GEMA v. Suno?

The Munich court's willingness to engage with US copyright law as applied to training activity that occurred in the United States, in its 31 July 2026 ruling against Suno, is significant because it suggests national courts may be willing to reason across jurisdictional lines when the underlying AI training conduct occurred elsewhere but the resulting model or its effects reach the court's own jurisdiction. That raises real, currently unresolved questions about how far any single national court's reasoning can practically extend into conduct occurring entirely within another country's borders, and it is the kind of cross-border legal question that is likely to recur as more courts around the world confront AI training cases involving multinational companies.

How are music rights holders approaching AI copyright litigation differently than book authors?

Music rights holders, through cases like the GEMA rulings against OpenAI and Suno in Germany and the Universal Music/Concord/ABKCO and Sony/Suno cases in the US, have generally pursued output-focused, memorization-based theories or direct litigation against specific AI companies, rather than the class-action settlement structure that resolved the bulk of the book-publishing claims in Bartz v. Anthropic. Part of this difference likely reflects the structure of the industries themselves — music rights are often held by a smaller number of large rights organizations and labels capable of pursuing coordinated, well-resourced litigation, compared to the much larger and more fragmented population of individual book authors for whom a class settlement made practical sense as a way to secure compensation without each author litigating individually.

What legal theory underlies the 'secondary copyright infringement' claim rejected in Getty v. Stability AI?

Getty's secondary copyright infringement theory argued that Stable Diffusion's model weights themselves constitute an infringing copy of the training images, such that simply having and using the trained model within the UK could be infringing, regardless of where or how the original training occurred. The High Court rejected this, finding that model weights do not meet the legal definition of a "copy" under the UK's Copyright, Designs and Patents Act, according to Latham & Watkins' analysis. The theory's appeal was that it offered a way to reach AI training conducted anywhere in the world, by focusing on the model's presence and use within the UK rather than on the training process itself — its rejection significantly narrows that avenue for future claimants.

Could Getty Images appeal the UK High Court's ruling?

The High Court ruling is a first-instance decision, and UK court procedure generally allows for appeal to a higher court, though whether Getty pursues that option, and with what specific grounds, was not confirmed in current reporting reviewed for this analysis. Given how significant the secondary copyright infringement theory was to Getty's broader legal strategy — and how narrowly it lost, receiving only a limited trademark-related finding in place of the sweeping infringement ruling it sought — an appeal would not be a surprising next step, but businesses should treat the current ruling as the operative precedent unless and until a higher court revisits it.

Does the outcome of Getty v. Stability AI affect copyright claims in the US?

UK court rulings are not binding precedent in US courts, so Getty v. Stability AI does not directly determine the outcome of US cases like NYT v. OpenAI or Thomson Reuters v. Ross Intelligence. That said, rulings from other major jurisdictions are frequently cited as persuasive authority or discussed in legal commentary and briefing even where they carry no binding weight, and a ruling as closely watched as Getty's UK decision is likely to inform how US litigants and courts think about analogous theories, even without any formal legal obligation to follow it.

Which creative industries lobbied against the UK's proposed TDM opt-out exception?

Current reporting on the UK's abandoned text-and-data-mining opt-out exception, including analysis from firms like Linklaters and Reed Smith, points to significant pushback from creative industries broadly during the government's consultation process, reflecting concern that a default opt-out structure would place the burden on individual creators to actively exclude their own work rather than requiring AI developers to obtain permission before using it. The specific industry coalitions and their detailed positions are not exhaustively catalogued in the sources reviewed here, but the overall consultation feedback was significant enough to be widely cited as a contributing factor in the UK government's ultimate decision to abandon the proposal in its March 2026 report.

What's the likely outcome if the Third Circuit affirms Thomson Reuters v. Ross Intelligence?

If the Third Circuit affirms the district court's summary judgment for Thomson Reuters following the 11 June 2026 oral argument, it would become the first appellate-level ruling specifically addressing fair use in AI training, and it would likely be cited heavily in every other pending US case working through similar questions, including NYT v. OpenAI. An affirmance favoring Thomson Reuters would generally be read as a signal that AI training on copyrighted material faces real fair-use limits in the Third Circuit, though its precedential weight outside that circuit would depend on how persuasive other circuits and the parties in other pending cases find its reasoning.

How are AI companies changing licensing practices in response to the wave of 2026 copyright rulings?

While specific company-by-company licensing strategies are not exhaustively detailed in the sources reviewed here, the general direction implied by the 2026 litigation wave is clear: settlements like Bartz v. Anthropic have produced a concrete, citable price point for training-data rights (roughly $3,113 per book), rulings like Germany's memorization-based findings have made output-side reproduction of copyrighted material a demonstrated, exploitable legal theory, and policy positions from Australia and the UK have made clear that a broad training exemption is not coming from either government. Together, those developments push in one direction: toward AI companies increasingly needing to negotiate licensed access to training data proactively, rather than assuming unlicensed use will remain unchallenged or legally uncontested indefinitely.