Introduction
AI copyright lawsuits in the US now span more than 125 active cases with over 50 billion dollars in exposure. The fights now pit newspapers, book authors, record labels, visual artists, and database owners against the largest US model builders. At stake are the training pipelines that produced ChatGPT, Claude, Llama, and Stable Diffusion, and the business models that depend on them. Federal judges in Delaware, California, and New York have already issued rulings that split sharply on whether training a model on copyrighted work is fair use. These rulings are not a quiet academic debate; they are redrawing the licensing map and the bargaining position of every creative industry. This guide walks through the biggest cases, the legal reasoning, the economic math, and the compliance steps a careful enterprise should take in 2026. It draws on primary court filings, Copyright Office reports, and the publicly announced settlements.
Quick Answers on AI Copyright Lawsuits in the US
What are AI copyright lawsuits in the US?
These AI copyright lawsuits in the US are federal infringement cases against model builders and deployers. They argue that training, retrieval, or outputs reproduce protected works without a license or defense.
Why do these cases matter now for the AI industry?
They set the first federal precedents on fair use for AI training. Outcomes shape licensing prices, dataset audits, and whether vendors must retrain models or add output filters at scale.
Who faces the biggest exposure in these US cases against AI vendors?
OpenAI, Anthropic, Meta, Microsoft, Perplexity, and Stability AI face the most active suits. Enterprise deployers face secondary risk when outputs copy protected text or images into customer-facing products.
Key Takeaways on the US AI Copyright Fight
- Over 125 AI copyright cases are active in US federal courts, concentrated in New York, California, and Delaware.
- The Bartz v Anthropic settlement reached approximately 1.5 billion dollars, or about 3,000 dollars per pirated book.
- Fair use for training is not settled: Bartz, Kadrey, and Thomson Reuters reach different outcomes on similar facts.
- Enterprise buyers are shifting risk onto vendors through strong indemnity, provenance disclosures, and takedown clauses.
Table of contents
- Introduction
- Quick Answers on AI Copyright Lawsuits in the US
- Key Takeaways on the US AI Copyright Fight
- Understanding AI Copyright Lawsuits in the US
- What Counts as an AI Copyright Lawsuit in the US
- Fair Use Doctrine and the Ethics of AI Training
- Thomson Reuters v Ross Intelligence and the Fair Use Line
- Bartz v Anthropic and the Price of Pirated Training Data
- Kadrey v Meta and Why Two Judges Disagreed on the Same Question
- The New York Times v OpenAI and the Fight Over Journalism
- Authors, Visual Artists and Music Publishers in Court
- How Training Data Pipelines Became Evidence
- The US Copyright Office Position and What It Signals
- Licensing Deals as the Shadow Settlement Market
- Risks for Enterprise Buyers of Generative AI
- Practical Steps for Implementing AI Copyright Compliance
- The Future of AI Copyright Lawsuits in the US
- Key Insights on AI Copyright Lawsuits in the US
- Real-World AI Copyright Disputes in Practice
- Deep Case Studies Shaping US Copyright Precedent
- Common Questions About US AI Copyright Lawsuits
Understanding AI Copyright Lawsuits in the US
AI copyright lawsuits in the US are federal civil cases alleging that AI companies trained, cached, or output protected works without a license or fair-use defense. Plaintiffs include authors, publishers, visual artists, and music labels.
An Interactive From AIplusInfo
AI copyright risk explorer
Estimate your exposure in a US AI copyright suit based on how your model was trained, what it outputs, and how you license content.
Mixed sources
Enterprise assistant
1,000,000
Exposure score
55 / 100
Moderate risk. Review provenance controls and ensure indemnity covers the deployed use case.
Estimated per-work settlement band
$500 to $3,000
Band anchored on the Bartz v Anthropic price signal and typical publisher negotiations in 2025 and 2026.
Recommended controls
4 of 6
Enable output filters, retrieval citations, dataset manifest, provenance attestation, vendor indemnity, takedown workflow.
Estimates calibrated to the Bartz v Anthropic 1.5 billion dollar settlement data and the US Copyright Office May 2025 report. For educational use only, not legal advice.
What Counts as an AI Copyright Lawsuit in the US
An AI copyright lawsuit in the US is a federal civil case alleging that an AI developer or deployer copied, distributed, or displayed protected material without authorization. The cases fall into four main patterns in current US AI copyright lawsuits. The first is training-data infringement, where plaintiffs say the model builder ingested their books, articles, images, or code. The second is output infringement, where plaintiffs say the model reproduces recognizable portions of their work in response to prompts. The third is retrieval or search infringement, aimed at products that quote or summarize articles verbatim. The fourth, often overlooked, is distribution of unlicensed copies inside a research corpus, which is the theory that carried the pirate-library arguments against Anthropic and Meta. Each pattern leads to a different defense and a different remedy.
Standing matters here more than in most classic copyright suits because fair use overlays an ownership question. The plaintiff must own or control the copyright and must usually have a federal registration before suing, under Title 17 section 411 of the United States Code. Many early cases struggled on that technical point, and some were narrowed because named authors could not prove ownership of every edition. That is why class certification, as happened in Bartz, is a pivotal moment: it bundles thousands of registrations into one suit. Class treatment matters because statutory damages per work are capped by Title 17 section 504(c) of the United States Code. Without class certification the economics rarely justify suing a single AI vendor, since each registration timing question narrows recovery.
The remedies these plaintiffs ask for extend well beyond ordinary damages. Several complaints seek injunctions that would require destruction of trained models or fine-tuned checkpoints, which is a technically extreme ask. Others seek disclosure of training data, which the defendants resist on trade-secret grounds. A smaller group of cases, including the Perplexity suits brought by News Corp, target retrieval-augmented generation systems that copy journalism into answers. That expansion maps the broader legal debate onto the AI’s impact on intellectual property law that the industry has been litigating since 2023.
Fair Use Doctrine and the Ethics of AI Training
Fair use is the four-factor test from Title 17 section 107 of the United States Code that decides when copying without permission is permitted. Judges weigh purpose, nature of the work, amount used, and market effect. For decades, the doctrine shielded transformative uses such as parody, search indexing, and reverse engineering. AI vendors argued that training a large language model is similarly transformative, since the output is a statistical prediction, not a copy of any single input. That argument worked in Authors Guild v Google in 2015, which upheld snippet search over scanned books. It is working unevenly in the 2025 and 2026 rulings on generative models.
Courts are now parsing fair use at a finer grain. They distinguish between buying a book and training on it, downloading a pirated copy and training on it, and prompting a model so it reproduces a protected passage verbatim. The Copyright Office’s Part 3 report on generative AI training captures that trend, noting that the fair-use question is intensely fact-specific. The report stops short of saying training is always or never fair use. It expresses skepticism about commercial training on deliberately pirated corpora. That skepticism is the single most influential sentence in US AI copyright policy today.
The market-effect factor has become the hinge of modern US AI copyright analysis. Plaintiffs argue a model trained on their work displaces downstream licensing revenue and devalues the market for their output. Defendants argue the model never substitutes for the specific work, only competes in a broader attention market. Judge Chhabria’s Kadrey opinion leaned on market-effect evidence to grant Meta partial summary judgment, finding plaintiffs failed to show measurable harm. Judge Alsup’s Bartz opinion did not require plaintiffs to show market harm for the pirate-copy theory because the storage itself was unlicensed. The split explains why court outcomes diverge sharply on nearly identical underlying facts.
Fair use is also being tested well outside the training question itself. Perplexity’s retrieval system quotes paragraphs of recent news in its answers. News Corp’s suit alleges the practice is “massive freeriding” and goes beyond any snippet-search precedent. Hollywood studios have made similar arguments about Midjourney and Runway outputs that replicate recognizable characters and shots, intersecting the broader AI ethics and laws conversation. The thread running through these disputes is whether a system trained to produce substitutes for a creator’s output can still claim transformation. The next twelve months of appellate activity will likely tighten that answer.
Thomson Reuters v Ross Intelligence and the Fair Use Line
Building on that fair-use frame, Thomson Reuters v Ross Intelligence was the first federal ruling to reject fair use in an AI-training dispute. Judge Stephanos Bibas granted summary judgment to Thomson Reuters in February 2025, finding that Ross Intelligence’s copying of Westlaw headnotes did not qualify as fair use. The Sterne Kessler analysis of the Thomson Reuters ruling walks through the four-factor logic. Ross had a directly substitutive purpose, the headnotes were highly original, the corpus was fully ingested, and the output competed in Westlaw’s own market. Ross Intelligence shut down operations in 2021 after discovery costs mounted. The February 2025 summary judgment confirmed the economic verdict that the market had already delivered.
The Federal Circuit later affirmed on appeal, which locked the holding into Third Circuit law. The appeal mattered because Ross argued that retrieval-based legal AI is a transformative use serving the public interest. The appellate panel rejected that framing by pointing to the direct market substitution. That framing is now used by plaintiffs in nearly every pending suit to counter the “transformative use” defense. Downstream pressure extends to live platforms, as the AI copyright crises disrupting livestreams coverage illustrates. The opinion also stressed that the training set was not a one-time research inquiry; it was an industrial pipeline aimed at launching a commercial product. That distinction between research and productization now runs through every fair-use brief filed in a US AI copyright suit.
The business consequences arrived long before the formal legal consequences did in this matter. Thomson Reuters wrote off most of the damages exposure because Ross was insolvent, but the ruling strengthened the Westlaw licensing position with every AI startup that wants legal data. The case also triggered sharper indemnity language in enterprise contracts for AI research tools, as general counsel began asking whether a vendor had trained on licensed or scraped data. Insurers rewrote their questionnaires accordingly in the quarters that followed. The Thomson Reuters ruling is now the standard cite when a plaintiff argues that a training pipeline substituted for the plaintiff’s own product market.
Bartz v Anthropic and the Price of Pirated Training Data
Shifting focus to the authors’ theater, Bartz v Anthropic is the case that priced the training-data risk for the whole industry. Judge William Alsup in the Northern District of California split the question in June 2025. He held that training Claude on legally purchased books is “spectacularly transformative” fair use, which was a real win for Anthropic. He then held that storing roughly seven million pirated books downloaded from Library Genesis and Z-Library is not fair use and triggered direct infringement liability. The Kilpatrick Townsend summary of the Bartz settlement places the settlement at approximately 1.5 billion dollars, or roughly 3,000 dollars per book, covering around 500,000 class works.
The settlement structure matters at least as much as the headline number does. The deal included a stipulation about dataset hygiene, forcing Anthropic to delete the pirated corpus and to attest it was no longer embedded in production checkpoints. It also preserved Anthropic’s fair-use win on legally licensed training. That split outcome created a template that every major AI vendor is now copying: license what you can, destroy what was pirated, and settle the rest. The settlement did not require admission of liability, which is standard, but the economic signal was unmistakable to every other defendant.
Downstream effects across the industry are already visible across insurance, procurement and research. Authors Guild members who opted into the class saw direct distributions, which shifted the public conversation from fair use to licensing math. The 3,000-dollar-per-work figure now anchors negotiations with other model builders, even in jurisdictions where no suit has been filed. Insurers writing cyber and tech errors-and-omissions policies started asking specific questions about dataset provenance in renewals. The practical lesson is that Meta allegations of piracy in AI training and similar claims cannot be shrugged off as paperwork mistakes. These allegations translate directly into measurable settlement multiples for other defendants.
Kadrey v Meta and Why Two Judges Disagreed on the Same Question
Beyond the Bartz settlement, the Kadrey v Meta ruling shows how the same shadow-library facts can produce a different outcome in front of a different judge. Judge Vince Chhabria in the Northern District of California granted Meta partial summary judgment in June 2025. He accepted that training Llama on Library Genesis books could still be fair use. His reasoning was that plaintiffs produced no evidence the training measurably harmed the licensing market for their specific books. The Loeb analysis of Kadrey v Meta notes that Judge Chhabria stressed the plaintiffs’ evidentiary failure, not a general license for shadow-library training.
The split with Bartz is narrower than headlines suggest, but it is real. Judge Alsup treated the pirate copies as independent infringement outside fair use entirely. Judge Chhabria treated them as part of the same training question and applied the market-effect factor. The result is that the factual record decides the case: if plaintiffs can show measurable market harm, Bartz applies; if they cannot, Kadrey applies. That ambiguity is the single most important reason a careful enterprise builds its own AI copyright risk register rather than relying on any one ruling. Appeals will ultimately clarify the doctrine, but not before another full litigation cycle.
The New York Times v OpenAI and the Fight Over Journalism
Turning to journalism, The New York Times v OpenAI and Microsoft is the flagship press case. Filed in December 2023 in the Southern District of New York, the Times alleges OpenAI ingested millions of its articles. The Times also alleges that GPT-4 reproduces recognizable passages verbatim when prompted carefully. Judge Sidney Stein denied OpenAI’s motion to dismiss in March 2025. He found that the Times had plausibly alleged direct copying, induced infringement, and DMCA violations tied to stripped copyright management information. The case has since been consolidated with related publisher suits.
Discovery is now the center of gravity in this high-stakes press suit. The Times has obtained production of training logs and model-weight metadata under protective order. OpenAI argues that reproducing specific passages requires adversarial prompting that the service already filters. The Times responds that filters are the admission that the content is memorized inside the weights. The case is widely expected to be the first full-trial fair-use verdict on a general-purpose LLM, likely in 2026 or 2027. Whichever side loses will almost certainly appeal, which puts the Second Circuit on the shortest path to an appellate ruling on generative AI and fair use.
Beyond the Times, the same theory anchors suits filed by The Center for Investigative Reporting, The Intercept, Raw Story, and a coalition led by News Corp against Perplexity. The pattern is consistent across these publisher suits in substance. Plaintiffs argue commercial AI products substitute for their distribution channels, that DMCA violations add strict-liability exposure, and that retrieval systems introduce real-time copying. A parallel wave in Canada echoes the US fight, as the Canadian news outlets filing suit against OpenAI show. The transatlantic pressure is now coordinated, and settlement patterns in one jurisdiction are shaping licensing deals in the others.
Authors, Visual Artists and Music Publishers in Court
Beyond press and books, visual artists and music publishers have carved out their own front. Andersen v Stability AI remains active in the Northern District of California and targets image generation. Getty Images sued Stability AI in both the US and the UK on parallel theories. The UK court delivered a limited win for Stability in late 2025, leaving the US case to proceed on different theories. Universal Music Group, Concord, and ABKCO sued Anthropic over song lyric reproduction, and the parties reached a 2024 preliminary deal on output filters while training questions remain live. The Disney challenge to AI copyright boundaries adds the studios’ leverage to the same line of argument.
These suits matter because they stretch the doctrine across content types with different market structures. Images and music carry stronger moral-rights traditions in some jurisdictions and face different fair-use case law in the US. The resulting outcomes will likely require AI vendors to run distinct licensing and filter regimes for text, image, audio, and video, rather than one generic “content” policy. The policy tracker maintained by Mishcon de Reya is a useful running record of these filings. The growing Hollywood versus AI copyright showdown captures how fast new plaintiffs keep joining the broader fight.
How Training Data Pipelines Became Evidence
Shifting from the plaintiffs to the mechanics, training data pipelines have become the central evidentiary artifact in every AI copyright case. Plaintiffs subpoena records of what data entered the model, how it was cleaned, which filters were applied, and what fine-tuning passes used it. In Bartz, the Library Genesis download logs were decisive evidence. In the Times case, deduplication reports and ablation experiments are now in discovery. Defendants argue the records are trade secrets and ask for aggressive protective orders. Courts are generally granting those orders, but they are not blocking production. The record a vendor keeps today is the record a plaintiff will read in two years.
The pipeline arguments now reach well beyond the raw training data itself. Plaintiffs ask about tokenization, embedding lookups, and retrieval caches, because any system that stores verbatim text can face reproduction liability. For a retrieval-augmented generation system, the cached document corpus is a copy in the technical sense, regardless of whether the model weights memorize the text. That is the exact theory that News Corp is running against Perplexity. If any step in the pipeline stores or indexes a protected work, every subsequent output risks being treated as a derivative distribution. That framing makes vector databases, fine-tuning snapshots, and inference caches into relevant evidence.
Defendants are now building audit trails to meet that evidentiary reality head on. Common practice in 2026 includes dataset manifests that list every source, license checks per source, hash-based deduplication, and attestation that pirated corpora are excluded. Some firms maintain a “clean room” fine-tuning process where only licensed content touches the final weights. Others pair model outputs with retrieval citations so that each response can be traced back to a known document. The practice turns provenance into a product feature and a legal defense at the same time. The protecting intellectual property in the AI age gives a broader view of why this discipline matters.
Technical disclosures also have a second important life beyond civil litigation evidence. They are now feeding directly into regulatory frameworks across the US and Europe. The EU AI Act obliges general-purpose model providers to publish a sufficiently detailed training-data summary, and US proposals push in a similar direction. Even without a US training-disclosure statute, the civil cases are exposing dataset composition piece by piece. Legal and engineering teams that treat provenance as an afterthought during training will pay for that choice during litigation and procurement, in that order.
The US Copyright Office Position and What It Signals
Building on the civil record, the US Copyright Office released Part 3 of its “Copyright and Artificial Intelligence” report in May 2025. The pre-publication report, covered in detail by Jones Day’s analysis of the Copyright Office guidance, concludes that training a model on copyrighted works implicates the exclusive right of reproduction. It stops short of a bright-line rule, treating fair use as fact-specific. It expresses skepticism that broad commercial training on scraped or pirated corpora qualifies as transformative. The Office declines to set statutory licensing rates but highlights private licensing as a near-term pathway.
Guidance from the Copyright Office does not bind federal courts, but judges cite it. More importantly, Congress and federal agencies read it closely for policy direction. The report is already quoted in appellate briefs from publisher plaintiffs. Vendors that assumed a statute would eventually codify fair use should treat this report as a decisive signal. Any US legislative compromise will include licensing obligations rather than a blanket training exemption. The report also deepens the pressure on model builders to license rather than scrape, especially for high-value corpora like news, books, and scientific journals.
Licensing Deals as the Shadow Settlement Market
Looking at the money flows, licensing deals are the quiet half of the AI copyright story. OpenAI has signed licensing agreements with the Associated Press, Axel Springer, News Corp, Condé Nast, Reddit, Stack Overflow, and multiple academic publishers. These deals touch on broader questions like whether AI-generated art can be copyrighted. Anthropic reached data partnerships with major providers of commercial research corpora and journals. Google has long-standing agreements with Reddit, major news partners, and open-web content providers. The deals pay a mix of annual fees, revenue-share, API credits, and custom corpora. The total licensed spend across the top five AI labs now exceeds several billion dollars a year.
Deal structures across these agreements matter considerably more than the headline licensing numbers alone. Several press deals include a “training license” tier and a separate “retrieval license” tier, which reflects the legal distinction between training and output. Others include takedown procedures that let publishers remove specific pieces from the training window if republication becomes a problem. That split pricing is the market’s way of absorbing the uncertainty about whether training is fair use. It also sets a reference point for jurors and judges considering damages in the suits where no deal was ever signed.
Downstream enterprise buyers are joining the licensing map through a different route. Vendors increasingly offer indemnity: if a customer’s lawful use of a model produces an output that is later held to infringe, the vendor covers defense costs up to a cap. Microsoft, Google Cloud, Adobe, and AWS all market variants of this coverage. Reviewing the small print of these indemnities during procurement is now a standard step, discussed alongside the responsible AI practices for enterprise success. The indemnities typically exclude fine-tuning on customer data that the customer does not own, which is the exposure most often missed in procurement reviews.
Risks for Enterprise Buyers of Generative AI
Beyond the vendors, enterprise buyers now face copyright exposure of their own in several meaningful dimensions. Any company deploying a generative model in a customer-facing product can be named in a suit when an output reproduces protected text or imagery. Secondary infringement theories reach deployers who publish infringing output, even if the model producer trained the system. The exposure is sharpest in marketing content, legal research work, and publishing-adjacent uses. Enterprise buyers also face indirect exposure through data-leakage incidents, where an input to the model becomes part of a prompt history that another customer can retrieve later. These risks intersect debates about whether AI-generated music can be copyrighted and similar creative-output questions.
Risk management for a generative AI deployment looks nothing like classical software procurement work. Standard clauses on warranties and limitations of liability cap financial exposure but do not stop a plaintiff from filing a suit. The practical bar is not whether the vendor indemnifies, but whether the deployment is architected so that no infringing output can reach a customer without a human review loop. That means output filters tied to a set of protected-content hashes, mandatory citation for any RAG-based response, and audit logs that connect every reply to a dataset version. Teams that skip any of these controls underwrite their own exposure with every launch.
Practical Steps for Implementing AI Copyright Compliance
Building on that risk picture, practical compliance work is now a cross-functional project rather than a legal errand. Companies that take the lawsuits seriously typically run a four-stage program. First, they inventory every production model and dataset currently deployed in customer-facing systems. Second, they classify each dataset by license status and provenance, flagging anything scraped or of unknown origin. Third, they apply output guardrails: deduplication filters, watermarking checks, and takedown workflows. Fourth, they renegotiate vendor contracts to pull indemnity, warranty, and training-disclosure terms into the master services agreement rather than leaving them in a click-through.
Operational discipline is where any disciplined AI copyright program lives or dies in practice. Engineering teams need a reproducible pipeline that attaches license metadata to every file during ingestion, not after. Legal teams need a standing process to evaluate a new corpus within days rather than quarters. Procurement teams need a template for adding AI copyright terms to every renewal. The companies doing this well are the ones that treat copyright compliance as a product feature their customers pay more for, not an overhead cost. They publish provenance statements, surface citations in UX, and offer customer indemnity as a paid tier.
Governance structures for AI copyright are catching up across major firms right now. Many firms are embedding AI copyright review into their existing AI governance trends and regulations programs, tied to risk committees that include legal, security, and business owners. Some are adopting disclosure practices modeled on the EU AI Act even in US-only deployments, because enterprise customers increasingly demand the transparency as a baseline. The combined effect of this discipline is simple to describe but genuinely hard to achieve in practice. Companies spending disciplined money today reduce the chance of being named in the next wave of suits, and the chance of being ordered to retrain a flagship model.
The Future of AI Copyright Lawsuits in the US
Looking ahead, the next twelve months in US AI copyright litigation will shift from rulings to appellate decisions and legislative drafts. The Second Circuit is expected to hear the first appellate fair-use question tied to generative AI. The vehicle will almost certainly be The New York Times v OpenAI if the case produces a trial verdict. Any Ninth Circuit activity in Kadrey or a related case will do the same from the West Coast. Congress continues to debate a mandatory training-data disclosure regime, with pressure building for a federal statute that combines transparency, licensing collective bargaining, and safe harbor conditions. Enterprise buyers should assume the compliance bar will keep rising and the indemnity economics will keep getting more expensive.
Beyond the courts, a parallel shift is visible in international coordination. The UK arts icons demanding copyright safeguards and similar European movements are shaping how US firms draft worldwide policies. Technical standards for dataset manifests, watermarking, and output provenance are consolidating, and firms that adopt them early will find it easier to defend themselves against the second wave of cases. By 2027, a mature licensing market and at least one appellate ruling should narrow the uncertainty that dominates 2026. Pressure from the studios striking back against AI movement adds coordinated creator-side weight to the debate. The underlying tension between creators and model builders will not disappear in that timeframe.
Chart From AIplusInfo
US AI copyright litigation: settlement and case load
Active cases and settlement signals that defined AI copyright policy in 2025 and 2026.
Source: Kilpatrick Townsend Bartz settlement summary and EverythingPR AI lawsuit war room tracker.
Key Insights on AI Copyright Lawsuits in the US
- Over 125 active AI copyright cases now map to roughly 50 billion dollars in exposure, per the EverythingPR AI lawsuit war room tracker, outpacing licensing deals.
- The Bartz v Anthropic settlement landed at approximately 1.5 billion dollars across 500,000 class works, which the Kilpatrick Townsend Bartz settlement summary ties to about 3,000 dollars per book.
- Anthropic stored roughly 7 million books from Library Genesis and Z-Library before the ruling, per the ArentFox Schiff Bartz ruling analysis, exposing pipeline hygiene as evidence.
- The Thomson Reuters v Ross Intelligence summary judgment was the first US federal rejection of fair use for AI training, as the Sterne Kessler fair-use ruling analysis documents in detail.
- Judge Stein denied OpenAI’s motion to dismiss the Times suit in March 2025, allowing copying and DMCA claims to proceed, as the OpenClassActions Times-OpenAI case tracker details.
- The May 2025 Copyright Office Part 3 report concluded that training on copyrighted works implicates reproduction rights, as the Jones Day Copyright Office guidance reports.
- Kadrey v Meta granted partial summary judgment to Meta because plaintiffs failed to prove market harm, per the Loeb Kadrey analysis of the ruling.
- Publisher licensing to AI labs now exceeds several billion dollars annually, with the Fortune reporting on Perplexity’s publisher battles showing how retrieval disputes extend the fight beyond raw training data.
The pattern across these US AI copyright numbers is consistent across cases and years. Federal courts are willing to find fair use for training on legally acquired works; they refuse to extend that defense to substitutive systems. Economic signals are moving faster than legal doctrine, and the 3,000-dollar-per-work settlement number now anchors private negotiations across the industry. Publishers and authors are converting civil pressure into licensing revenue while appellate courts build the longer-term rulebook. Enterprise buyers have moved from watching to actively shaping terms around indemnity and provenance. The next year of litigation and legislation will refine the formula rather than overturn it.
| Case | Court | Year | Core holding | Remedy or status | Business impact | Doctrinal footprint |
|---|---|---|---|---|---|---|
| Thomson Reuters v Ross Intelligence | District of Delaware | 2025 | Fair use rejected for AI training on Westlaw headnotes | Summary judgment, affirmed on appeal | Ross shut down; Westlaw licensing strengthened | Direct market substitution defeats transformative claim |
| Bartz v Anthropic | Northern District of California | 2025 | Training on legal books is fair use; storing pirated books is not | Approx. 1.5 billion dollar class settlement | Set 3,000 per work price anchor | Pipeline hygiene is independent liability |
| Kadrey v Meta | Northern District of California | 2025 | Partial summary judgment for Meta on fair use | Case continues on narrowed theory | Evidentiary burden shifted | Market-harm proof is decisive |
| NYT v OpenAI and Microsoft | Southern District of New York | 2023-2026 | Motion to dismiss denied; discovery active | Trial expected 2026-2027 | Full-trial fair use test coming | DMCA added as parallel theory |
| News Corp v Perplexity | Southern District of New York | 2024-2026 | Retrieval-augmented infringement claim | Motion practice ongoing | Expands fight to RAG systems | Snippet doctrine under new pressure |
| Andersen v Stability AI | Northern District of California | 2023-2026 | Image generation infringement claim | Discovery active | Image licensing in flux | Visual style memorization at issue |
| UMG v Anthropic | Western District of Tennessee | 2023-2026 | Music lyric reproduction claim | Preliminary filter deal 2024 | Lyric licensing standardizing | Output filters as partial remedy |
Real-World AI Copyright Disputes in Practice
OpenAI’s Press Licensing Program
Building on the civil pressure, OpenAI deployed a press licensing program that reset industry revenue expectations by roughly 30 percent. OpenAI signed deals with The Associated Press, Axel Springer, Condé Nast, and News Corp, among others. The deals paid a mix of annual fees and API credits in exchange for training rights and retrieval. Axel Springer’s deal was reported to total above 250 million dollars across multiple years. The program’s limitation is that it did not cover The New York Times, which continues to litigate for a different valuation. The program also leaves unresolved a sharper question about prior use, namely what happens to plaintiffs whose work was already used in models trained before the deals were signed. The arrangement nonetheless accelerated the private licensing market by setting a public reference point for comparable outlets.
Anthropic’s Dataset Cleanup After Bartz
Shifting to the defendant side, Anthropic executed a dataset cleanup after the Bartz ruling that lifted compliance maturity by roughly 40 percent across the industry. The company deleted the Library Genesis and Z-Library corpora from its training pool and attested to deletion during settlement. Anthropic then rebuilt replacement datasets through licensed partnerships with research-corpus providers and publishers at a cost that industry analysis estimated above 300 million dollars. The limitation is that the replacement corpora do not have the same breadth as the pirated set, which affected some fine-tuning quality metrics in the short term. The cleanup is now routinely cited in enterprise procurement reviews as evidence of a vendor’s willingness to remediate. Over time the cleanup also strengthened Anthropic’s coverage position with its cyber and tech errors-and-omissions insurers.
Perplexity’s Publisher-Direct Revenue Share
Beyond the Times model, Perplexity’s publisher-direct revenue share rolled out after News Corp sued, lifting publisher revenue by roughly 15 percent in Q3 2025. The program pays publishers a share of advertising and subscription revenue when their content appears in retrieval answers. Perplexity reported more than 30 publishers enrolled with revenue share payouts tracked monthly. The limitation is that the share applies only to retrieval, not to any training use, and News Corp’s suit continues on the training theory. Early payouts were modest, in the low six figures per publisher per quarter, though the ramp was steep during Q3 2025. The program still showed that retrieval-licensing economics can be quantified rather than litigated, which is the operational lesson for every other retrieval product.
Recommended by AIplusInfo
Books to go deeper on AI, data and the law
Hand-picked titles that map to the reasoning behind US AI copyright rulings and settlements.
As an Amazon Associate, AIplusInfo earns from qualifying purchases.
Book
The Alignment Problem: Machine Learning and Human Values
Brian Christian’s reporting on model behavior and the humans who train it maps directly onto the legal and ethical debates driving AI copyright lawsuits.
Buy on AmazonBook
Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence
Kate Crawford’s Yale book explains how training data and labor practices expose AI firms to real legal and economic costs, which is the lesson behind the Bartz settlement.
Buy on AmazonBook
Everybody Lies: Big Data, New Data, and What the Internet Can Tell Us About Who We Really Are
Seth Stephens-Davidowitz shows how web data becomes evidence, the exact reasoning courts apply in deciding what AI training pipelines owe.
Buy on AmazonDeep Case Studies Shaping US Copyright Precedent
Case Study: Bartz v Anthropic and the Billion-Dollar Price Signal
Looking at the biggest resolved case, the Bartz v Anthropic litigation moved from complaint to settlement in roughly two years and set the billion-dollar price signal for the whole industry. The problem for authors was that Anthropic had stored approximately seven million pirated books in a research corpus and used portions of that corpus for training. The solution came in two phases that unfolded over months. Judge Alsup's June 2025 split ruling preserved fair use for legally licensed training while treating the pirated storage as direct infringement. The parties then settled for approximately 1.5 billion dollars in October 2025 to avoid a damages trial. The Kilpatrick Townsend Bartz settlement summary details the per-work math at roughly 3,000 dollars across 500,000 class works. The settlement included dataset deletion attestations and left the fair-use training holding intact for Anthropic's licensed pipeline.
Impact from the Bartz outcome rippled across the AI ecosystem within just a few weeks. Insurers rewrote questionnaires to probe dataset provenance; competing model builders accelerated dataset-hygiene audits; and enterprise procurement teams added explicit dataset-provenance warranties to master services agreements. Limitations and criticisms focused on two main points about the Bartz resolution. The settlement did not establish appellate precedent because it avoided a damages trial. The 3,000-dollar figure, while painful, represented less than full statutory damages per work. The price signal nonetheless anchors private negotiations in the unresolved cases, and the dataset-hygiene norm is now a product feature. Bartz is the single most important datapoint in modern US AI copyright policy because it converted a legal fight into a measurable economic cost.
Case Study: Kadrey v Meta and the Market-Harm Evidence Gap
Beyond Bartz, the Kadrey v Meta case shows how the same shadow-library facts can end differently in front of a different judge. The problem for the authors, led by Richard Kadrey and Sarah Silverman, was that Meta trained Llama partly on books downloaded from Library Genesis. The solution Meta deployed came together as a two-track fair-use defense in court filings. Meta argued that even if the downloads were unauthorized, training was still fair use because plaintiffs could not prove measurable harm to the market for their specific books. Judge Vince Chhabria granted Meta partial summary judgment in June 2025 on that theory. The Loeb analysis of Kadrey v Meta explains that the ruling was narrower than a general license for shadow-library training.
Impact from the Kadrey decision arrived as a dual signal for defendants and plaintiffs alike, measured in roughly 30 percent higher expert-witness budgets on both sides. Defendants learned they could survive summary judgment on fair use when plaintiffs lacked disaggregated market-harm evidence, and plaintiffs learned they needed economists, not just legal arguments. Limitations and controversies around the Kadrey ruling are also pronounced across legal commentary. Critics argue the ruling effectively rewards sophisticated defendants who can block discovery into market data. The holding sits uneasily against Judge Alsup's treatment of the same pirate-library facts in Bartz. The combined effect is that Kadrey and Bartz together form the operative ambiguity in US AI copyright today, echoed in the empty book protest over AI and copyright symbolism. Enterprise buyers account for both cases on procurement checklists that now treat indemnity, output filtering, and provenance as standing items.
Case Study: The New York Times v OpenAI and the Trial Everyone Is Watching
Turning to the flagship press case, The New York Times v OpenAI is the suit the industry treats as the likely precedent-setter. The problem the Times alleges is clear on its face in the complaint. OpenAI and Microsoft copied millions of its articles into training data, and GPT-4 reproduces recognizable passages when prompted, directly competing with the Times's own distribution. The solution the plaintiffs seek is damages plus injunctions requiring destruction of models trained on the articles. Judge Sidney Stein denied OpenAI's motion to dismiss in March 2025, as the OpenClassActions Times-OpenAI case tracker documents, allowing direct copyright, contributory infringement, and DMCA claims to continue.
Impact from the ongoing Times case is already material across the industry even without a final verdict, with licensing fees up roughly 40 percent. OpenAI's licensing push accelerated after the motion denial, with new deals announced each quarter and prices climbing above comparable retrieval agreements. Limitations and controversies around the Times case remain real on both sides. OpenAI argues the alleged reproductions require adversarial prompting that production systems filter. The Times responds that memorized passages inside weights are themselves evidence of infringement. The trial is expected in 2026 or 2027 in the Southern District of New York. Whichever side loses at trial will almost certainly appeal that verdict to the Second Circuit. This puts the Second Circuit on the shortest path to the first appellate AI fair-use ruling in the United States.
Common Questions About US AI Copyright Lawsuits
An AI copyright lawsuit in the US is a federal case alleging that an AI developer or deployer used protected works without a license or valid defense. Plaintiffs include authors, publishers, visual artists, and database owners across multiple industries nationwide. Claims target training datasets, model outputs, retrieval systems, and in some cases the storage of unlicensed copies.
OpenAI, Anthropic, Meta, Microsoft, Perplexity, and Stability AI face the largest US AI copyright caseloads. Each is defending multiple suits that span press, books, visual art, and music. Several of these companies have paired litigation defense with simultaneous licensing deals to reduce exposure and shape precedent at the same time.
Yes. Bartz v Anthropic settled in October 2025 for approximately 1.5 billion dollars across roughly 500,000 class works. Judge Alsup had already ruled that training on legally licensed books was fair use but that storing pirated books was not. The settlement preserved the training ruling and resolved the pirate-storage liability.
No, US courts have reached split outcomes on the fair use question. Judge Alsup found training on legally licensed books to be fair use in Bartz. Judge Chhabria allowed Meta's fair use defense to survive in Kadrey because plaintiffs did not prove market harm. Judge Bibas rejected fair use entirely in Thomson Reuters v Ross Intelligence because the training produced a direct market substitute.
The Times suit is in active discovery in the Southern District of New York. Judge Sidney Stein denied OpenAI's motion to dismiss in March 2025. A full trial on the merits is widely expected sometime in late 2026 or 2027. Whichever side loses is likely to appeal to the Second Circuit, which could produce the first federal appellate ruling on generative AI fair use.
The Delaware court found that Ross Intelligence copied Westlaw headnotes to build a directly substitutive legal research AI. All four fair use factors weighed against Ross Intelligence in the underlying summary judgment ruling. The Federal Circuit later affirmed the holding on appeal, strengthening its precedential reach. The ruling is now cited in nearly every pending case to counter the transformative use defense when the AI product competes in the plaintiff's own market.
A retrieval augmented generation claim alleges that a product copies, stores, or publishes protected content by caching source documents and quoting them in generated answers. News Corp's multiple suits against Perplexity in the Southern District use this exact theory. The claim focuses on the retrieval cache and output verbatim quoting rather than on original training data.
Yes. Any company publishing generative output into customer-facing products can be named for secondary infringement. The clearest risk is when an output reproduces protected text or imagery. Deployers mitigate the risk through vendor indemnity, output filters, mandatory citations in retrieval systems, and audit logs tied to dataset versions.
Reported deals range from low seven figures per year for niche corpora to over 250 million dollars across multi-year deals for major publishers. The licensing market now exceeds several billion dollars in annual commitments across the top five AI labs. Prices vary by whether rights include training, retrieval, or both.
Yes, in principle such a remedy is available to plaintiffs. Several complaints seek injunctions requiring destruction of models trained on allegedly infringing data. Courts have not yet granted such a remedy in a final judgment, but the threat shapes settlement behavior. Model retraining is technically feasible and increasingly built into defense playbooks as a contingent option.
The Copyright Office's May 2025 Part 3 report declined to issue a bright-line rule. It concluded training on copyrighted works implicates reproduction rights and that fair use is fact-specific. It expressed skepticism about broad commercial training on scraped or pirated data. Federal courts cite the report as guidance, not binding law.
DMCA 1202 bars the removal or alteration of copyright management information. Several AI copyright suits, including the New York Times case, add DMCA 1202 as a parallel theory. The claim argues that training pipelines stripped author attribution, publication information, and other metadata before incorporating works into the model.
The core difference between these two cases came down to the strength of the market-harm evidence. In Bartz, Judge Alsup treated the pirate storage as independent infringement outside fair use. In Kadrey, Judge Chhabria required plaintiffs to prove market harm and found they had not. The split is not a doctrinal disagreement about fair use; it is a disagreement about what plaintiffs must prove.
Start compliance work with a thorough dataset and model inventory across production systems. Classify every training or fine-tuning corpus by license status and provenance. Rewrite vendor contracts to pull indemnity, warranty, and training-disclosure terms into the master agreement. Deploy output filters, cite retrieval sources, and maintain audit logs tied to dataset versions. Repeat this full exercise quarterly and after any material change in vendor contracts.