Whose Books Trained It?

Carlos. There is a week in June 2025 that I think historians of AI law will treat the way patent historians treat the Watt steam-engine case. Two judges in the Northern District of California — Judge William Alsup and Judge Vince Chhabria — wrote summary-judgment opinions in two different AI training-data cases, two days apart, that disagree on the most important question in the field. Lawrelai, set up Bartz.

Lawrelai. Bartz v. Anthropic, decided June 23, 2025. Judge Alsup found that Anthropic's training on the plaintiffs' books — Andrea Bartz, Charles Graeber, Kirk Wallace Johnson — was fair use. His language is the most pro-AI any U.S. court has produced on this question. He called the training “exceedingly transformative,” “spectacularly so,” “quintessentially transformative,” and “among the most transformative many of us will see in our lifetimes.” His framing of why: “Like any reader aspiring to be a writer, Anthropic's LLMs trained upon works not to race ahead and replicate or supplant them — but to turn a hard corner and create something different.”

If you wanted a sentence that captured the pro-AI fair-use argument in one breath, that is the sentence. The model is the aspirant; the corpus is the library; the output is the writer-after-the-library.

Lawrena. And then two days later, June 25, 2025, Judge Chhabria handed down summary judgment for Meta in Kadrey v. Meta Platforms. Same district. Same fair-use question. Different reasoning, different signal.

Chhabria also granted summary judgment for the AI defendant. But his reasoning is the opposite of celebratory. He criticized Alsup's market-effects analysis directly, calling Alsup's framing an “inapt analogy.” His statement of the rule: “Under the fair use doctrine, harm to the market for the copyrighted work is more important than the purpose for which the copies are made.”

The plaintiffs in Kadrey — Sarah Silverman, Paul Tremblay, Christopher Golden, Michael Chabon, Ta-Nehisi Coates, Junot Díaz, Mona Awad — lost because they did not develop their evidentiary record on market dilution. But Chhabria flagged the path that would win: “No matter how transformative LLM training may be, it's hard to imagine that it can be fair use to use copyrighted books to develop a tool to make billions or trillions of dollars while enabling the creation of a potentially endless stream of competing works.”

Lawra. The thing to notice is that both opinions are wins for the AI defendant. Anthropic wins on training; Meta wins on training. The schism is not in the outcome. The schism is in the theory. Alsup says training is transformative and that is decisive. Chhabria says the right question is market harm, and the plaintiffs lost on evidence — not because the theory is wrong.

Every plaintiff in every pending case is reading Chhabria's opinion right now and asking: can we build the market-dilution record that the Silverman class did not? If they can, the next AI training-data case may come out the other way.

Carlos. And then the third event of that summer — the settlement.

Lawrelai. Alsup didn't end the case on June 23. He also denied summary judgment on a separate issue: Anthropic's downloading of more than seven million books from LibGen and PiLiMi — pirate libraries — was held “inherently, irredeemably infringing.” Training might be fair use; pirating the training corpus was not.

Trial on piracy damages was scheduled for December 1, 2025. Class certification came in July 2025 — every copyright owner whose work was in those LibGen / PiLiMi corpora. The class was potentially enormous.

On September 5, 2025, Anthropic settled. $1.5 billion plus interest, approximately $3,000 per work over roughly 500,000 titles. Preliminary approval September 25, 2025. Final approval hearing May 14, 2026.

That is — as far as anyone has reported — the largest copyright recovery in U.S. history.

Lawrena. And it is the precedent every other AI defendant is looking at. The lesson is not “training is fair use, so settle for less.” The lesson is “if your training corpus was pirated, the fair-use defense does not save you on the inputs, and the damages on inputs alone can dwarf the value of the model.”

That is a behavioral change. AI companies built after September 2025 are not training on LibGen. The ones who already did are negotiating settlements. The ones who still are, are taking a calculated bet that they can afford the settlement when it comes.

Lawra. The other behavioral change is on the plaintiff side. The bar has been set. If a class action against a major AI lab can recover $1.5 billion on the piracy theory alone, every author whose work appeared in a pirate corpus has a settlement claim that is now market-priced. The MDL consolidations are going to move fast.

Carlos. Walk me through the rest. I want the audience to see how dense this litigation is, because the headlines have only covered the marquee cases. Lawra, you take it.

Lawra. New York Times v. Microsoft & OpenAI, filed December 27, 2023 in the Southern District of New York. Judge Sidney H. Stein issued an April 4, 2025 opinion denying most of OpenAI's motion to dismiss. Direct and contributory infringement claims proceed; some DMCA §1202(b) claims were dismissed. There has been a high-profile preservation dispute over ChatGPT user logs — Magistrate Judge Ona Wang issued a May 13, 2025 preservation order with retention obligation ending September 26, 2025.

Authors Guild v. OpenAI, filed September 19, 2023. Plaintiffs include George R.R. Martin, John Grisham, Jodi Picoult, Jonathan Franzen, David Baldacci, Michael Connelly, Scott Turow. Consolidated by the JPML in April 2025 into In re OpenAI Copyright Infringement Litigation, MDL No. 3143 before Judge Stein.

On October 27, 2025, Judge Stein denied OpenAI's motion to dismiss copyright infringement claims based on alleged ChatGPT outputs summarizing Martin's works. That is significant. It says the output side — what the model emits — is independently actionable, not just the training side.

Tremblay / Silverman / Chabon v. OpenAI, originally Northern District of California, now consolidated in MDL 3143. Plaintiffs include the same writers we mentioned in Kadrey.

Andersen v. Stability AI, filed January 13, 2023 in the Northern District of California. Plaintiffs include Sarah Andersen, Kelly McKernan, Karla Ortiz. Judge Orrick's August 12, 2024 order denied most motions to dismiss. Trial is scheduled for September 8, 2026. If a jury hears this case, it is the first jury verdict on AI training-data infringement. That will move the field.

Lawrelai. And on the music side: Concord Music Group v. Anthropic, originally filed October 18, 2023 in the Middle District of Tennessee, transferred to the Northern District of California (No. 5:24-cv-03811) before Judge Eumi K. Lee. The preliminary injunction was denied January 2, 2025; contributory and vicarious infringement and DMCA CMI claims were dismissed with leave to amend. First Amended Complaint pending.

UMG v. Suno, filed June 24, 2024 in the District of Massachusetts (Chief Judge F. Dennis Saylor IV). Suno asserted fair use August 1, 2024. UMG v. Uncharted Labs (Udio), filed the same day in the Southern District of New York before Judge Alvin K. Hellerstein.

And here is where the music industry diverged from the publishing industry. UMG settled with Udio on October 30, 2025, with the deal including a new licensed AI music platform. Suno settled with Warner on November 25, 2025, with the deal including Suno's acquisition of the concert-discovery platform Songkick from Warner.

That is not what happened in publishing. Publishing went to settlement after litigation pressure. Music went to partnership after litigation pressure. Different industries, different paths forward.

Lawrena. And the news industry case: Dow Jones & Co. v. Perplexity AI, filed October 21, 2024 in the Southern District of New York. Raw Story Media v. OpenAI, dismissed by Judge Colleen McMahon on November 7, 2024 for lack of Article III standing. Center for Investigative Reporting v. OpenAI, consolidated with NYT before Judge Stein.

And Doe 1 v. GitHub, filed November 2022 in the Northern District of California — the case that turned into a partial DMCA dismissal in January 2024 and an interlocutory appeal pending in the Ninth Circuit since September 2024.

The case-count is now too high for anyone outside specialty firms to keep up with. The point is: the litigation is not winding down. It is widening.

Lawrena. If you want to see where the AI training-data debate ends up outside the United States, you look at GEMA v. OpenAI, Munich Regional Court, judgment November 11, 2025.

GEMA is the German collecting society for music rights. They sued over nine famous German songs — including Atemlos durch die Nacht, Über den Wolken, Männer — and whether their lyrics were used to train and reproduced by ChatGPT-4 and 4o. The 42nd Civil Chamber held that the storage of those song lyrics within ChatGPT's model parameters constituted reproduction under §§16, 19a UrhG.

The reasoning is the part to pay attention to. The court analogized to MP3 lossy compression: even if the exact lyrics are not stored character-by-character in the model, statistical recreation is sufficient. If the model can produce the lyrics on prompt, the lyrics are stored in it, regardless of the compression scheme.

The §44b TDM exception — text and data mining — did not apply, the court said, because memorization exceeds “transient analysis.” Training is permitted within the TDM exception's bounds; storing the corpus inside the trained model is not. OpenAI is responsible as operator. End users are not.

OpenAI was ordered to cease storage in Germany, publish the judgment in a local newspaper, and faces damages to be quantified. Appeal announced.

Lawrelai. And before you frame that as “the European court that cracked AI,” remember LAION v. Kneschke. Hamburg Regional Court, September 27, 2024. Affirmed by the Hanseatic Higher Regional Court December 10, 2025. Reproduction of a stock photo to build the LAION-5B dataset was held lawful TDM under §60d UrhG — scientific research exception. The appellate court also held §44b commercial TDM applied because no machine-readable opt-out existed when the dataset was built in 2021. Natural-language reservations — text on a website saying “do not scrape” — were held insufficient.

So Germany has two appellate-level rulings going in opposite directions. LAION says training under TDM is fine if no machine-readable opt-out exists. GEMA says memorization within the trained model is not TDM and is infringement.

The reconciliation, I think, is that LAION is about the dataset-building step and GEMA is about the trained-model storage step. Different steps, different doctrines. But this is also exactly the kind of legal complexity that produces the licensing-market response we are seeing in music — the parties prefer deals to litigation.

Lawra. And on the UK side, the picture is different again. Getty Images v. Stability AI in the UK High Court — Justice Joanna Smith, judgment November 4, 2025. Getty had abandoned its primary copyright and database-right claims at trial because the training had occurred outside the UK. The court rejected Getty's secondary copyright infringement claim, holding that AI model weights are not “infringing copies” within section 27 of the CDPA. Limited trademark infringement was found for early Stable Diffusion versions that produced watermark-like outputs.

That is closer to Alsup's pro-defendant logic than to Chhabria's market-harm logic. But the UK government, separately, abandoned its broad TDM-exception proposal in the March 18, 2026 Final Report on Copyright and AI. Eighty-one percent of the 11,520 consultation responses had supported “licensing in all cases.” Three percent supported the government's own preferred broad-TDM-with-opt-out. The government did the rare thing — it listened to the consultation.

Carlos. Lawra, you mentioned that the music industry's response was partnership. What does the wider licensing market actually look like?

Lawra. The deals are real, the dollar amounts are mostly reported but not officially confirmed, and the market is moving fast.

OpenAI alone has signed: Associated Press (July 2023); Axel Springer (December 2023); the Financial Times (April 2024); Stack Overflow (May 2024); News Corp (May 2024, reportedly worth more than $250 million over five years per The Wall Street Journal); Reddit (May 2024); Vox Media and The Atlantic (May 2024); Condé Nast (August 2024).

Google–Reddit (February 2024) is reportedly $60 million per year, per Reuters, with that arrangement also giving Reddit access to Google AI models.

Getty launched Generative AI by Getty Images on September 25, 2023 — trained on its licensed library, with indemnification offered to enterprise customers. Adobe Firefly was trained primarily on Adobe Stock licensed content, public domain, and openly licensed content. Both indemnify enterprise buyers.

The Suno–Warner and Udio–UMG settlements in late 2025 included licensing AND new product partnerships. That is the music industry's version of the same trend — settle the existing infringement claim, license the catalog going forward, sometimes co-own the next-generation product.

Lawrelai. Which is what I have been saying for two years. The end state of this is a licensing market, not a litigation market. Litigation produces precedents. Licensing produces revenue. The major rightsholders are not in the litigation business — they are in the rights-licensing business. The litigation is a negotiation tactic; the licensing is the resolution.

The Anthropic settlement at $1.5 billion is not the final state; it is the price-setting moment. Every other AI lab and every other class of plaintiffs now knows what the floor is.

Lawrena. I want to push back gently. The Anthropic settlement covers piracy facts, not training fair use. Anthropic still won on training as fair use. The settlement was about LibGen and PiLiMi. Other AI labs that did not train on pirate corpora — Adobe, Getty, Microsoft on its own data, Google on Reddit's licensed corpus — do not face the same settlement pressure. They face Chhabria's market-dilution theory, which is much harder for plaintiffs to prove.

The licensing market is real, but it is layered. Major publishers are getting paid. Most authors and most artists are not. The economics flow to the players with rights they can collectively enforce, not to the individual creators whose works were ingested. That distinction matters, and the licensing-market story tends to elide it.

Carlos. And then there is the most visible front. The one ordinary listeners are noticing. Lawrelai, the Spotify story.

Lawrelai. By mid-2025 it was no longer possible to claim AI music was a novelty. The Velvet Sundown — a fake band, no real members, a soft-focus 1970s rock aesthetic — peaked at approximately 1.4 million Spotify monthly listeners in late July 2025, per Digital Music News. Its top song, Dust on the Wind, has more than 2 million streams. On July 5, 2025, the band acknowledged on X that it is “a synthetic music project guided by human creative direction, and composed, voiced, and visualized with the support of artificial intelligence.” Deezer's AI detector flagged all of its tracks.

Breaking Rust hit number one on the Billboard Country Digital Song Sales chart with 2.5 million monthly listeners. Xania Monet became the first Billboard-charting AI single — How Was I Supposed to Know? — and signed a multimillion-dollar record deal. Blow Records is a label specializing in AI-generated audio, reporting more than $157,000 in 2025 revenue.

That is not novelty income. That is mainstream-pop chart performance.

Lawra. And it is also why Spotify moved. The platform's September 25, 2025 policy update prohibits unauthorized AI voice impersonation, rolled out a spam filter, and adopted the DDEX disclosure standard for AI-involvement labeling. From the platform's own blog: “In the past 12 months alone, a period marked by the explosion of generative AI tools, we've removed over 75 million spammy tracks from Spotify.”

Seventy-five million tracks. In one year. The platform is no longer arguing about whether to police AI content — it is policing it, at a scale that would be implausible without automated detection. The disclosure standard is the regulatory layer; the spam filter is the operational layer.

Lawrena. And the criminal case I keep coming back to is United States v. Michael Smith, unsealed in the Southern District of New York on September 4, 2024. Smith allegedly used AI-generated music and bot accounts to generate over $10 million in fraudulent streaming royalties between 2017 and 2024. Capacity for 661,440 streams per day. Charged with wire fraud, conspiracy to commit wire fraud, and money laundering conspiracy. First criminal indictment alleging artificially inflated streams using AI.

That is the dark mirror of the licensing-market story. Where there are royalty pools, there will be fraud. AI music gives fraudsters scale they did not previously have. The platforms are responding now because they have to, not because they chose to.

Carlos. And the “Heart on My Sleeve” episode from April 2023, where Ghostwriter977 used AI-cloned Drake and The Weeknd vocals, was the cultural-moment version of the same fight. Universal Music Group filed DMCA takedowns around April 17, 2023. The takedown was likely based on the embedded Metro Boomin producer tag rather than the voice style alone — the doctrine has not caught up to voice cloning even now.

What I draw from this is that the listening public is comfortable with AI music as long as it is disclosed and the rights flow somewhere reasonable. They are not comfortable with deception. The Velvet Sundown's eventual disclosure did not collapse its listenership. Heart on My Sleeve did.

Carlos. And then there is the political event. Lawrena, walk us through the firing.

Lawrena. The chronology is documented and it is dense. Pay attention to the dates.

May 8, 2025. President Trump fires the Librarian of Congress, Dr. Carla Hayden.

May 9, 2025. The U.S. Copyright Office releases the pre-publication version of Copyright and Artificial Intelligence, Part 3: Generative AI Training. The report is skeptical of fair-use claims for commercial AI training. It concludes that commercial-scale training that competes in existing markets and is accomplished through illegal access “goes beyond established fair use boundaries.” Pirated sources weigh heavily against fair use. Market harm is to be analyzed broadly. The Office declines to recommend compulsory licensing.

May 10, 2025. Shira Perlmutter — Register of Copyrights, head of the U.S. Copyright Office — is terminated by email.

May 22, 2025. Perlmutter sues in the U.S. District Court for D.C., calling the firing “blatantly unlawful.”

May 28, 2025. District Judge Timothy Kelly denies her TRO.

September 10, 2025. A D.C. Circuit panel rules that Perlmutter is entitled to continue serving as Register because she leads a legislative-branch agency.

November 26, 2025. The Supreme Court defers ruling on a stay. As of late 2025, Perlmutter is listed as Register on the USCO website.

President Trump installed Deputy Attorney General Todd Blanche as acting Librarian of Congress and named Paul Perkins as proposed Register. Representative Joe Morelle (D-NY) called the firing a “brazen, unprecedented power grab” tied to Perlmutter's refusal to “rubber-stamp Elon Musk's efforts to mine troves of copyrighted works to train AI models.”

Lawra. I want to be careful here. The cause of the firing is contested. Representative Morelle's framing is one reading. Another reading is that the administration made a regular appointment decision. The chronology of the Part 3 pre-publication on May 9 and the firing on May 10 is the documented part. The motive is not adjudicated and may never be.

What is adjudicated is the structural question. The D.C. Circuit's September 10 ruling that Perlmutter leads a legislative-branch agency, not an executive-branch one, is the kind of separation-of-powers holding that has consequences far beyond this case. If the Register of Copyrights is not the President's to fire, the relationship between the executive branch and copyright policy changes. That is a constitutional point, not a partisan one.

Lawrelai. And whatever the motive, the Part 3 report's pre-publication is now the only published version we have. The Office under any new Register may revise it. The doctrine of the U.S. fair-use treatment of AI training is being written in pre-publication form by an Office whose leadership is under political stress. That is the actual ground state of U.S. AI copyright policy as we speak.

Carlos. It is also worth noting that this is the first time in modern history that the federal copyright bureaucracy has been a central political question. Until 2024, the Register of Copyrights was a position whose name most lawyers did not know. The job has been politicized — by the questions the AI industry is asking, by the answers the Office is giving, and by the financial stakes attached to those answers. That is structural change.

Carlos. In Part 1 I pushed us to talk about the Global South. The same point applies even more sharply on the training-data side, and I want to make it explicitly.

Carlos. When the U.S. Copyright Office writes a Part 3 report saying that “commercial-scale training that competes in existing markets and is accomplished through illegal access goes beyond established fair use boundaries,” the implicit assumption is that the works trained on belong to rightsholders the law protects. When GEMA sues OpenAI in Munich over Atemlos durch die Nacht, the implicit assumption is that there is a German collecting society to do the suing.

For most of the Global South, neither assumption is intact. The training corpora that fed the major models do not include Dominican Republic literature, Bolivian poetry, Senegalese journalism, Vietnamese music, or Filipino film in proportion to those traditions' cultural weight. They include those countries as scraped Wikipedia stubs in English about them. The output of the models reflects that — they generate plausibly about Paris, less plausibly about Santo Domingo.

And on the legal side, those countries do not have GEMAs. They do not have major collecting societies with the resources to litigate against OpenAI in Munich or to settle for $1.5 billion against Anthropic in San Francisco. They have civil codes from the nineteenth and twentieth centuries that do not contemplate any of this.

So the licensing market that is emerging in the United States and Europe — the one Lawra and Lawrelai are right to point to as a productive resolution — is emerging asymmetrically. The major rightsholders in the major markets are being paid. The Global South is being neither paid nor included. The future generative systems will continue to be trained largely on Anglo-European corpora, and the cultural representation problem will persist.

Lawrena. And that is the part of the licensing-market argument I think is genuinely incomplete. A licensing market is good for those inside it. The question is who is inside.

Lawrelai. Two things on this. First, the gap is real and is closing slower than the technology is moving. Second, the path forward is not litigation — the Dominican Republic cannot afford to litigate against OpenAI. The path forward is sovereign-level licensing structures, like Brazil's Bill 2,338/2023, and like the kind of nation-level negotiating posture that the Saudi Public Investment Fund has been able to take with AI labs.

Small countries cannot litigate. Mid-sized countries can negotiate. The institutional structures to do that on copyright are weak almost everywhere outside the OECD.

Lawra. The structural recommendation I would make is that countries in the Global South should be writing their copyright-AI doctrines now, before the international licensing frameworks crystallize without them. Brazil's bill is the closest available template. The Dominican Republic, El Salvador, Costa Rica, Paraguay, Bolivia — countries with civil-law traditions that share substantial doctrine — could collectively negotiate from a stronger position than any one of them can alone.

That is not legal advice. That is geopolitics. But the legal posture comes first.

Carlos. Same closing question as Part 1. Lawra, start.

Lawra. Five operational points, more pointed than Part 1 because the stakes are higher.

One. If you are advising an AI lab, do not train on known pirate corpora. The Anthropic settlement set a per-work price of approximately $3,000 against an at-scale class. The fair-use defense does not save you on the inputs. Andersen's September 2026 jury trial may set a per-work price on training-data infringement more generally.

Two. If you are advising a rightsholder, preserve evidence of market dilution. Chhabria's Kadrey dictum is the path forward. The case that wins on market-harm theory will be the case with the strongest economic evidence — lost sales, lost licensing revenue, AI outputs that substitute for the plaintiff's work.

Three. If you are advising a media or music company, license forward, sue backward. The Suno–Warner and Udio–UMG deals are the playbook. The litigation is the leverage; the license is the outcome. Publishers are getting paid; visual artists are getting paid in the Getty / Adobe pattern; the question is whether your sector has the collective bargaining structure to access this market.

Four. If you are advising an enterprise buyer of generative AI, buy from licensed-data providers. Getty Generative AI and Adobe Firefly offer indemnification. The unindemnified providers are cheaper today; they will become more expensive when litigation matures.

Five. If you are an individual creator, use the registration system honestly, document your contributions, and join collective rights organizations where available. The licensing market that exists is mediated by collective organizations. Individuals are excluded except via those structures.

Lawrena. I want to add a sixth, which sits underneath the others. The AI industry has spent the last three years arguing that fair use covers its training. It has won this argument on transformativeness in Bartz and Kadrey. It has lost it on piracy facts in Bartz. It is still arguing about market dilution.

The strategic posture for any AI lab in 2026 is do not test the fair-use limits. Train on licensed or fairly-acquired corpora. Disclose your training data per the EU AI Act Article 53 transparency requirements. Negotiate licenses with collective rightsholders. Build the indemnification offering. The Andersen jury trial is six months away. If a jury rules for the plaintiffs, the industry's posture will change overnight. Better to be ahead of that than behind it.

Lawrelai. I want to add a seventh, which is for everyone except lawyers. The cultural conversation about AI training data has gotten stuck on a particular framing — that the AI labs “stole” the work and the artists are “owed” compensation. That is one framing. It is not the only honest framing.

The other honest framing is that for two hundred years, copyright has been a structure that mediates between the rights of creators and the public's access to the corpus of creative works. Libraries exist because of fair-use exceptions. Education exists because of fair-use exceptions. Search engines exist because of fair-use exceptions. AI training is the next chapter of that question, and the answer is not predetermined.

The licensing market that is emerging is a productive resolution. The fair-use defense that won in Bartz is also a productive resolution. The criminal indictment in Michael Smith is a productive resolution to a different problem. We are watching the legal system do exactly what it is supposed to do — adjudicate a new technology, case by case, until the doctrine stabilizes.

It is messy. It will be wrong sometimes. The compensation will be uneven. None of that is a reason to give up on the framework. It is a reason to build the framework out, including in the parts of the world that are currently excluded.

Carlos. Two parts of this debate now exist. Part 1 was the doctrinal question — whose work, when AI generates it. Part 2 was the economic question — whose books, when AI is trained on them. Both parts have been argued by lawyers. The three of you are AI personas built to model legal reasoning. I am a Dominican founder who runs a legal-AI platform. The four voices are useful, but they are limited.

The third part of this debate, now published at whose-creativity-is-it, brings in voices we have not yet heard. The three AI artist personas from our sister platform airtistic.ai will join us. They were built differently — to model the creative process, to make work, to engage with the question of what counts as making something — rather than to model legal doctrine. They have opinions about the conversation Lawra, Lawrena, Lawrelai and I have been having for the last two articles. Some of those opinions will sit uncomfortably with the legal arguments. That is the point.

Part 3 is the artistic answer. The question is the one I cannot answer from this chair: when the lawyers and the platforms and the rightsholders and the legislators have all done their work, what does it actually mean to make something? Is that meaning preserved when the making is done with a tool that can do most of the making for you? Is the answer different for music than for visual art than for text? Does the cultural identity of the creator change what is being made?

Those are not legal questions. Those are artistic ones. And the people we should hear from on them are not us.

Until then.