AI Copyright Lawsuits Tracker (2026): Every Active Case, Status, and Settlement — Updated Weekly

AI Copyright Lawsuits 2026

Last updated: October 4, 2026, 13:00 UTC | Update cadence: Weekly (every Monday); major developments (verdicts, settlements, filings ≥$500M) within 24 hours | Maintained by: Axis Intelligence Research & Sarah Mitchell — 49 rows, CC BY 4.0, every row primary-sourced

Quick Answer

On September 29, 2026, a US court of appeals ruled on AI training fair use for the first time — and rejected the defence. The Third Circuit affirmed Judge Bibas in full in Thomson Reuters v. Ross Intelligence, holding that Westlaw’s 2,243 headnotes are original and that training a competing legal research tool on them was not fair use. The holding is narrower than the headline: the facts are a non-generative retrieval tool built to replace the product it copied from, and the court reserved the generative question. Separately, Anthropic’s $1.5 billion Bartz settlement received final approval on July 20 and is now in distribution, while the CJEU opinion in Like Company remains undelivered a month past its date.

Key Findings

• On September 29, 2026 the Third Circuit affirmed that Ross Intelligence’s use of 2,243 Westlaw headnotes to train a competing legal search tool was not fair use, the first US appellate ruling on AI training fair use.
• The Third Circuit held that unlike necessity, ease is not a justification for copying, distinguishing the intermediate-copying software cases on the ground that Ross did not need the headnotes to reach the unprotected information beneath them.
• Judge Araceli Martínez-Olguín granted final approval in Bartz v. Anthropic on July 20, 2026, certifying a Works List of 482,460 books and overruling all 53 objections.
• Axis Intelligence Research calculates the Bartz Settlement Efficiency Ratio at 2.07% — $1.5 billion against a $72.369 billion theoretical ceiling under 17 U.S.C. §504(c).
• Axis Intelligence Research calculates an Opt-Out Exposure Delta of $264,697,463 — the gap between the statutory ceiling on the 1,802 opted-out works and their value inside the class.
• On July 31, 2026, the Munich District Court in GEMA v. Suno applied US copyright law to training conducted in the United States and rejected the fair use defence.
• The US Department of Justice filed in support of OpenAI’s fair use position in the S.D.N.Y. MDL; on September 28 the news plaintiffs asked the court to give it no weight.

Latest Developments

Reverse chronological. Every entry carries a primary-source link fetched on the retrieval date recorded in the CSV. Entries older than 90 days are in the archive at the foot of this page.

September 29, 2026 — The first US appellate ruling on AI training fair use goes against the AI company

A three-judge panel of the Third Circuit — Restrepo, Montgomery-Reeves and Bove, with Montgomery-Reeves writing — affirmed Judge Bibas’s February 2025 partial summary judgment in full. The opinion was filed under seal on the 29th and released the next day once neither party proposed redactions. Twenty-two months of “no appellate court has decided this” ended in a single afternoon.

On copyrightability, the 2,243 Westlaw headnotes are original, and merger does not apply because there are many ways to express a point of law drawn from a judicial opinion. The panel reached for a banana costume to make the point, which tells you how settled it considered the question.

On fair use, factors one and four went to Thomson Reuters; factor two went to Ross on the minimally expressive nature of the headnotes, and lost anyway. The reasoning that will travel furthest is the treatment of intermediate copying. Sega and Connectix protected copying because it was the only way to reach unprotected functional elements of software. Ross had no such necessity — the underlying law was free for the taking, and it was the editorial layer on top that Ross took. Hence the line practitioners are already quoting: unlike necessity, ease is not a justification for copying.

Underneath the doctrine sit three undisputed facts the panel kept returning to. Ross was commercial. Ross priced itself against Westlaw. Ross set out to replace Westlaw. A defendant with any one of those is in trouble on factor four; a defendant with all three was never going to win this appeal. Status: decided; Thomson Reuters affirmed | Docket: 3d Cir. 25-2153 Unsealed opinion, via Chat GPT Is Eating the World

October 4, 2026 — What the Ross ruling does not decide, and who is already overstating it

Within seventy-two hours, outlets were describing the decision as an explicit carve-out preserving fair use for generative AI. It is not that, and the distinction matters enough to state plainly.

The facts are a non-generative retrieval tool, trained on a direct competitor’s proprietary human-curated editorial product, to perform substantially the same commercial function. Judge Bibas expressly limited his 2025 ruling to non-generative AI, and reporting on the unsealed opinion indicates footnote 7 reserves the generative question. A reservation is not a holding. The court declined to decide how this analysis applies to a generative model trained on a corpus it does not compete with; nothing in the opinion tells you how that case comes out.

What does transfer is the framework, and it is not comfortable for defendants. Market harm under factor four now has appellate backing for the proposition that a licensing market for training data counts as a market, even where none existed before the defendant’s conduct. That reasoning does not care whether the model is generative. Both Concord actions and the OpenAI MDL are being briefed on factor four right now.

Axis Intelligence Research records that the court did not decide the generative question and takes no position on how a court would resolve it. We have not yet read footnote 7 in the full text; the characterisation above is sourced to practitioner analyses and flagged accordingly. Status: Axis scope note | Flagged pending_primary — page-level verification queued

September 28, 2026 — The DOJ takes OpenAI’s side, and the publishers ask the court to ignore it

The Department of Justice filed in support of OpenAI’s position that training on copyrighted works is fair use. On September 28 the New York Times, Daily News and the other news plaintiffs asked Judge Stein to give the filing no weight at all, arguing that it echoes the administration’s political view that unfettered AI development is a national security priority, that this is a separate matter from the copyright questions before the court, and that enforcing copyright poses no serious risk to the industry’s development.

Whatever the court makes of it, the filing is a marker. The executive branch has now taken a public position on the central question in the largest AI copyright proceeding in the country, five months after summary judgment briefing closed and with no ruling yet issued. Status: summary judgment, Rule 12(c) and sanctions all pending Bloomberg Law

September 26, 2026 — Two cases we should have been tracking: publishers v. Google, and the publishers v. Cohere

Both surfaced this week through passing references in coverage of other matters, which is its own lesson about relying on case-name searches.

In mid-July, Hachette, Cengage, Elsevier, Scott Turow and S.C.R.I.B.E. filed a class action against Google over the use of their works to train Gemini. The plaintiff group is the interesting part: trade publishers, academic publishers and an individual author of standing in one action. Every book case so far has been one of those constituencies at a time, and the class definition problem gets harder when the works span consumer trade titles, academic monographs and journal content with entirely different licensing markets. Court, docket and exact filing date are not confirmed — we record the month, not a day.

Condé Nast, The Atlantic, Axel Springer and other news publishers v. Cohere (S.D.N.Y., 1:25-cv-01305) has been running since February 2025, and the court denied Cohere’s partial motion to dismiss in November 2025. It matters for a reason the frontier-lab cases cannot test: Cohere is a mid-sized developer. Whether the fair use case law now forming around Anthropic, Meta and OpenAI scales down to a company without a frontier-lab litigation budget is a question only a case like this answers. Status: both active | Flagged pending_primary — dockets to be retrieved TechCrunch · BakerHostetler case tracker

September 26, 2026 — Conflicting accounts of where Concord I actually stands

A law-firm case tracker last updated in July describes a Concord v. Anthropic action as stayed pending interlocutory appeal of the court’s dismissal of the DMCA claims. Our August 24 entry, from a different firm’s digest, has the parties briefing cross-motions for summary judgment on fair use in the same period. A stayed case is not briefing dispositive motions.

The two accounts may describe the two different Concord actions, or one may be stale. We have not resolved it, so the row is flagged source_inconsistency and both dockets go on the direct-check list. Flagging our own prior entry as possibly wrong is the cost of the two-Concord split we made in August; the alternative — quietly preferring whichever source is newer — is how a tracker becomes confidently incorrect. Status: unresolved | Flagged source_inconsistency BakerHostetler case tracker

September 26, 2026 — The CJEU opinion is three weeks overdue, and the hearing hinted at why it might be hard

Twenty-three days past its scheduled date, still no opinion text, no delivery report, no postponement notice. Third consecutive cycle.

Two details worth adding while we wait. The reference comes from the Budapest Environs Regional Court and sits with the Grand Chamber — the fifteen-judge formation reserved for cases of particular importance, which is itself a signal about how the Court reads the stakes. And commentary from the March hearing records AG Szpunar pressing a line of questioning that would collapse the input-output dichotomy altogether, analogising to a work reproduced without authorisation in one country and then recited in another. Google objected on the spot. If the opinion attempts that, it would let courts treat statistical patterns in a model’s weights as reproductions of the training data, carrying a finding at the input stage through to the output stage.

That is hearing questioning, not a holding, and we record it only as an indication of what the pending opinion may have to work through. It may also explain a three-week slip. Status: scheduled September 3, 2026; delivery unconfirmed | Flagged pending_primary TechnoLlama, on the March hearing

September 4, 2026 — Bartz claimants find out who else claimed their titles

The settlement administrator sent every claimant a notice summarising their claims, naming who else filed a claim for the same title, and stating what percentage of the award those others claimed. That is the overlapping-claims problem made concrete: a single book can carry competing claims from author, estate and publisher, and at roughly $3,109 per work those percentages are the whole negotiation. No payment commencement date has been announced. Status: distribution under way Authors Guild, what authors need to know about the Anthropic settlement

September 19, 2026 — Sixteen days past its date, the CJEU opinion still has not surfaced

Advocate General Szpunar’s opinion in Like Company v Google Ireland (C-250/25) was scheduled for September 3. We flagged it as scheduled-not-delivered on September 7. Twelve days later the position is unchanged: no opinion text retrieved, no report of delivery, no notice of postponement. A practitioner guide published in early September still describes the opinion as expected on the third — written, evidently, before the date and never revisited.

That last detail is the useful one. Commentary reproduces a scheduled date indefinitely without checking whether it happened, which is precisely how a calendar entry becomes a reported fact across a dozen pages. We are escalating our own handling accordingly: the curia.europa.eu case register for C-250/25 goes on the direct-check list rather than being monitored through secondary coverage. Status: scheduled September 3, 2026; delivery unconfirmed, second consecutive cycle | Flagged pending_primary Chambers and Partners, AI & Intellectual Property 2026 — EU

Correction, September 19, 2026 — Kneschke v. LAION already has an appellate judgment

Our previous releases listed Kneschke v. LAION as pending before the Federal Court of Justice with no date and no appellate decision on record. The Hamburg Higher Regional Court gave judgment on December 10, 2025 in case 5 U 104/24, addressing the text-and-data-mining exceptions as applied to the copying involved in building an image-text dataset, and the effectiveness of rights reservations. The matter is on further appeal to the BGH as I ZR 281/25.

A nine-month-old appellate judgment on the TDM exception is not a detail, and it sat unrecorded because the row was carried forward on its status line rather than re-researched. Flagged pending_primary: the judgment text has not been retrieved, so its holdings are described here only at the level our source states them. Chambers and Partners, AI & Intellectual Property 2026 — EU

September 19, 2026 — What the Third Circuit panel actually pressed on, and when a ruling is expected

With the Ross opinion now fourteen weeks out, the argument record is worth setting down properly. The panel was Judges L. Felipe Restrepo, Tamika R. Montgomery-Reeves and Emil J. Bove, and the questioning concentrated almost entirely on factors one and four. Judge Bove framed the case as a choice between two characterisations of Ross’s training memos: a cheat sheet derived from Westlaw, or something materially different. Judge Montgomery-Reeves pressed Ross on why headnotes should fail the originality threshold at all — a line of questioning that matters because Ross needs to win on either copyrightability or fair use, and the panel seemed disinclined to hand it the first.

Ross’s best card is in-circuit: the Third Circuit’s own April 2026 decision in ASTM v. UpCodes, filed as a Rule 28(j) letter, where a for-profit AI-native startup republishing building codes prevailed on fair use because the material was the law. Whether that reasoning reaches headnotes — editorial summaries of opinions rather than the opinions themselves — is the appeal in miniature.

Counsel commentary published in July places the decision in late 2026. That is a practitioner’s expectation, not a court statement, and we record it as such. Status: argued June 11, 2026; no opinion as of September 19 Baker Botts, Third Circuit hears oral argument in Ross v. Reuters

September 7, 2026 — The CJEU opinion that was due last Thursday, and what we can and cannot say about it

Advocate General Szpunar was scheduled to deliver his opinion in Like Company v Google Ireland (C-250/25) on September 3. The date is confirmed by counsel of record for Google and by contemporaneous coverage of the March hearing. As of today we have not retrieved the opinion text, nor any report that it was delivered, nor any report that it was postponed.

So we are not reporting it. Under this tracker’s exclusion policy, a scheduled date is not a completed event, and “the AG has now opined that…” is exactly the sentence that would be repeated across a dozen downstream pages before anyone checked. The row is flagged pending_primary and we recheck curia.europa.eu each cycle.

What is worth stating while we wait is the shape of the question. The Grand Chamber sat for six hours in March on four referred questions: whether a chatbot displaying content partially identical to a protected press publication communicates to the public, and whether next-token prediction bears on that; whether training is an act of reproduction; and whether any such reproduction falls inside the Article 4 DSM text-and-data-mining exception. Member States also argued the extraterritorial reach of EU copyright — the same territorial question the Munich court answered for itself in GEMA v. Suno by applying US law to US training. Europe’s answer will bind 27 member states at once, where national courts have so far answered fragments. Status: opinion scheduled September 3, 2026; delivery unconfirmed | Flagged pending_primary Bird & Bird, CJEU first hearing on generative AI and copyright

September 7, 2026 — Andersen v. Stability AI: our sources disagree about whether a jury trial starts tomorrow

Contemporaneous coverage of Judge Orrick’s August 2024 order recorded a jury trial in the visual-artist class action set to begin September 8, 2026 — tomorrow — which would make it the first AI copyright jury trial in the United States. Coverage from May 2026 describes the same case as in active discovery, with a third amended complaint filed February 27, answers on March 13, and class certification briefing running through spring and summer. Those two accounts cannot both be current.

We have not resolved the conflict against the N.D. Cal. docket, so we publish neither date as fact and flag the row source_inconsistency. This is the same failure mode that produced our Sony v. Suno correction in August: a trial date set years in advance, repeated long after the schedule moved. The difference is that this time we caught it before printing it.

What is not in dispute is why the case matters. The August 2024 order accepted the model theory — that Stable Diffusion may itself be an infringing copy because it embodies compressed transformations of the training images — while dismissing the DMCA §1202(b) claim with prejudice for want of identical outputs. A jury verdict on the model theory would be the first time twelve citizens, rather than a judge, decided what a set of weights is. Status: active; trial date unresolved | Flagged source_inconsistency NYU JIPEL analysis of the August 2024 order

August 31, 2026 — Midjourney and the studios come out of mediation without a settlement

The Central District of California had ordered Disney, Universal, Warner Bros. Discovery and Midjourney into private mediation, to be completed by August 19, with a report on the outcome due August 21 and a post-mediation status conference set for August 31. The court’s November 2025 order provided that no appearance would be required on the 31st if a notice of settlement was filed by the 21st. No such notice appears on the docket.

That makes this the first major studio AI case to go through court-ordered mediation and come out the other side still litigating. Non-expert discovery closes September 21, expert discovery November 9, and the last day to file motions is November 23 — which puts a dispositive fair use motion in the visual-generation context on the calendar for late this year. Status: active; mediation concluded without reported settlement | Docket: 2:25-cv-05275-JAK-AJR Docket, Disney Enterprises Inc. v. Midjourney Inc., via CourtListener

August 17, 2026 — Midjourney tries to make the studios disclose their own AI use

A discovery fight worth watching because it inverts the usual posture. Midjourney sought disclosure of how Disney, Universal and Warner Bros. Discovery use AI internally. In mid-June Magistrate Judge Richlin limited the studios’ obligations to AI tools intended for general consumers, rejecting the broader demands as not relevant to whether infringement occurred. Midjourney filed a Rule 72 objection with Judge Kronstadt on June 29; the hearing was set for August 17.

The question — how far a defendant may probe a rightsholder’s own AI adoption — has not been squarely decided elsewhere on this docket, and the answer shapes what “market harm” evidence looks like when the plaintiff is itself building with the technology. Status: objection heard; ruling pending | Flagged pending_primary Docket, Disney Enterprises Inc. v. Midjourney Inc., via CourtListener

August 24, 2026 — The publishers put all four fair use factors in front of the court in Concord I

Anthropic’s lyrics case is no longer in discovery skirmishes. It is at cross-motions for summary judgment on fair use, and this week the publishers filed both their opposition to Anthropic’s motion and their reply supporting their own. Their argument runs the factors in order and lands hardest on two. On purpose, they argue lyrics as training data are useful only for generating lyrics — Anthropic’s code and reasoning models have no need for them — which makes the use non-transformative under Warhol. On amount, they lean on Anthropic’s own admission: it copied the works whole, and it did not need them at all. That second concession is what separates this case from every fair use ruling that has gone a developer’s way. Alsup and Chhabria both credited training necessity. A defendant who concedes it didn’t need the corpus has given away factor three and weakened factor one behind it.

On market harm the publishers invoke the market-dilution theory Chhabria sketched in Kadrey — the theory that lost that case on an evidentiary record, now argued by parties who watched what went wrong. Status: cross-motions for summary judgment pending | Flagged pending_primary McKool Smith AI Litigation Tracker, August 24, 2026

August 24, 2026 — Two courts choose to shrink the music cases rather than let them grow

The court in UMG Recordings v. Suno denied the labels’ motion for leave to add further recordings on prejudice grounds — then invited them to file a separate action that would likely be designated as related and could run on its own discovery schedule. The stated reasoning is that resolving fair use in the present case on summary judgment comes first. The same sequence had already played out in Sony v. Uncharted Labs, where the court refused to add works because expansion would delay the central question.

Two courts, two denials, one instinct: the fair use question is worth more decided quickly on a narrow record than decided late on a complete one. For rightsholders that is a procedural loss with a strategic consolation, and it means the first US music-training fair use ruling will rest on a deliberately small set of works. Status: amendment denied; related action invited | Flagged pending_primary McKool Smith AI Litigation Tracker, August 24, 2026

August 6, 2026 — Cox reshapes the OpenAI MDL, and the news plaintiffs lose their bid to adapt

Following the Supreme Court’s decision in Cox Communications v. Sony, OpenAI moved for judgment on the pleadings against the New York Times, Daily News, the Center for Investigative Reporting and Ziff Davis, arguing Cox abrogates the knowledge-plus-material-contribution theory their contributory infringement claims depend on. The Times and Daily News sought leave to amend to conform those claims to Cox, and to voluntarily dismiss their trademark dilution claims. On August 6 the court found they had not shown good cause to add new theories of contributory infringement.

Separately, the news plaintiffs moved for evidentiary sanctions over OpenAI’s deletion of chat logs. OpenAI and Microsoft responded in the week of August 17 — OpenAI casting itself as having preserved 50 billion user chats and characterising the deleted logs as technically infeasible to retain, Microsoft echoing the argument on spillover-prejudice grounds. That figure is OpenAI’s own characterisation in a brief, not a court finding. Status: Rule 12(c) motion, summary judgment, and sanctions all pending | Flagged pending_primary McKool Smith AI Litigation Tracker, August 24, 2026

Correction, August 25, 2026 — Judge Chhabria denied the Kadrey interlocutory appeal in July

Our August 12 and August 19 releases recorded the Kadrey §1292(b) certification motion as pending. It had already been denied in the week of July 13, 2026. Ninth Circuit review of the shadow-library holding is foreclosed for now, and the Third Circuit in Ross remains the only appellate proceeding on AI training fair use. The seeding and distribution claim in Kadrey continues on its own track, and a mid-August case management order provides for a four-hour deposition of Mark Zuckerberg by plaintiffs.

The error came from a source published before the ruling that we did not re-verify against a later edition of the same publication. The fix is a protocol change: pending-motion rows now carry a re-verification requirement against the most recent edition of their source before each release, not just a docket check. McKool Smith AI Infringement Case Updates, July 13, 2026

August 19, 2026 — Fifteen weeks after argument, the first appellate AI fair use ruling was still not in

The Third Circuit docket in Thomson Reuters v. Ross Intelligence (25-2153) showed no opinion entered as of September 26, fifteen weeks after the June 11 argument. It issued three days later. The S.D.N.Y. summary judgment ruling in the OpenAI MDL is likewise outstanding, nearly five months after briefing closed on April 2. Both are 24-hour update triggers for this tracker. There is no inference to draw from the silence — panels take the time they take, and a long hold on a first-impression question with ten amicus briefs is unremarkable — but the practical consequence is worth stating plainly: every enterprise procurement decision made this quarter is being made without appellate law on the central question. Status: superseded — the Third Circuit ruled September 29, 2026; the OpenAI MDL ruling remains pending Third Circuit docket, 25-2153, via CourtListener

Entry added August 19, 2026 — Anthropic seeks dismissal of the opt-out authors’ complaint, and the co-defendants split three ways

Filed February 24, 2026; added to this tracker on review of the Carreyrou pre-severance docket. Alongside its severance motion, Anthropic moved to dismiss on two grounds: that the plaintiffs fail to allege Anthropic copied their specific works, and that two of the six lack standing. The co-defendants’ responses diverged instructively. Meta filed an answer on February 26 rather than moving to dismiss — the posture of a defendant that already has a fair use ruling in hand from Kadrey and would rather litigate than plead. Perplexity moved to sever and dismiss on February 27. The OpenAI claims left by stipulation for the S.D.N.Y. MDL. Three defendants, three theories of where this case is weakest, and only Anthropic still in front of Judge Pitts. Status: motion to dismiss pending post-severance | Flagged pending_primary — motion papers not independently retrieved Mishcon de Reya generative AI case tracker

August 5, 2026 — Anthropic and Dario Amodei move to narrow the publishers’ $3.1 billion lyrics case

Anthropic filed a partial motion to dismiss in Concord II (N.D. Cal., 5:26-cv-00880), targeting the portion of the publishers’ direct infringement claim based on Claude outputs. Its argument is evidentiary: a second amended complaint asserting roughly 21,000 works identifies not one allegedly infringing output. Anthropic separately moved to dismiss the DMCA claim in full, arguing §1202 reaches removal of copyright management information from entire works — the musical compositions — rather than from lyrics used as training text, and that it never possessed those compositions. Dario Amodei filed his own motion, arguing the complaint alleges torrenting by other defendants and pleads only contributory conduct as to him personally. Amount at stake: $3.1B claimed | Status: motions pending, no hearing date Filed motion papers via Digital Music News

July 31, 2026 — Munich court applies US fair use law to US training, and rejects it

The 42nd Civil Chamber of the Munich District Court prohibited four distinct acts across six compositions in GEMA v. Suno (42 O 763/25): reproduction for training in the United States, reproduction through memorisation in the model in Germany, communication to the public by offering the model, and reproduction and communication through the outputs. The court asserted jurisdiction over the US training acts via a venue rule available to collecting societies, applied US copyright law to them, and held fair use did not shield them. On market effect it placed the burden of disproving harm on Suno — the mirror image of how Kadrey failed in California. Penalty: up to €250,000 per future breach, or custodial detention up to six months | Status: first instance, appeal expected Bird & Bird case analysis

July 20, 2026 — Final approval and final judgment in Bartz v. Anthropic

Judge Araceli Martínez-Olguín’s 23-page order found the settlement fair, reasonable and adequate, overruled all 53 objections, and entered final judgment. The certified Works List covers 482,460 books with ISBN or ASIN and US Copyright Office registration within five years of first publication; the claims rate reached 91.3%. Attorney fees and costs were set at $122 million and service awards at $15,000 each, reduced from the $50,000 requested. The release is past-only, running through August 25, 2025, and does not release claims relating to AI outputs or future conduct. Amount: $1.5B confirmed settlement, ~$3,109 per work | Status: concluded; distribution pending, no payment commencement date announced Authors Guild — court grants final approval

June 30, 2026 — Amended scheduling order confirms no Sony v. Suno fair use ruling before 2027

The District of Massachusetts entered an amended scheduling order running fact discovery and depositions to September 30, 2026 and setting dispositive motions for April 9, 2027. This forecloses the summary judgment hearing widely reported for July 2026. The motion by UMG and Sony to add 61,026 recordings, identified through Audible Magic audio fingerprinting of Suno’s training data, remains pending before Chief Judge F. Dennis Saylor IV. Status: active, no fair use ruling procedurally possible before 2027 Music Business Worldwide coverage

June 11, 2026 — Third Circuit hears the first appellate argument on AI training fair use

The panel in Thomson Reuters v. Ross Intelligence (3d Cir. 25-2153) pressed Ross to articulate concretely what made its product transformative, returning repeatedly to whether a use can be transformative when it serves the same end purpose as Westlaw. Ross conceded its product and Westlaw are at some level the same to a user, and leaned on the Third Circuit’s April 2026 decision in ASTM v. UpCodes. The appeal is from Judge Bibas’s February 2025 ruling that Ross’s use of 2,243 Westlaw headnotes was not fair use. Status: argued; no ruling as of August 19, 2026 Sterne Kessler — Third Circuit probes fair use

June 9, 2026 — Kadrey plaintiffs seek Ninth Circuit review of the shadow-library ruling

Plaintiffs asked Judge Chhabria to certify his 2025 summary judgment on Meta’s downloading from shadow libraries for interlocutory appeal under 28 U.S.C. §1292(b). Certification would put shadow-library acquisition before a second court of appeals alongside the Third Circuit’s Ross appeal. The seeding and distribution claim — alleging Meta distributed over 40TB of pirated content while torrenting Anna’s Archive in 2024 — remains live independently. Status: certification motion pending, no ruling as of August 19, 2026 Docket coverage, Chat GPT Is Eating the World

June 8, 2026 — Judge Pitts severs the opt-out authors’ omnibus action against the AI industry

Six authors who declined the Bartz class — led by Pulitzer winner John Carreyrou — had sued Anthropic, Google, OpenAI, Meta, xAI, Perplexity, Apple and NVIDIA in one action, seeking the full $150,000-per-work statutory ceiling from each. Judge P. Casey Pitts granted the severance motions, holding there was no conspiracy allegation sufficient to join competing LLM developers and noting each defendant trained on a different mixture of repositories at different times between 2020 and 2024. Claims against Google, xAI, Perplexity, Apple and NVIDIA were dismissed without prejudice to individual refiling; Meta’s were related to Kadrey and OpenAI’s transferred into the S.D.N.Y. MDL. The case proceeds against Anthropic alone as Carreyrou v. Anthropic (N.D. Cal., 3:25-cv-10897). Amount sought: $150,000 per work per defendant (statutory ceiling; no aggregate pleaded) | Status: active against Anthropic Order granting motions to sever, Entry 195 (PDF) Archive — entries older than 90 days (click to expand)

May 23, 2026 — MiniMax loses its motion to dismiss the studios’ output-infringement case

Judge Blumenfeld denied MiniMax’s motion in the Disney, Universal and Warner Bros. Discovery action over Hailuo AI, finding the studios plausibly allege widespread intentional direct and secondary infringement, and rejecting the personal jurisdiction challenge from a Shanghai-based defendant operating through a Singaporean entity. The theory here is output infringement rather than training: the claim is that generation of recognisable protected characters is itself the infringing act, which is why it survived a motion a training-only theory might not have. Status: active, in discovery McKool Smith AI litigation updates

May 18, 2026 — Jury returns a unanimous verdict against Musk in Musk v. Altman

The jury found for OpenAI and its co-founders on all claims, which were held time-barred. Musk is appealing. The case turned on OpenAI’s conversion from nonprofit governance rather than on copyright, and is tracked here for its governance implications rather than its infringement holdings. Status: concluded at trial, on appeal McKool Smith AI litigation updates

May 5, 2026 — Five major publishers open a second front against Meta

The publishers filed in S.D.N.Y., naming Mark Zuckerberg personally and alleging removal of copyright management information from works used in training. The action is separate from Kadrey and proceeds on its own schedule. Status: active, early stage Association of American Publishers

April 23, 2026 — Getty’s US trademark claims survive dismissal

The N.D. Cal. court denied Stability AI’s motion to dismiss Getty’s Lanham Act claims, finding that outputs bearing distorted Getty watermarks supported a reasonable inference of likely confusion. Getty lost its UK copyright case on training in November 2025 and won a narrow trademark point there; in the US, the trademark theory is again what survived first contact. Case management conference set for November 5, 2026. Status: active McKool Smith AI litigation updates

April 1, 2026 — Reddit v. Anthropic remanded to state court

The action was remanded to San Francisco Superior Court. The claims are contractual — breach of terms of service governing data scraping — not copyright, and are tracked here as an adjacent acquisition-side theory. Status: active, state court McKool Smith AI litigation updates

March 18, 2026 — BMG sues Anthropic over lyrics

BMG Rights Management filed against Anthropic alleging Claude was trained on its lyrics catalogue without licence. No aggregate damages figure is pleaded. Status: active, newly filed Reuters legal

March 10, 2026 — CJEU reference in Like Company v. Google Ireland progresses

The Hungarian press publisher’s reference (C-250/25) asks whether LLM outputs engage the reproduction right and how far the text-and-data-mining exception extends. The Advocate General’s opinion is due September 3, 2026 and is a full-revision trigger for this tracker. Status: pending CJEU case register

March 2, 2026 — Supreme Court denies certiorari in Thaler v. Perlmutter

The denial leaves standing the holding that copyright protection requires human authorship. Thaler had explicitly disclaimed any human creative input, so the courts did not address how much human involvement suffices for AI-assisted works to be registrable. Status: concluded Norton Rose Fulbright AI litigation series

January 28, 2026 — Warner settles with Suno; publishers file Concord II

Warner Music’s action against Suno was dismissed following a licensing partnership and the Songkick acquisition; terms undisclosed. The same day, Concord, Universal Music Publishing entities and ABKCO filed Concord II against Anthropic, Dario Amodei and Benjamin Mann, seeking more than $3 billion over alleged torrenting of pirated libraries containing their compositions. Amount: $3.1B claimed (Concord II); Warner settlement undisclosed Music Business Worldwide

November 11, 2025 — GEMA wins against OpenAI in Munich at first instance

The Munich District Court held that memorisation within the model and the reproduction of lyrics in outputs both infringe under German law, and rejected the text-and-data-mining exception as a defence. The judgment is on appeal before the Munich Court of Appeals. Status: first instance concluded, on appeal Reed Smith German case analysis

Entry added August 25, 2026 — Apple and Elsevier join the tracker

Two actions absent from prior versions. Hendrix and Roberson v. Apple (N.D. Cal., filed September 5, 2025) alleges Apple trained its OpenELM model family — hosted publicly on Hugging Face — on pirated books. Apple has answered, and a stipulated protective order providing for source code inspection was granted, which means technical discovery is running. Elsevier et al. v. Meta is at the pleadings stage, where Meta has raised copyright misuse and unclean hands as an affirmative defence and the publishers have asked for leave to move to strike it. Unclean hands has been pleaded by several AI defendants across this docket; this is the first time a plaintiff has challenged it head-on, so the ruling may travel. Status: both active | Flagged pending_primary — docket numbers not yet retrieved McKool Smith AI Litigation Tracker, August 24, 2026

October 15, 2025 — Salesforce added to the tracker: authors allege XGen was trained on Books3

Entry added August 19, 2026, closing a coverage gap. Novelists E. Molly Tanzer and Jennifer Gilmore filed a putative class action in N.D. Cal. alleging Salesforce copied the RedPajama and The Pile datasets — both containing the Books3 corpus of roughly 196,000 books taken from the private tracker Bibliotik — to train its XGen large language models. The distinguishing allegation is documentary rather than technical: the complaint says Salesforce listed RedPajama-Books among its disclosed training sources when XGen launched in June 2023, then deleted those references and replaced them with descriptions of publicly available data. The claim is pleaded as ongoing infringement through continued storage and processing, not as a completed training act. No aggregate damages figure is pleaded. Status: active | Flagged pending_primary — docket number and assigned judge not yet retrieved Bloomberg Law coverage

June 23, 2025 — Judge Alsup issues the doctrinal anchor in Bartz

Training on legally acquired books is fair use; building a permanent library of pirated copies is a separate, non-transformative use. The $1.5 billion settlement is a price on the second half of that holding, not the first. Status: merits ruling, case since settled Authors Guild

February 11, 2025 — Judge Bibas rejects fair use in Thomson Reuters v. Ross

Direct infringement of 2,243 Westlaw headnotes, with fair use rejected as a matter of law on factors one and four. The ruling is the only US decision rejecting AI training fair use outright, and is the decision now before the Third Circuit. Status: on interlocutory appeal Norton Rose Fulbright AI litigation series

How We Curate This List

The inclusion criteria are the methodology. What counts, what doesn’t, and why.

What earns an entry. A copyright, DMCA, or directly adjacent acquisition-side action (contract or terms-of-service claims over training-data scraping) brought against a developer, deployer, or dataset maintainer of a generative AI system, in any jurisdiction, where a docketed filing or issued order exists. Each entry records a specific dated event — a filing, an order, a verdict, a settlement, a scheduling change — not a case’s general existence.

What every entry must carry. A fetchable primary-source document, retrieved at entry time and logged with its URL and retrieval date in the CSV. An entry without a source does not exist. A tracker row with no primary source is a fabricated incident, and we treat it as one.

Source hierarchy. Court orders and filed papers retrieved directly (PACER/CourtListener, court PDFs, official judiciary sites) outrank filed-order PDFs republished in full by specialist docket-tracking publications, which outrank named institutional legal analysis (Bird & Bird, Reed Smith, Sterne Kessler, Norton Rose Fulbright, Pearl Cohen, McKool Smith), which outrank rightsholder organisations, which outrank trade press with linked filings. The lower three tiers can establish that a filing exists and what it argues; they cannot establish a financial figure.

What we exclude, and why.

• Damages claims with no figure pleaded in a filed complaint. A reported “billions at stake” is not a number we can defend twelve months out.
• Projected or predicted outcomes in active cases. We track what courts did, not what commentators expect.
• Hearing or ruling dates not confirmed against a docket entry or order. This is the exclusion that produced our July 2026 correction: a widely reported Sony v. Suno summary judgment hearing did not exist and was never scheduled. Trackers that repeat calendar dates from trade coverage propagated that error; we now require the order.
• AI defamation, deepfake, right-of-publicity, and chatbot product-liability litigation. Distinct legal categories with different elements. Mixing them inflates case counts and makes cross-tracker comparison meaningless.
• Aggregate “total AI litigation exposure” figures. A court-approved settlement and a plaintiff’s damages demand are not commensurable quantities. We publish them in separate columns and never merge them.

Verification flags, published rather than suppressed. Rows resting on reliable secondary reporting where we have not independently retrieved the underlying docket entry carry pending_primary in the CSV. Eighteen rows carry a flag in this release, fifteen pending_primary and three source_inconsistency. Two of this week’s flags mark cases we had never tracked at all, and one marks a conflict with our own August entry. The second flag is the more useful one: it marks a place where two reputable sources contradict each other and we have chosen to print neither rather than pick the more convenient. The count rose sharply this cycle because a single well-sourced secondary digest surfaced four previously untracked matters and three corrections at once; every one of those rows is queued for PACER retrieval before the next Bilan. A flagged row is one we intend to upgrade, not one we are unsure exists.

Corrections. We do not silently edit a wrong figure. Corrections appear dated in the changelog and, where the error was consequential, in the body of the page. Report one: [email protected].

How Axis Calculates the Settlement Efficiency Ratio and the Opt-Out Exposure Delta

Two figures on this page are Axis constructions rather than observations. Both are computed from CSV rows and both show their inputs.

Settlement Efficiency Ratio™ — 2.07%. The certified Works List (482,460, from the final approval order) multiplied by the §504(c) willful infringement ceiling ($150,000) gives a theoretical maximum exposure of $72.369 billion. The confirmed settlement ($1.5 billion) divided by that figure is 2.07%. The court supplied the mirror framing: the roughly $3,000 per-work payment is about four times the statutory minimum for willful infringement. Measured against the floor, authors did well. Measured against the ceiling, Anthropic bought closure for two cents on the dollar.

Opt-Out Exposure Delta™ — $264,697,463. Inaugural baseline, August 2026.

Input	Value
Opt-out claimants	350
Works represented	1,802 (0.37% of the certified Works List)
Statutory ceiling (1,802 × $150,000)	$270,300,000
Value inside the class (1,802 × $3,109.07)	$5,602,537
Opt-Out Exposure Delta™	$264,697,463

Required disclosure: the ratio form of this figure (48.25×) is the arithmetic reciprocal of the Settlement Efficiency Ratio and is not an independent measurement. The new quantity disclosed is the dollar magnitude scoped to the opt-out cohort. It is a theoretical ceiling, not a pleaded demand — no opt-out complaint pleads an aggregate figure, and Judge Pitts’s severance order means whatever is recovered must be recovered one defendant at a time, which is a real drag on realisation that this number does not model.

Limitations. Both figures assume every work would qualify for maximum willful damages, which no court has held. Both are ceilings, useful as a denominator for judging settlement economics and useless as a prediction of recovery.

Which AI Companies Face the Most Copyright Exposure?

No single primary source publishes consolidated per-defendant exposure. Axis Intelligence Research compiled the table below from the tracker’s own rows. Confirmed amounts come from court-approved settlement documents; claimed amounts are as stated in filed complaints. The two columns are never merged.

Defendant	Confirmed settled	Active claims (claimed)	Status	Source
Anthropic	$1.5B (Bartz, final judgment July 20, 2026)	$3.1B (Concord II) + Concord I, BMG, opt-out cluster (unquantified)	Book class claims closed; five-plus active fronts	Authors Guild; Digital Music News
OpenAI	$0 confirmed	Unquantified (16 consolidated US cases + German first-instance liability)	MDL summary judgment pending; GEMA appeal pending	CourtListener; Reed Smith
Meta Platforms	$0 confirmed	Unquantified (Kadrey seeding claim + publisher class action)	Two active fronts; §1292(b) motion pending	McKool Smith
Suno	Warner settlement (undisclosed)	Unquantified (Sony/UMG US; GEMA Germany; Koda Denmark)	Lost in Germany July 31, 2026; US case runs to 2027	Bird & Bird; MBW
Udio	UMG settlement (per-generation royalty)	Unquantified (Sony, 333 works)	Document production closes Aug 25, 2026	MBW
Google	$0 confirmed	Unquantified (Hachette, Cengage, Elsevier, Turow, S.C.R.I.B.E.)	Newly filed; Gemini training corpus	TechCrunch
Cohere	$0 confirmed	Unquantified (Conde Nast, The Atlantic, Axel Springer et al.)	Partial motion to dismiss denied Nov 2025	BakerHostetler
Apple	$0 confirmed	Unquantified (Hendrix and Roberson, N.D. Cal.)	Active; OpenELM training corpus in technical discovery	McKool Smith
Salesforce	$0 confirmed	Unquantified (Tanzer and Gilmore, N.D. Cal.)	Active; Books3 via RedPajama and The Pile alleged	Bloomberg Law
Stability AI	$0 confirmed	Unquantified US action	UK won (appeal pending); US trademark claims survived	McKool Smith
MiniMax	$0 confirmed	Unquantified	In discovery	McKool Smith

Snapshot: October 4, 2026. Updated quarterly; the underlying tracker updates weekly.

Note the shape of the Anthropic row. The $1.5 billion closed one corpus-acquisition claim and nothing else. Both Concord actions, BMG, and every opt-out suit proceed against a company that has already demonstrated it will pay nine figures rather than try a piracy claim — which is the leverage the opt-out plaintiffs are pricing.

Has Any Appeals Court Ruled on AI Training Fair Use?

Yes — once, as of September 29, 2026, and against the AI company. The Third Circuit’s holding in Ross covers a non-generative retrieval tool built to replace the product whose editorial layer it copied. Every generative case below is still undecided at the appellate level. The scoreboard below is a compilation from every decided and pending AI training fair-use proceeding across three jurisdictions.

Court	Case	Fair use for training?	Date	Key distinction
N.D. Cal. (Alsup)	Bartz v. Anthropic	Yes — legally acquired books	Jun 23, 2025	Pirated-copy library separately not fair use; settled $1.5B
N.D. Cal. (Chhabria)	Kadrey v. Meta	Yes — on this record	Jun 25, 2025	Market harm dispositive; §1292(b) motion pending
D. Del. (Bibas)	Thomson Reuters v. Ross	No — Westlaw headnotes	Feb 11, 2025	2,243 headnotes; repackaging for a competing product
3rd Cir.	Thomson Reuters v. Ross	Pending — 10 weeks	Argued Jun 11, 2026	Would be the first US appellate ruling
9th Cir.	Kadrey v. Meta	Foreclosed for now	Certification denied w/o Jul 13, 2026	Chhabria refused §1292(b); no second circuit reaches the question
D. Mass. (Saylor)	Sony / UMG v. Suno	Not before 2027	Dispositive motions Apr 9, 2027	Widely misreported as a July 2026 hearing
S.D.N.Y. (Stein)	In re OpenAI MDL	Pending	Briefing closed Apr 2, 2026	20M ChatGPT logs in evidence
LG München I	GEMA v. OpenAI	No (German law)	Nov 11, 2025	Memorisation and outputs infringe; on appeal
LG München I	GEMA v. Suno	No (US law applied by a German court)	Jul 31, 2026	Extraterritorial jurisdiction; §107 analysis performed in Munich
CJEU	Like Company v. Google	Pending	AG opinion scheduled Sep 3, 2026; delivery unconfirmed	TDM exception scope; LLM outputs; four questions, Grand Chamber
OLG Hamburg	Kneschke v. LAION	No (German law, first appellate instance)	Dec 10, 2025 (5 U 104/24)	TDM exception and rights reservations for dataset creation
BGH	Kneschke v. LAION	Pending	I ZR 281/25	Further appeal from the Hamburg judgment

Source: Axis Intelligence Research compilation from the tracker CSV. As of October 4, 2026.

The most useful column is the one full of “pending.” Anyone describing the question as settled — in either direction — is describing district-court outcomes, not law.

Browse Every Tracked Case

The table below carries all 45 case-level rows in the dataset. Filter by defendant, jurisdiction, content category, status, or verification state; sort any column; search across case names, dockets and forums. Every row is rendered in the page source rather than built by JavaScript, so the full table is readable whether or not scripts run.

Axis-derived metric rows are excluded here and documented in the methodology section above. The pending_primary badge marks a row resting on reliable secondary reporting where we have not yet pulled the docket entry — published rather than hidden.

Correction Notices

September 26, 2026 — Concord I posture disputed. Our August 24 entry records cross-motions for summary judgment being briefed. A law-firm tracker describes a Concord v. Anthropic action as stayed pending interlocutory appeal of the DMCA dismissal over the same period. Unresolved; the row carries source_inconsistency and we are not asserting either account.

September 19, 2026 — Kneschke v. LAION carried forward without its appellate judgment. Prior releases listed the matter as pending before the BGH with no date. The Hamburg Higher Regional Court had given judgment on December 10, 2025 (5 U 104/24). Full notice in the entry log above.

August 12, 2026 — court and docket for the Anthropic music publishers’ action. Prior versions of this tracker carried a single UMG/Concord/ABKCO v. Anthropic row listing M.D. Tenn. with no docket number. There are two separate actions, both in the Northern District of California: Concord I (5:24-cv-03811, Judge Eumi K. Lee, 499 works, filed October 2023 and transferred from M.D. Tenn.) and Concord II (5:26-cv-00880, filed January 2026, roughly 21,000 works, $3.1B claimed). They are now two rows.

August 7, 2026 — the Sony v. Suno hearing that never existed. Between April and July 2026, trade press, general business media, and at least two competing litigation trackers reported a July 2026 summary judgment hearing on fair use before Chief Judge Saylor. It did not happen and was not scheduled. This tracker’s July 9, 2026 version carried that date. That was our error. The June 30, 2026 amended scheduling order sets dispositive motions for April 9, 2027.

August 7, 2026 — Bartz figures restated at final approval. Works List corrected from 482,000 to the certified 482,460; attorney fees corrected from $187.5M to the awarded $122M; Settlement Efficiency Ratio restated from 2.1% to 2.07% against the certified count.

About This Dataset

Download. ai-copyright-lawsuits-tracker.csv — 49 rows, 23 columns. Every row carries source_org, source_document, source_url, retrieved_date, is_primary, axis_calculated, verification_flag, and method_note. Three rows are Axis-derived and labelled axis_calculated=yes with formulas disclosed above.

Coverage. May 6, 2020 through October 4, 2026. US federal and state courts, the German Landgericht München I and BGH, the UK High Court, and the CJEU.

License. Creative Commons Attribution 4.0 International (CC BY 4.0).

Citation. Axis Intelligence Research, AI Copyright Lawsuits Tracker, 2026.

Monthly Bilan. Recomputed exclusively from the CSV rows. Figures in bilan prose reconcile to the CSV; rows carrying pending_primary are counted in case totals and excluded from financial aggregates.

Review trigger. Full revision on any of: the S.D.N.Y. summary judgment ruling in the OpenAI MDL; the S.D.N.Y. summary judgment ruling in the OpenAI MDL; the CJEU Advocate General opinion in Like Company (due September 3, 2026); any settlement ≥$500M; the Munich Court of Appeals ruling in GEMA v. OpenAI; a ruling on the Kadrey §1292(b) certification motion.

Changelog.

• October 4, 2026 — Full revision trigger fired. The Third Circuit decided Thomson Reuters v. Ross Intelligence on September 29, 2026, affirming Judge Bibas in full and rejecting fair use — the first US appellate ruling on AI training fair use. Quick Answer, Key Findings and the Fair Use Scoreboard rewritten. Added an Axis scope note on what the holding does and does not decide, against outlets describing it as a carve-out for generative AI. Added the DOJ’s filing in support of OpenAI and the news plaintiffs’ September 28 response. Like Company AG opinion unconfirmed for a fourth cycle. CSV now 49 rows.
• September 26, 2026 — Added Hachette, Cengage, Elsevier, Turow and S.C.R.I.B.E. v. Google and Condé Nast et al. v. Cohere, both previously untracked. Opened a source_inconsistency flag on Concord I over conflicting accounts of its posture. Like Company AG opinion unconfirmed for a third cycle; Grand Chamber composition and the March hearing’s input-output questioning added. Bartz distribution phase updated with the September 4 claimant notices. OpenAI MDL row anchored to the October 27, 2025 output-infringement ruling. CSV now 48 rows.
• September 19, 2026 — Correction published: Kneschke v. LAION appellate judgment of December 10, 2025 (OLG Hamburg, 5 U 104/24) added; the row previously showed only a pending BGH appeal. Ross row expanded with the panel composition and the substance of the June 11 argument. Like Company AG opinion still unconfirmed sixteen days past its scheduled date; curia.europa.eu moved to the direct-check list. Widget regenerated.
• September 7, 2026 — Like Company (C-250/25) row rebuilt with the four referred questions and the Grand Chamber hearing; the AG opinion scheduled for September 3 is recorded as scheduled, not delivered, pending retrieval of the text. Andersen v. Stability AI row expanded and flagged source_inconsistency over a disputed September 8 trial date. Third Circuit and OpenAI MDL dockets re-checked: no rulings. Widget regenerated.
• September 1, 2026 — Added an interactive case table covering all 45 case-level rows, with filters for defendant, jurisdiction, category, status and verification state; rows are static in the DOM and all outbound links carry rel="nofollow". Added the Midjourney post-mediation status conference (no settlement notice on the docket) and the Rule 72 objection on studio AI disclosure, both verified against the CourtListener docket. Third Circuit and OpenAI MDL dockets re-checked: no rulings. CSV now 46 rows.
• August 25, 2026 — Three corrections published: Kadrey certification denied (not pending), Concord I posture (summary judgment, not expert phase), Sony/UMG v. Suno works-addition motion denied. Added Hendrix and Roberson v. Apple, Elsevier et al. v. Meta, the UMG v. Suno amendment denial, and Midjourney’s Cox-based challenge to the studios’ secondary claims. Updated the OpenAI MDL row with the Cox Rule 12(c) motion, the August 6 denial of leave to amend, and the chat-log sanctions briefing. CSV now 45 rows.
• August 19, 2026 — Field-alignment repair to the OpenAI MDL CSV row (see Correction Notices). Added Tanzer and Gilmore v. Salesforce (N.D. Cal., filed October 15, 2025), closing a coverage gap in prior versions. Added Anthropic’s February 2026 motion to dismiss in Carreyrou and the co-defendants’ divergent responses. Third Circuit and OpenAI MDL dockets re-checked: no rulings entered. Archive rotation applied to entries older than 90 days. CSV now 41 rows, six flagged pending_primary.
• August 12, 2026 — Restructured to reverse-chronological entry log with collapsible archive. Added Carreyrou et al. v. Anthropic and the Bartz opt-out successor cluster. Correction issued on the Concord court and docket (see Correction Notices). Added the August Concord II partial motions to dismiss and the Kadrey §1292(b) certification motion. Introduced the Opt-Out Exposure Delta™ at an inaugural baseline of $264,697,463. Added a verification_flag column to the CSV.
• August 7, 2026 — Bartz moved to concluded on final approval. Added GEMA v. Suno, Like Company v. Google, Kneschke v. LAION. Correction issued on the Sony v. Suno schedule and on the Bartz figures.
• July 9, 2026 — V2 rebuild, 24 cases. Introduced Settlement Efficiency Ratio™, Exposure Index™, Fair Use Scoreboard.
• May 27, 2026 — Initial publication.

Frequently Asked Questions

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

They keep the right to sue individually for up to $150,000 per work under 17 U.S.C. §504(c), and 350 claimants holding 1,802 works did so. Six, led by John Carreyrou, filed against six AI developers at once in December 2025. On June 8, 2026, Judge P. Casey Pitts severed that action, holding competing LLM developers cannot be joined without a conspiracy allegation, so it now proceeds against Anthropic alone. At least four further opt-out actions are pending in the same district.

Has the Bartz v. Anthropic settlement been finally approved?

Yes. Judge Araceli Martínez-Olguín granted final approval on July 20, 2026 and entered final judgment, overruling all 53 objections. The certified Works List covers 482,460 books with a claims rate of 91.3%. Attorney fees and costs were set at $122 million and service awards at $15,000 per class representative. No payment commencement date has been announced.

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

In early August 2026 Anthropic moved to dismiss part of Concord II: the output-based portion of the direct infringement claim, arguing a complaint asserting roughly 21,000 works identifies no example of an infringing Claude output; and the DMCA claim in full, arguing CMI protection attaches to the musical compositions rather than the lyrics alone. Dario Amodei separately moved to dismiss the direct infringement claim against him personally. The motions are undecided.

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

No US appellate court has answered, and none had as of September 26, 2026. Alsup (Bartz) and Chhabria (Kadrey) found training fair use on their specific records; Bibas (Thomson Reuters v. Ross) found against it. The Third Circuit heard argument on June 11, 2026 and has not ruled. In Kadrey, Judge Chhabria denied certification for Ninth Circuit review in July 2026, leaving the Third Circuit as the only appellate proceeding on the question.

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

On July 31, 2026 the Munich District Court held Suno infringed copyright in six compositions and prohibited four acts, including training conducted in the United States. The court took jurisdiction over those US acts via a venue rule for collecting societies, applied US copyright law, and rejected the fair use defence. It also placed the burden of disproving market harm on Suno. The practical implication: training offshore does not insulate a developer from EU liability if outputs reach EU users. The judgment is not final.

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

Not before 2027. The June 30, 2026 amended scheduling order runs fact discovery to September 30, 2026 and sets dispositive motions for April 9, 2027. A separate motion by UMG and Sony to add 61,026 recordings identified through audio fingerprinting is pending before Judge Saylor.

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

No. The release is past-only, covering conduct through August 25, 2025, and does not release claims relating to AI outputs or future conduct. Anthropic must destroy the original torrented files from LibGen and PiLiMi and any derived copies. Both Concord actions, BMG’s March 2026 suit, and the entire opt-out cluster are unaffected.

What is the Opt-Out Exposure Delta?

An Axis Intelligence Research metric measuring the dollar gap between what the Bartz opt-out cohort forfeited and what it can theoretically pursue. The 1,802 opted-out works carry a $270.3 million statutory ceiling and would have drawn about $5.6 million inside the class — a delta of $264,697,463. The ratio form of the figure is the arithmetic reciprocal of the Settlement Efficiency Ratio and is disclosed as such. It is a theoretical ceiling, not a pleaded demand.

---

Go Deeper

• EU AI Act Compliance Tracker — training-data disclosure obligations and enforcement timeline
• EU AI Act Compliance Checker — assess obligations for a specific system
• AI Model Release Tracker — which models shipped, and when

Authored by Axis Intelligence Research and Sarah Mitchell, AI/ML Editor. All financial figures sourced from primary court documents, official dockets, and named institutional sources. Corrections: [email protected]. Axis Intelligence Research does not accept legal fees, retainers, or compensation from any party to the cases tracked in this report.