AI Copyright Lawsuits Over Training Data: Case Tracker With Rulings and Settlements

AI copyright lawsuits over training data are suits and regulator proceedings that challenge how an AI developer obtained or used data to train its models. This tracker records 70 of them in 12 jurisdictions as of 2026-11-09, each with its court, docket, filing date and latest decided entry, including the Third Circuit's ruling of 2026-09-29 that ROSS Intelligence's use of 2,243 Westlaw headnotes was not fair use and the Bartz v. Anthropic settlement that reached final judgment on 2026-07-20.

Rows run by jurisdiction, United States first and the others alphabetically, then by filing date. The cases form the court record behind AI training data licensing: they show which acquisition routes courts have accepted, which they have questioned, and which disputes ended in licenses.

What the case tracker covers

The tracker lists every proceeding over AI training data that has, or will have, its own page here, with the court, docket and status on a stated date. Its 70 rows sit on 51 case pages, and 11 further case pages about related conduct appear in a separate table, for 62 case pages in total. No worldwide total of AI copyright suits is given, because the published totals come from unofficial trackers that disagree with each other.

The five proceeding types in the tracker are listed below, each with example rows.

• Copyright suits over training copies: Bartz v. Anthropic, Kadrey v. Meta and GEMA v. Suno.
• DMCA 1202 claims over removed copyright management information: Raw Story v. OpenAI and Doe v. GitHub.
• Contract and scraping claims: Reddit v. Anthropic, narrowed in September 2026 to contract, interference and unfair competition claims.
• Privacy and biometric suits: In re Otter.AI Privacy Litigation and the Clearview AI settlement under Illinois BIPA.
• Regulator decisions: PIPEDA Findings #2026-002 in Canada and the ANPD's 2024 order against Meta in Brazil.

Seven rows record proceedings that a plaintiff or authority has announced but for which no court or authority source publishes a docket, such as KBS, MBC and SBS v. Naver in the Seoul Central District Court. Those rows name the forum only where the authority or plaintiff states it, and they give no status or ruling.

The case tracker: every training-data case, docket and status

The tracker orders proceedings by jurisdiction, United States first and the rest alphabetically, then by filing date within each jurisdiction. Each case page is reached from the first row that belongs to it, and every status is stated as of 2026-11-09.

The four reading keys below explain the codes and markers in the table.

• Claim codes: CT copyright over training copies; CO copyright over training and outputs; DM DMCA 1202 (copyright management information); 1201 DMCA anti-circumvention; CN contract and scraping; PR privacy or data protection; BI biometric privacy; VP voice and publicity; DL training-data disclosure law; TM trademark.
• n/r: the filing date or claim is not on record in a court source, which does not mean that no filing exists.
• Not stated: no docket, status or ruling is published on a court or authority source for that row.
• Sorting: partial dates sort to the end of their month or year, "In re" and "The" are ignored, and undated rows sort last, alphabetically.

Court rulings on training so far: fair use, fair dealing and TDM outcomes

Three US rulings have decided fair use for AI training, and they split: two found for the AI developer on their records and one held a non-generative use not fair use. The table compares the nine training rulings on record as of 2026-11-09 by court, legal test and outcome.

In Bartz v. Anthropic, the Order on Fair Use of 2025-06-23 (Dkt. 231) called training "exceedingly transformative", held the scanning of purchased print books fair use and refused to excuse pirated library copies. Two days later, in Kadrey v. Meta, summary judgment went to Meta on its record (Dkt. 598), and the court stated that the ruling "does not stand for the proposition that Meta's use ... is lawful", calling market dilution the "potentially winning argument" but finding it undeveloped. The Third Circuit then held on 2026-09-29 that ROSS's use was "minimally transformative, at best", noting that ROSS's AI "was not a generative AI", and ROSS petitioned for rehearing on 2026-11-03 (Doc. 225).

US fair use for AI training therefore separates generative from non-generative uses so far. The ROSS opinion cites a Statement of Interest that the United States filed in MDL 3143 on 2026-09-01, stating that training that can generate original responses is transformative. Three summary-judgment rulings are pending as of 2026-11-09: MDL 3143, where OpenAI filed its opposition on 2026-10-29 (Dkt. 2312); Concord I, heard on 2026-10-21 after the court asked for separate analysis of training and three output types (Dkt. 918); and In re Mosaic LLM Litigation, where cross-motions were filed on 2026-06-30. The doctrine is set out in fair use and AI training in US law.

Courts outside the United States apply their own exceptions. The Munich Regional Court held in GEMA v. OpenAI on 2025-11-11 that § 44b UrhG covers copies made to assemble a training corpus but not memorization in the model, and in GEMA v. Suno on 2026-07-31 it applied US law to training copies made in the United States and found them not fair use; neither judgment is final. The Hamburg Higher Regional Court held on 2025-12-10 that LAION's dataset copying was justified by §§ 44b and 60d UrhG, and the Federal Court of Justice pronounces on 2026-12-17. The Delhi High Court refused ANI an interim injunction on 2026-07-24, holding training prima facie fair dealing under s. 52(1)(a). In London, Getty abandoned its training claim, and the High Court held on 2025-11-04 that a model which never stored the works is not an infringing copy. Advocate General Szpunar's opinion of 2026-11-03 in C-250/25 found the training questions inadmissible, and the CJEU judgment is pending.

Settlements recorded in the tracker and what they paid

Two training-data settlements in the tracker received court approval, Bartz v. Anthropic and the Clearview biometric settlement, and the Clearview approval was later vacated on appeal; the record label settlements with Suno and Udio were filed or announced without terms.

The Bartz settlement is a USD 1.5 billion non-reversionary fund, granted final approval and judgment on 2026-07-20 (N.D. Cal. 3:24-cv-05417-AMO, Dkt. 680). The final approval order of 2026-07-20 puts the Works List at 482,460 works, estimates approximately USD 3,000 gross per work and sets attorneys' fees at USD 101,561,111. Class counsel's status report filed on 2026-09-02 (Dkt. 688) states an initial distribution of approximately USD 2,203.56 per claimed work across 449,731 claimed works, before Anthropic's final USD 450 million payment, so that figure is not the final amount per work.

The Clearview settlement, approved on 2025-03-20 in the Northern District of Illinois (1:21-cv-00135), gave the class a stake in Clearview's equity instead of a cash fund. On 2026-07-13 the Seventh Circuit (No. 25-1673, Doc. 69) vacated that approval and remanded: it accepted that a stock-based settlement carries some uncertainty, but found that no class representative had adequately represented the nationwide class, which received the smallest benefits next to state-residency subclasses (opinion, Doc. 69, checked 2026-11-10). The case is back before the district court, and no replacement settlement is recorded here.

The record labels' suits against the two music generators ended partly in settlements paired with licenses. Universal Music Group announced on 2025-10-29 that it and Udio "settled copyright infringement litigation" alongside new licenses, Warner informed the Udio court of its settlement on 2025-11-25 (Doc. 139), and Warner's claims against Suno were dismissed by stipulation after a Suno and Warner Music Group license announced on 2025-11-25 (Doc. 174, letter of 2025-12-09). None of these filings or announcements discloses a value. Sony's claims against Udio continue, and UMG and Sony filed a second action against Suno over 60,202 recordings. The licenses that followed are logged among AI data licensing deals.

Privacy and regulator proceedings over training data

Data protection authorities and privacy plaintiffs challenge the legal basis for training on personal data, and their outcomes so far run both ways.

Canada's Privacy Commissioner and the Quebec, British Columbia and Alberta offices published PIPEDA Findings #2026-002 on 2026-05-06, finding the training-data complaint against OpenAI well-founded and conditionally resolved. The offices held that OpenAI lacked implied consent for web-scraped and licensed data and that licensed datasets require that individuals were given a choice; OpenAI stated that licensed data made up less than 1% of its pre-training datasets, a figure the offices could not verify. Brazil's ANPD suspended Meta's processing of user data for AI training on 2024-07-02 (Voto nº 11/2024/DIR-MW/CD) and set a daily fine of BRL 50,000 for non-compliance.

Other outcomes favor the developer. The Higher Regional Court of Cologne rejected a consumer association's application for a preliminary injunction against Meta's AI training on 2025-05-23 (15 UKl 2/25), finding that training can serve a legitimate interest. The Tribunale di Roma upheld OpenAI's opposition to the Garante's EUR 15 million fine decision of 2024-11-02 (judgment n. 4153/2026, published 2026-03-18), and the Garante temporarily removed that decision from its website.

US privacy suits split the same way. The Otter.ai court let ECPA, CIPA 631 and BIPA claims proceed on 2026-08-13 (ECF 68), while the Google Cloud Contact Center AI court held on 2026-10-22 that training use alone is not a CIPA 631(a)(3) claim without a predicate violation. The Irish Data Protection Commission's 2024 High Court application and 2025 inquiry over Grok training appear without a status, because neither a court record number nor a decision is published. How the legal basis question applies to a holder's records is covered under GDPR and AI training data.

Dated next steps in the tracked cases

The next dated steps in the tracker are court listings and deadlines, not predicted outcomes. The six steps below run in calendar order.

1. 2026-11-11: the Getty Images v. Stability AI appeal (CA-2026-000007) has a fixed hearing in the Court of Appeal in London, estimated at 4.5 hours (Court of Appeal case tracker, checked 2026-11-10).
2. 2026-12-11: conference in Sony v. Uncharted Labs (Udio) in the Southern District of New York (Doc. 190).
3. 2026-12-17, 08:45: Federal Court of Justice pronouncement in Kneschke v. LAION (I ZR 281/25).
4. 2027-01-18: fact discovery cut-off in Getty Images v. Stability AI (US), with trial set for 2028-01-18 (Dkt. 74).
5. 2027-05-03: first trial among the cases coordinated with Kadrey v. Meta, the Entrepreneur Media case (Dkt. 879).
6. 2027-07-12: trial in Disney v. MiniMax (C.D. Cal. 2:25-cv-08768-SB-E).

How the tracker is built and which proceedings sit outside it

Every docketed row comes from a court document, an authority's own page or a party's own filing, and every status carries the date it was checked. Rows without a published docket cite the plaintiff's or authority's own announcement where one exists and show n/a where none does. Press-reported values, unofficial tracker counts and secondary-only dates stay out, following how we source, date and check every figure.

Statuses come from the latest docket entry read on the check date, and an entry that was not read is marked "not checked" rather than inferred. Settlement values appear only with the court entry that approved them, and license values are not given because none of the parties disclosed one.

Eleven related proceedings concern data acquired for AI but not training itself, so the table below lists them with the reason each sits outside the tracker, as of 2026-11-09.

The five Perplexity and Reddit rows concern copying content for AI search answers or circumventing access controls, not training a model, as of 2026-11-09. The Mercor breach litigation, McKinney v. Scale AI and noyb's five complaints concern the contributors who produce human data rather than the data used for training. EDPS v SRB sets the CJEU's test for when pseudonymized data stop being personal data, which matters for de-identified datasets, and In re Character Technologies is a chapter 11 asset sale rather than a training claim.

What the cases mean for a data holder deciding whether to license

The cases do not tell a holder whether to license, but they show what buyers' training practices have cost them and which acquisition routes courts have questioned. The three benefits, three risks, two downsides and four alternatives below are drawn from the rulings and settlements in the tracker.

Benefits:

• Lawful copies count: the Bartz court held the scanning of purchased print books fair use but did not excuse pirated library copies, and the case then settled.
• A licensing market can weigh in court: the Third Circuit stated on 2026-09-29 that "the market for licensing headnotes as text to train AI is rapidly developing".
• Disputes can end in licenses: UMG's and Warner's settlements with Udio and Warner's with Suno came with new licenses.

Risks:

• No assured licensing market: Kadrey v. Meta held on its record that the authors were "not entitled to the market for licensing their works as AI training data" (Dkt. 598).
• Training found fair use: Bartz and Kadrey decided training for the developer, which lowers an unlicensed holder's leverage in the United States.
• Personal data travels with a license: PIPEDA Findings #2026-002 require that individuals in licensed datasets were given a choice.

Downsides:

• Law still moving: the ROSS rehearing petition, the Federal Court of Justice pronouncement and the CJEU judgment in C-250/25 are pending, so a license priced now precedes them.
• No public price benchmark: none of the music settlements or their licenses discloses a value.

Alternatives:

• Not licensing at all, which keeps every right and any later claim with the holder.
• Reserving text and data mining rights in the EU, where Art. 4(3) of the DSM Directive requires a machine-readable reservation.
• Collective licensing or enforcement through a collecting society, the route GEMA took in Munich against OpenAI and Suno.
• Waiting for the pending summary judgments in MDL 3143, Concord I and Mosaic before pricing a license.

Each route, including declining every offer, is weighed in the guide on how to sell or license company data to AI companies.

Where to read the law, the deals and the buyers behind these cases

The tracker answers who sued and what was decided; the pages below answer which law applies, which deals followed and who buys data.

• AI training data law: copyright, privacy, employees and contracts by jurisdiction.
• EU rules for AI training data: GDPR, AI Act, TDM exception and Data Act.
• Who buys AI training data: every buyer and program, without fee ranking.
• AI training data market: size, prices, payouts and history.
• Data assets in insolvency: chapter 11 sales such as Character.AI's.
• Human data for AI: contributor work, pay and platforms.

Three narrower questions have their own pages: the EU text and data mining exception behind the German rulings, pirated books and shadow libraries at issue in Bartz and Kadrey, and contributor lawsuits against human data platforms such as McKinney v. Scale AI.

Frequently asked questions

Has any court ruled that AI training is fair use?

Yes, two US district court rulings found training fair use on their records, and one appellate ruling found a non-generative use not fair use. Bartz v. Anthropic (2025-06-23) and Kadrey v. Meta (2025-06-25) went for the developer, while the Third Circuit ruled against ROSS on 2026-09-29. Tests in other countries are compared under copyright and AI training across jurisdictions.

How much is the Anthropic settlement?

The Bartz v. Anthropic settlement is a USD 1.5 billion non-reversionary fund, approved with final judgment on 2026-07-20 (Dkt. 680). The order estimates approximately USD 3,000 gross per work on a Works List of 482,460 works. The claims process and distribution are covered on the Bartz v. Anthropic case page.

Which AI lawsuits are not in this tracker?

AI lawsuits that are not about training data, such as suits over chatbot harm, hiring tools, securities disclosures or defective products, are not tracked. Breaches and contributor-pay cases that involve data acquired for AI appear in the related-proceedings table. Dated breaches and complaints against buyers are logged among data buyer incidents.

How do I get tracker updates?

Tracker updates by email go out through the DataPacts newsletter, with no promised frequency. Each row also carries the date its status was checked, so a reader can see how current every entry is.

Is this tracker legal advice?

No, this tracker is information, not legal advice. It reports what courts and authorities recorded on stated dates and does not assess any holder's own position or claim.