Loading
Technology
10 min read
Areas of Expertise :
Artificial Intelligence, Machine Learning, Digital Forensics, Copyright, Economic Damages
Published Date:
Last Modified Date:
Jurisdiction:
Federal. Ninth Circuit
Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson filed Bartz v. Anthropic PBC on August 19, 2024, in the U.S. District Court for the Northern District of California. The authors alleged that Anthropic obtained unauthorized digital copies of copyrighted books, reproduced those files, and used selected works while developing its Claude large language models. Evidence produced during the litigation described a central research library containing millions of books acquired through purchases, print-book scanning, and downloads from unauthorized online libraries.
On June 23, 2025, the court divided Anthropic’s conduct into separate categories rather than applying one conclusion to every act of copying. It held that the use of the named authors’ works to train the large language models at issue qualified as fair use on the summary-judgment record. The court also found fair use for certain digital replacement copies made from purchased print books. Anthropic did not establish fair use, however, for books downloaded from pirate websites and retained in its central library. The parties later reached a $1.5 billion class settlement covering 482,460 works, and the court granted final approval and entered judgment on July 20, 2026.
Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson filed a proposed class action against Anthropic PBC in federal court in San Francisco. The complaint alleged that Anthropic obtained unauthorized copies of copyrighted books and reproduced those works while developing and training its artificial intelligence systems.
The claims concerned complete books rather than isolated excerpts. The authors alleged that Anthropic acquired copyrighted works through unauthorized online sources and used copies of selected books in connection with its Claude models. They asserted that the downloading and copying violated their exclusive rights under the Copyright Act.
The litigation examined how Anthropic acquired, organized, converted, stored, and used books within its internal systems. The record addressed downloads from Books3, Library Genesis, and Pirate Library Mirror, as well as print books purchased for scanning and conversion into digital copies.
Anthropic maintained a central research library that served purposes extending beyond immediate model training. Some books were selected from that collection for use in training datasets, while others remained stored in the library. These facts required the court to examine acquisition, reproduction, storage, format conversion, and model training as separate acts.
Anthropic argued that training involved computational analysis of books to develop models capable of producing new text. The authors responded that a downstream training use did not excuse the earlier acquisition and retention of unauthorized source files.
U.S. District Judge William Alsup granted Anthropic summary judgment on its use of the named authors’ books to train the large language models at issue. The court found the training use transformative because the books were used to develop systems capable of generating new text rather than to distribute substitute copies of the authors’ works to readers.
The ruling was tied to the claims and evidence before the court. The authors had not alleged that Claude outputs reproduced protected passages from their books. The order therefore did not resolve claims involving allegedly infringing model outputs or establish that every use of copyrighted material for AI development qualifies as fair use.
The court also addressed print books that Anthropic lawfully purchased, scanned, and destroyed after creating one-to-one digital replacements. It found that the format conversion qualified as fair use under the circumstances presented because Anthropic had purchased each physical copy and retained a replacement rather than both the print and digital versions.
The result differed for books downloaded from pirate websites and retained in Anthropic’s central library. The court concluded that Anthropic had not established fair use for acquiring and permanently storing those unauthorized copies. The later selection of some books for model training did not justify obtaining pirated source files or keeping them in a broader research collection.
Only Anthropic had moved for summary judgment on these issues. The court therefore did not enter judgment establishing infringement liability for the authors. The claims involving the pirated library copies, along with questions of liability and damages, remained for later proceedings.
Anthropic agreed to establish a settlement fund of at least $1.5 billion to resolve defined class claims involving books downloaded from unauthorized online libraries. The covered-works list ultimately contained 482,460 works, producing a gross allocation of approximately $3,100 per work before deductions and ownership allocations.
The court granted preliminary approval on September 25, 2025. The settlement established procedures for notice, claims, objections, exclusions, ownership disputes, competing submissions, and distribution of the fund.
The agreement did not constitute a judicial ruling that Anthropic was liable for infringement. It resolved specified past claims through a negotiated settlement and did not grant Anthropic a prospective license to use class members’ books.
The deadline to object to the settlement or request exclusion from the class was February 9, 2026. Eligible rights holders were required to submit claims by March 30, 2026.
Claimants could search the official Works List to determine whether a title was included. Eligibility depended on the work’s presence on that list and the claimant’s qualifying copyright interest. When multiple parties claimed rights in the same work, the allocation process accounted for ownership divisions and competing claims.
U.S. District Judge Araceli Martínez-Olguín conducted the final approval hearing on May 14, 2026. The hearing addressed the settlement amount, class-member participation, objections, exclusions, distribution procedures, attorneys’ fees, litigation expenses, service awards, and settlement-administration costs.
An updated claims report stated that claims had been submitted for 440,490 of the 482,460 works on the Works List as of April 16, 2026. That figure represented approximately 91.3% of the listed works.
After the hearing, the court took the approval requests under submission.
Judge Martínez-Olguín granted final approval of the $1.5 billion settlement on July 20, 2026, and entered final judgment. The settlement resolved covered claims involving qualifying books allegedly downloaded from unauthorized sources and retained by Anthropic.
The court awarded $101,561,111 in attorneys’ fees, with 10% withheld pending post-distribution accounting and possible adjustment. It approved reimbursement of $2,635,197.46 in incurred litigation expenses and established an $18.22 million reserve for anticipated settlement-administration and related costs, subject to additional court approval. Each of the three class representatives received a $15,000 service award.
The action was dismissed with prejudice, subject to the court’s continuing authority over settlement implementation and administration. Dismissal with prejudice prevents the covered class claims from being filed again, but the scope of that effect is governed by the settlement’s release provisions.
The settlement’s contractual Effective Date does not necessarily occur on the date of final approval. Under the agreement, the Effective Date follows entry of final approval and judgment and the expiration or resolution of applicable appellate proceedings.
Court:
United States District Court for the Northern District of California
Case Name:
Bartz et al. v. Anthropic PBC
Copyright Infringement
Data Analysis
Computer Data
Data Management
Publishing
Licensing
Damages Calculations
Big Data
Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson filed Bartz v. Anthropic PBC on August 19, 2024, in the U.S. District Court for the Northern District of California. The authors alleged that Anthropic obtained unauthorized digital copies of copyrighted books, reproduced those files, and used selected works while developing its Claude large language models. Evidence produced during the litigation described a central research library containing millions of books acquired through purchases, print-book scanning, and downloads from unauthorized online libraries.
On June 23, 2025, the court divided Anthropic’s conduct into separate categories rather than applying one conclusion to every act of copying. It held that the use of the named authors’ works to train the large language models at issue qualified as fair use on the summary-judgment record. The court also found fair use for certain digital replacement copies made from purchased print books. Anthropic did not establish fair use, however, for books downloaded from pirate websites and retained in its central library. The parties later reached a $1.5 billion class settlement covering 482,460 works, and the court granted final approval and entered judgment on July 20, 2026.
Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson filed a proposed class action against Anthropic PBC in federal court in San Francisco. The complaint alleged that Anthropic obtained unauthorized copies of copyrighted books and reproduced those works while developing and training its artificial intelligence systems.
The claims concerned complete books rather than isolated excerpts. The authors alleged that Anthropic acquired copyrighted works through unauthorized online sources and used copies of selected books in connection with its Claude models. They asserted that the downloading and copying violated their exclusive rights under the Copyright Act.
The litigation examined how Anthropic acquired, organized, converted, stored, and used books within its internal systems. The record addressed downloads from Books3, Library Genesis, and Pirate Library Mirror, as well as print books purchased for scanning and conversion into digital copies.
Anthropic maintained a central research library that served purposes extending beyond immediate model training. Some books were selected from that collection for use in training datasets, while others remained stored in the library. These facts required the court to examine acquisition, reproduction, storage, format conversion, and model training as separate acts.
Anthropic argued that training involved computational analysis of books to develop models capable of producing new text. The authors responded that a downstream training use did not excuse the earlier acquisition and retention of unauthorized source files.
U.S. District Judge William Alsup granted Anthropic summary judgment on its use of the named authors’ books to train the large language models at issue. The court found the training use transformative because the books were used to develop systems capable of generating new text rather than to distribute substitute copies of the authors’ works to readers.
The ruling was tied to the claims and evidence before the court. The authors had not alleged that Claude outputs reproduced protected passages from their books. The order therefore did not resolve claims involving allegedly infringing model outputs or establish that every use of copyrighted material for AI development qualifies as fair use.
The court also addressed print books that Anthropic lawfully purchased, scanned, and destroyed after creating one-to-one digital replacements. It found that the format conversion qualified as fair use under the circumstances presented because Anthropic had purchased each physical copy and retained a replacement rather than both the print and digital versions.
The result differed for books downloaded from pirate websites and retained in Anthropic’s central library. The court concluded that Anthropic had not established fair use for acquiring and permanently storing those unauthorized copies. The later selection of some books for model training did not justify obtaining pirated source files or keeping them in a broader research collection.
Only Anthropic had moved for summary judgment on these issues. The court therefore did not enter judgment establishing infringement liability for the authors. The claims involving the pirated library copies, along with questions of liability and damages, remained for later proceedings.
Anthropic agreed to establish a settlement fund of at least $1.5 billion to resolve defined class claims involving books downloaded from unauthorized online libraries. The covered-works list ultimately contained 482,460 works, producing a gross allocation of approximately $3,100 per work before deductions and ownership allocations.
The court granted preliminary approval on September 25, 2025. The settlement established procedures for notice, claims, objections, exclusions, ownership disputes, competing submissions, and distribution of the fund.
The agreement did not constitute a judicial ruling that Anthropic was liable for infringement. It resolved specified past claims through a negotiated settlement and did not grant Anthropic a prospective license to use class members’ books.
The deadline to object to the settlement or request exclusion from the class was February 9, 2026. Eligible rights holders were required to submit claims by March 30, 2026.
Claimants could search the official Works List to determine whether a title was included. Eligibility depended on the work’s presence on that list and the claimant’s qualifying copyright interest. When multiple parties claimed rights in the same work, the allocation process accounted for ownership divisions and competing claims.
U.S. District Judge Araceli Martínez-Olguín conducted the final approval hearing on May 14, 2026. The hearing addressed the settlement amount, class-member participation, objections, exclusions, distribution procedures, attorneys’ fees, litigation expenses, service awards, and settlement-administration costs.
An updated claims report stated that claims had been submitted for 440,490 of the 482,460 works on the Works List as of April 16, 2026. That figure represented approximately 91.3% of the listed works.
After the hearing, the court took the approval requests under submission.
Judge Martínez-Olguín granted final approval of the $1.5 billion settlement on July 20, 2026, and entered final judgment. The settlement resolved covered claims involving qualifying books allegedly downloaded from unauthorized sources and retained by Anthropic.
The court awarded $101,561,111 in attorneys’ fees, with 10% withheld pending post-distribution accounting and possible adjustment. It approved reimbursement of $2,635,197.46 in incurred litigation expenses and established an $18.22 million reserve for anticipated settlement-administration and related costs, subject to additional court approval. Each of the three class representatives received a $15,000 service award.
The action was dismissed with prejudice, subject to the court’s continuing authority over settlement implementation and administration. Dismissal with prejudice prevents the covered class claims from being filed again, but the scope of that effect is governed by the settlement’s release provisions.
The settlement’s contractual Effective Date does not necessarily occur on the date of final approval. Under the agreement, the Effective Date follows entry of final approval and judgment and the expiration or resolution of applicable appellate proceedings.
Artificial Intelligence and Machine Learning Expert Witness
An artificial intelligence expert may explain how training files are selected, tokenized, processed, and used to modify model parameters. The analysis may address model architecture, memorization, output generation, and whether a generated response reproduces protected expression from a particular work.
Digital Forensics and Data Analysis Expert Witness
A digital forensics expert may trace download records, metadata, access histories, storage locations, deletion activity, and transfers between internal systems. This work can help determine where a file originated, how long it was retained, whether it was copied, and whether it entered a training dataset.
Publishing and Content Licensing Expert Witness
A publishing and licensing expert may examine copyright ownership, royalty practices, customary license terms, digital-edition markets, and collective licensing arrangements. The analysis may also address whether licenses for AI training were available during the period covered by the claims and how comparable rights were valued.
Copyright Damages and Economic Expert Witness
An economic expert may assess actual damages, licensing values, market effects, and profits attributable to particular acts of infringement. The analysis may account for copyright-registration dates, ownership divisions, the number of works involved, and whether damages should be evaluated by work, edition, file, or asserted act of copying.
The Fair-Use Ruling Concerned Particular Uses, Not Every Act of Copying
The court found fair use for Anthropic’s use of the named authors’ books in the large language model training processes addressed by the record. This ruling did not automatically extend to how Anthropic acquired, stored, or reproduced all source files. Fair use is evaluated case by case, so different uses of the same work can lead to different legal outcomes.
Lawful Acquisition and Fair Use Were Separate Questions
The court accepted that model training could be transformative, but it did not treat that as justification for obtaining books from pirate sources. It distinguished between how a work is used and how it is acquired. As a result, storing pirated books in a central library was treated as a separate issue from training.
The Print-Book Ruling Depended on a Specific Replacement-Copy Process
Anthropic scanned purchased print books, created digital replacements, and destroyed the originals. The court treated this as a limited format conversion tied to those specific facts. It did not create a general rule that all book scanning is lawful.
Failure to Establish Fair Use Was Not a Final Infringement Judgment
The court rejected Anthropic’s fair-use defense for pirated library copies, but it did not enter final judgment for the authors. Liability and damages were left unresolved because no final cross-motion was decided. The settlement later resolved those claims without a trial or admission of liability.
The Settlement Release Was Narrower Than the Entire Dispute Over AI and Copyright
The settlement covered specific past uses of qualifying books obtained before August 25, 2025. It did not authorize future use of the works or resolve claims about AI-generated outputs. Conduct after the cutoff date also remained outside the release.
Bartz v. Anthropic separated the use of books for large language model training from the methods used to acquire and retain those books. The June 2025 order found fair use for the training uses before the court and for a defined process involving digital replacements of purchased print books. Anthropic did not establish fair use for unauthorized digital copies kept in its central library, leaving those claims for further litigation.
The $1.5 billion settlement resolved the remaining covered claims without a trial on infringement or damages. On July 20, 2026, the court granted final approval, entered judgment, dismissed the covered action with prejudice, and retained authority over settlement administration. The agreement addressed specified past inputs and source files while excluding claims involving past AI outputs and conduct after the release cutoff.
What Books and Online Libraries Were Involved in Bartz v. Anthropic?
The record addressed books obtained through sources including Books3, Library Genesis, and Pirate Library Mirror. The settlement was limited to qualifying works on the approved Works List associated with LibGen or PiLiMi copies downloaded by Anthropic.
Did the Court Hold That Training Claude Was Fair Use?
Yes, for the training uses and summary-judgment record before it. The ruling did not decide claims based on Claude outputs reproducing protected material because the named authors had not brought that type of output claim.
Did the Court Find Anthropic Liable for Pirating Books?
The court held that Anthropic had not established fair use for pirated copies retained in its central library. It did not enter final infringement judgment for the authors because they had not moved for summary judgment on liability.
Why Were Scanned Print Books Treated Differently?
Anthropic had purchased the print copies, scanned them, and destroyed the physical books after creating digital replacements. The court treated the process as a one-to-one format conversion under the facts presented rather than as the creation of additional copies for distribution.
Did the Settlement Give Anthropic a License for Future AI Training?
No. The settlement resolved defined claims based on specified past conduct. It did not grant Anthropic a license to use the works after the settlement’s cutoff date.
When Will Settlement Payments Be Issued?
Payments can be issued only after the settlement becomes effective under the agreement. The Effective Date follows final approval and judgment and the expiration or resolution of applicable appellate proceedings.
Artificial Intelligence and Machine Learning Expert Witness
An artificial intelligence expert may explain how training files are selected, tokenized, processed, and used to modify model parameters. The analysis may address model architecture, memorization, output generation, and whether a generated response reproduces protected expression from a particular work.
Digital Forensics and Data Analysis Expert Witness
A digital forensics expert may trace download records, metadata, access histories, storage locations, deletion activity, and transfers between internal systems. This work can help determine where a file originated, how long it was retained, whether it was copied, and whether it entered a training dataset.
Publishing and Content Licensing Expert Witness
A publishing and licensing expert may examine copyright ownership, royalty practices, customary license terms, digital-edition markets, and collective licensing arrangements. The analysis may also address whether licenses for AI training were available during the period covered by the claims and how comparable rights were valued.
Copyright Damages and Economic Expert Witness
An economic expert may assess actual damages, licensing values, market effects, and profits attributable to particular acts of infringement. The analysis may account for copyright-registration dates, ownership divisions, the number of works involved, and whether damages should be evaluated by work, edition, file, or asserted act of copying.
The Fair-Use Ruling Concerned Particular Uses, Not Every Act of Copying
The court found fair use for Anthropic’s use of the named authors’ books in the large language model training processes addressed by the record. This ruling did not automatically extend to how Anthropic acquired, stored, or reproduced all source files. Fair use is evaluated case by case, so different uses of the same work can lead to different legal outcomes.
Lawful Acquisition and Fair Use Were Separate Questions
The court accepted that model training could be transformative, but it did not treat that as justification for obtaining books from pirate sources. It distinguished between how a work is used and how it is acquired. As a result, storing pirated books in a central library was treated as a separate issue from training.
The Print-Book Ruling Depended on a Specific Replacement-Copy Process
Anthropic scanned purchased print books, created digital replacements, and destroyed the originals. The court treated this as a limited format conversion tied to those specific facts. It did not create a general rule that all book scanning is lawful.
Failure to Establish Fair Use Was Not a Final Infringement Judgment
The court rejected Anthropic’s fair-use defense for pirated library copies, but it did not enter final judgment for the authors. Liability and damages were left unresolved because no final cross-motion was decided. The settlement later resolved those claims without a trial or admission of liability.
The Settlement Release Was Narrower Than the Entire Dispute Over AI and Copyright
The settlement covered specific past uses of qualifying books obtained before August 25, 2025. It did not authorize future use of the works or resolve claims about AI-generated outputs. Conduct after the cutoff date also remained outside the release.
Bartz v. Anthropic separated the use of books for large language model training from the methods used to acquire and retain those books. The June 2025 order found fair use for the training uses before the court and for a defined process involving digital replacements of purchased print books. Anthropic did not establish fair use for unauthorized digital copies kept in its central library, leaving those claims for further litigation.
The $1.5 billion settlement resolved the remaining covered claims without a trial on infringement or damages. On July 20, 2026, the court granted final approval, entered judgment, dismissed the covered action with prejudice, and retained authority over settlement administration. The agreement addressed specified past inputs and source files while excluding claims involving past AI outputs and conduct after the release cutoff.
What Books and Online Libraries Were Involved in Bartz v. Anthropic?
The record addressed books obtained through sources including Books3, Library Genesis, and Pirate Library Mirror. The settlement was limited to qualifying works on the approved Works List associated with LibGen or PiLiMi copies downloaded by Anthropic.
Did the Court Hold That Training Claude Was Fair Use?
Yes, for the training uses and summary-judgment record before it. The ruling did not decide claims based on Claude outputs reproducing protected material because the named authors had not brought that type of output claim.
Did the Court Find Anthropic Liable for Pirating Books?
The court held that Anthropic had not established fair use for pirated copies retained in its central library. It did not enter final infringement judgment for the authors because they had not moved for summary judgment on liability.
Why Were Scanned Print Books Treated Differently?
Anthropic had purchased the print copies, scanned them, and destroyed the physical books after creating digital replacements. The court treated the process as a one-to-one format conversion under the facts presented rather than as the creation of additional copies for distribution.
Did the Settlement Give Anthropic a License for Future AI Training?
No. The settlement resolved defined claims based on specified past conduct. It did not grant Anthropic a license to use the works after the settlement’s cutoff date.
When Will Settlement Payments Be Issued?
Payments can be issued only after the settlement becomes effective under the agreement. The Effective Date follows final approval and judgment and the expiration or resolution of applicable appellate proceedings.
Copyright Infringement
Data Analysis
Computer Data
Data Management
Publishing
Licensing
Damages Calculations
Big Data
Technology
Patent, +4
Technology
Artificial Intelligence, +4
Technology