Naming Amodei and Mann turns Anthropic's data-acquisition approval chain into discovery

The structure of the pleading is the part to read closely. Direct infringement through torrenting is pleaded against all three defendants, and contributory infringement through torrenting is pleaded against Amodei and Mann without the company attached ^[9]. Counts built that way run on facts about individuals: Sony and Warner allege that Amodei authorized the acquisition of pirated material, and that Mann personally downloaded some of it before directing employees to collect more ^[3]. Those are assertions about an approval chain, and approval chains live in ticket queues, chat logs, and whoever clicked create on the bucket.

Bartz explains why acquisition is the target rather than training. According to mezha, citing TechCrunch, the court in Bartz v. Anthropic ordered Anthropic to pay $1.5 billion, holding that using protected works to train models was permissible while obtaining that content through piracy was not ^[16]. That draws the line between two steps in the same pipeline. Sony and Warner accordingly ask for an accounting of Claude's training data and methods, and for destruction of infringing copies under court supervision ^[11]. That accounting is a report, and a pipeline either produces one per corpus or it does not.

The damages arithmetic is worth doing yourself. The publishers say the case covers tens of thousands of compositions, and they seek up to $150,000 per work for willful infringement plus up to $25,000 for each removal or alteration of copyright-management information ^[6]. Ten thousand works at the willful ceiling is $1.5 billion ^[15], the same figure as the Bartz order ^[16]. That ceiling only transfers if a court finds willfulness and counts each composition as its own work, and for now it is a request in a complaint whose allegations have not been adjudicated ^[1].

The book counts carry a provenance note of their own. The complaint alleges Mann used BitTorrent in June 2021 to take at least 5 million books from LibGen, and that employees took at least 2 million more from Pirate Library Mirror in July 2022, which is at least 7 million volumes ^[5][14]. Those totals lean on the record in Bartz, where filings described a co-founder downloading 196,640 books from Books3 before the larger acquisitions ^[10]. Discovery produced in one case became the factual spine of the next.

For anyone running ingestion, the fields that matter are the unglamorous ones: source identifier, acquisition method, license or authorization reference, approving human, date. Those map directly onto what the publishers say they will pursue, which is where training material came from, who approved acquiring it, and what licenses covered its use ^[7]. If corpora sit commingled in one archive with no per-record lineage, neither the accounting nor the destruction request can be scoped narrowly ^[11]. Provenance metadata reads as a routine record-keeping detail, but it stops being routine the moment it becomes an exhibit.

Ingestion hygiene does not cover the whole complaint. The publishers also allege scraping from authorized lyric services including MusixMatch and LyricFind, scanning of second-hand books, and collection from Common Crawl, The Pile, and Books3 ^[8], and separately that Claude reproduces protected lyrics verbatim or nearly verbatim, including after repeated prompts ^[12]. Output filtering is a different control from intake logging, and this pleading reaches both. Anthropic had not commented when the suit was filed ^[13]. In my context I would rather pay the metadata overhead at ingestion than reconstruct 2021 from chat history.