Methodology and sources

1. Library Genesis catalogs, 25 snapshots, 2015–2025. The non-fiction ("updated" table) and fiction database dumps that Library Genesis published, as preserved by the Internet Archive on the dates below. Each catalog entry describes one file; entries describing the same edition are combined into one line, and each line carries the list of snapshots in which it appears, with the count of entries in each. The April 2019 snapshot contains the non-fiction catalog only. The February 2020 snapshot combines the non-fiction dump of February 13, 2020 with the fiction dump of February 8, 2020, which survives only as a Wayback Machine capture of libgen.rs/dbdumps/; both are dated February 13, 2020 in reports. Only catalog metadata is used; no book files are held.
2. Pirate Library Mirror (PiLiMi) index of Z-Library, 2022. The metadata table of the PiLiMi mirror of Z-Library (11,783,153 records), in the metadata-only copy re-hosted on Hugging Face as P1ayer-1/annas-archive-index. Each record names the PiLiMi release that carried the file: 2,179,884 in the first release (July 2022, the release the Bartz v. Anthropic order describes Anthropic downloading that month), 3,818,910 in the second (September 2022), and 5,784,359 not mirrored (most already in LibGen). This index carries no ISBNs, so PiLiMi records match registrations by title and name only. No files from the mirror are held.
3. Z-Library records added after PiLiMi, 2022–2025. Two later catalogs of Z-Library. (a) Anna’s Archive’s index of Z-Library records (“zlib3”), in the metadata-only copy on Hugging Face (P1ayer-1/annas-zlib3-index): 2,630,955 records with Z-Library IDs 22,430,000 to 25,639,999, added to Z-Library from August 24, 2022 to August 6, 2023, each with the date added, title, author, publisher, ISBNs and the MD5 Z-Library reported for the file. (b) The Z-Library extract in the MajinBook catalogue (Mazières and Poibeau, Zenodo, CC-BY 4.0; zlibrary.jsonl.gz, 8,097,488 records), which its authors describe as Z-Library’s metadata as of January 2025, obtained through Anna’s Archive, limited to EPUB-type records (EPUB, MOBI, AZW and similar); it carries IDs, titles, authors, years, languages, publishers and ISBNs, but no dates or MD5s. Records already in PiLiMi are not repeated (a PiLiMi line notes whether the January 2025 extract still lists it); 2,352,562 records are known only from the extract, and their IDs place them after August 2023. Anna’s Archive is a search engine and mirror that combines other collections; the court record on Meta’s early-2024 download of it is summarized below. No Z-Library or Anna’s Archive files are held.
4. Internet Archive lending library. The catalog of the Internet Archive’s lending library (its inlibrary collection of scanned print books that signed-in users can borrow), read through the Archive’s public scrape API on the date given under Provenance: identifier, title, creator, ISBNs, publication year, publisher, language and the date the Archive added the item. The Internet Archive is a nonprofit library. In Hachette Book Group, Inc. v. Internet Archive, 115 F.4th 163 (2d Cir. 2024), the court held that the Archive’s scanning and lending of the 127 works at issue was not fair use. That AI developers obtained Internet Archive books is alleged in pending cases against Meta and NVIDIA; no court has found it.
5. Journal articles (DOI lists). For the free article check: the list of 62,835,101 DOIs Sci-Hub posted on March 19, 2017 (Internet Archive item scihub_doi_list_2017, doi.7z); the DOI column of LibGen’s scimag database as published in February 2020 (libgen-meta-20200213, scimag_id_doi1_doi2_isbn_md5_pmid_pmcid.tsv.gz); and the scimag database dump of August 6, 2021 (libgen-meta-20210806, libgen_scimag_dbbackup-2021-08-06.rar), which also gives the date each article was added. Each DOI is lowercased, stripped of any doi: or doi.org/ prefix, and stored only as the first 8 bytes of its SHA-256 hash with a source mask; the article files were never fetched. Article lists for a person come from Crossref’s open metadata API (by ORCID iD where the publisher deposited it, or by author name). The free check shows list membership only. The paid Journal Article Registration Report (v6.2) adds two kinds of U.S. registration that may cover an article. Publisher side: the journal issue registered as a serial, from the Office’s serial registrations (the tabular file reg_serial_2026_01.csv, registrations 2006–2025, and the serial records of its MARC XML bulk file, registrations from 1978; 3,504,756 issue registrations across 174,707 serials). An article is matched to a registration through the journal’s ISSN (or its exact title when Crossref gives no ISSN) and the issue’s volume and number, or its volume or month where the registration gives only those; a serial registration covers the articles first published in that issue that the claimant fully owned when it filed (Circular 62B). Author side: a TX registration whose main title equals the article’s, with one of the article’s author surnames among the registration’s authors or claimants and a publication year within a year before or three years after the article’s; registrations carrying an ISBN (books) are left out. The § 412 timing condition is then compared with the dated article-copying events listed on the report, each graded. Many authors assigned their copyright to the publisher, so the publisher may be the owner who can sue.
6. Books3 file-name list. A list of 196,608 Books3 file names published by Peter Schoppert (psmedia/Books3Info, filenames.txt). Books3 as released is reported to contain 196,640 books; the published list is 32 names shorter. No Books3 texts are held.
7. Books3 ISBN list. 129,635 unique ISBN-13s that Schoppert extracted from the Books3 texts (Books2_ISBNs.txt, same repository). His README describes a pattern search over the texts that collected ISBN-13s only.
8. U.S. Copyright Office registrations, 1978–2025. The Office's public tabular data for non-dramatic literary works (TX) as published in January 2026 (reg_non_dramatic_literary_work_2026_01.csv, 9.28 million rows, registrations from January 1, 1978 through 2025). These are public U.S. government records; PayAuthors is not affiliated with, or endorsed by, the U.S. Copyright Office.

LibGen snapshot dates and Internet Archive items

Each report lists the SHA-256 checksum of every source file it was built from, and the dataset build it came from. The full provenance table (archive item, file name, size, the archive's own SHA-1, our SHA-256, and the checksum of each parsed catalog) is published with each dataset build at the end of this page.

A checksum identifies the exact bytes of a file. By itself it does not establish a file's origin, completeness or authenticity.

An MD5 identifies exact file contents, so an entry can be compared with a defendant’s file lists, hashes or logs produced in discovery. A match there would tie a work to a specific file. A catalog listing alone does not show that anyone obtained that file.

Torrents. LibGen distributes its files in torrents of 1,000 consecutive IDs, each named for the first ID in its block: non-fiction r_000, r_1000, r_2000 and so on, fiction f_0, f_1000 and so on. We checked this against the Internet Archive’s crawls of both torrent directories (November 20, 2020: 2,817 non-fiction and 2,223 fiction torrent names, every one a multiple of 1,000) and against a court filing that places two fiction files, identified by MD5, in fiction torrent “1000.” Our catalog lists both, as fiction IDs 1,286 and 1,291 (class plaintiffs’ Rule 56.1 statement ¶¶249, 251, In re OpenAI Copyright Litigation, S.D.N.Y. 1:25-md-03143, Dkt. 1987; an allegation). Each LibGen file line shows its torrent, worked out from its ID; confirm it against the torrent itself. PiLiMi lines show the PiLiMi torrent named in the PiLiMi index.

Torrent pieces. A torrent file lists every file it carries, with its size, and divides the combined data into pieces of a fixed length; BitTorrent peers exchange whole pieces. From the torrent files’ own metadata (file names, sizes and piece length; no book text), each file line shows the file’s size, the piece length, how many pieces the file spans, and the largest share of the file that lies inside one piece. A file that lies inside one piece is sent whole whenever that piece is sent. LibGen torrents come from the Internet Archive’s November 20, 2020 crawls and from 1,920 later block torrents captured one by one by the Wayback Machine at libgen.rs and libgen.is; PiLiMi torrents come from the PiLiMi torrent archive (pilimi-zlib-meta). Each LibGen file is located in its torrent by MD5 and each PiLiMi file by Z-Library ID. Piece data covers 99.6% of the non-fiction and 98.9% of the fiction files in the June 5, 2021 catalogs, and 95.2% and 84.0% of those in the June 3, 2025 catalogs, because some later torrents were never captured. Every PiLiMi release 1 file has piece data; release 2 was distributed as whole archive files, so its pieces do not follow book files and none is shown. Of the files located, 79% of fiction files, 26% of non-fiction files and 74% of PiLiMi release 1 files lie inside a single piece, and 97%, 61% and 93% have at least half of their data inside one piece. No torrent was joined.

EPUB and MOBI. File lines mark EPUB and MOBI files, because OpenAI’s 2018 download script is alleged to have taken only those two types (same filing, ¶147). The mark is information only; it does not change the statutory-damages timing mark.

Added dates. Each LibGen entry carries the date LibGen says the file was added, shown as “added to the catalog … (LibGen’s own record).” Between May 12, 2017 and February 12, 2021 the only fiction catalog that survives is the February 8, 2020 dump (a Wayback Machine capture of libgen.rs/dbdumps/fiction_2020-02-08.rar, 801,097,717 bytes, whose SHA-1 matches the digest the Wayback Machine recorded for the capture). Directory listings captured in 2019 show weekly fiction backups, but none of those files was captured. For fiction entries between those dates, the added date and the February 2020 dump are the only records of when an entry appeared. Entries deleted before a catalog we hold do not appear at all. As a check on the added dates: LibGen assigns IDs in order, and in our February 12, 2021 fiction catalog the added dates pass October 31, 2019 at about ID 2,190,000, below which the catalog holds 2,189,999 entries; the class plaintiffs’ filing puts LibGen fiction at 2,182,642 files in October 2019 (¶¶137, 236). In the February 13, 2020 non-fiction catalog the added dates pass January 31, 2020 at about ID 2,470,000, with 2,469,999 entries below it; the filing puts non-fiction at about 2.46 million as of January 2020 (¶137). Both agree within about half a percent (the ID cut-off is estimated to the nearest 10,000). The February 8, 2020 fiction dump allows a direct count: 2,185,033 of its 2,206,073 entries carry an added date on or before October 31, 2019, 0.1% above the filing’s 2,182,642, and 2,205,667 carry one on or before January 31, 2020.

Records are matched to Copyright Office TX registrations by the v6.3 rule (October 2, 2026), strongest first: (1) the same ISBN on both sides, kept only when the registration also shares an author or claimant surname, pen name or publisher name, or most of the title, because one ISBN is often printed on several editions, formats or an anthology; (2) the same full title plus an author or claimant surname; (3) the same main title (before a colon, slash or dash), or the book’s own title after a series name or number, plus surname, with safeguards against series names and disagreeing subtitles; (4) the same title words with stop words ignored, plus surname, only when nothing else matched; (5) the same full title plus the catalog’s publisher named on a company registration (a work made for hire). Titles are compared after routine normalization (case, punctuation, leading articles and edition notes), and common first names and role words are not treated as surnames. A match is set aside when the registration is for a translation, a study guide or an adaptation of the work (a play or graphic novel, say), or when volume, issue or edition numbers disagree. A record lists up to eight registrations of the work, chosen by how close their dates are to the record’s. Measured accuracy is below. Tiers 2 to 5 are heuristic: a match means a registration exists for that title and name, not that the catalog copy is that exact edition. Every match is a candidate to confirm against the linked Office record. TX registrations from 1978 through 2025 are covered, as published by the Office in January 2026; sound recordings such as audiobooks and renewals are not. Statutory-damages timing mark. For each matched registration the report compares the registration's effective date with each documented download date on the timeline (Anthropic: Books3 early 2021, LibGen June 2021, PiLiMi July 2022; Meta: October 2022 and after) and with the publication date in the registration. The screen sorts each comparison into one of four cases. (1) Known start: a court record, company statement or complaint gives when the download period began. (2) Catalog lower bound: the library’s own record of when that copy was added (LibGen’s or Z-Library’s record date; for Books3, its October 2020 release). This is catalog metadata as the library reported it, not independent proof of when anyone acquired the file. The earliest date the copying could have begun is the later of (1) and (2); for an event known only by an outer bound (a paper, model release or complaint date), only (2) is available. A registration dated before that earliest date passes the screen. (3) Publication route: a registration made within three months of a publication that came before that earliest date also passes (§ 412(2)). (4) Insufficient dates: a registration dated after the earliest date but before the end of the download period, or a comparison with no earliest date at all, is shown as “dates overlap or download date unknown” and never counted. The screen is deliberately conservative: it may leave out works for which § 412 would still allow statutory damages, for example a registration made within three months of publication where no earliest copying date is known. The condition comes from 17 U.S.C. § 412, which has exceptions this screen does not assess. The journal-article report uses the same four cases, with LibGen scimag’s record date as the catalog lower bound. For counts and dollar figures, copies, formats, editions and translations of one book count as one work, timed from its earliest copy in the dataset. It is a date screen, not a finding that the registration covers the listed edition, that the work was in the download, or that damages are available. Nor does it address whether a claim is timely: copyright claims must generally be brought within three years after they accrue (17 U.S.C. § 507(b)). In suits by authors who opted out of the Bartz v. Anthropic settlement, Anthropic argues (motion to dismiss, September 24, 2026) that claims based on its 2021–2022 downloads are time-barred and that the class action did not suspend the period for all of them; the court has not ruled. Limitations, the discovery rule, class-action tolling and the Bartz release should be checked with counsel for each work. Plaintiffs have begun to define classes by the same timing: the textbook-author classes proposed in Sullivan v. Meta (N.D. Cal., filed July 2, 2026) and Sullivan v. OpenAI Foundation (S.D.N.Y., filed August 14, 2026) cover only works registered within five years of first publication and before the defendant's copying, or within three months of first publication. Neither class has been certified.

For an event dated as a period (for example, OpenAI’s alleged torrenting from September 2019 through January 2020), the timeline also reports the first LibGen dump after the period ends, if it falls within 45 days, and how many of the lines it lists had a catalog entry added on or before the period’s last day (LibGen’s own record). For OpenAI’s alleged period that is the February 2020 pair of dumps (fiction February 8, non-fiction February 13), 8 and 13 days after the period ends. For the event dated only to 2018, the timeline also counts the EPUB and MOBI files listed in the snapshot compared, because the 2018 download script is alleged to have taken only those two types.

The Atlantic’s LibGen search uses a snapshot “taken in January 2025, after Meta is known to have accessed the database, so some titles here would not have been available to download,” and notes that “it’s impossible to know exactly which parts of LibGen Meta used to train its AI, and which parts it might have decided to exclude.” The Atlantic (a subscription may be required); the same caveat is quoted on Bruce Schneier’s blog. For Books3, The Atlantic’s Alex Reisner explained that the books are stored “as large, unlabeled blocks of text,” so he “extracted ISBNs from these blocks of text and looked them up in a book database”; of the 191,000 titles he identified, 183,000 had associated author information. The Atlantic, September 25, 2023.

The two approaches answer different questions. Ours documents what each dated catalog snapshot and a published file list contained, so a title added after a download shows up with the date it first appeared; a single later snapshot cannot separate those. A defendant’s own acquisition records, file manifests or logs would be more direct evidence of what that defendant obtained. Counts also differ because sources count different things: files, catalog records, identified titles, editions, or books with author metadata.

Huckabee v. Bloomberg (S.D.N.Y., November 24, 2025). The court denied Bloomberg’s motion to dismiss. The plaintiffs identified their works, alleged the works were in Books3, and relied on Bloomberg’s own paper saying it trained BloombergGPT on Books3. The court rejected the argument that the plaintiffs had to specify which files were theirs. This was a pleading-stage ruling, not a finding of liability. Decision.

Bartz v. Anthropic (N.D. Cal.). A June 2025 order described evidence that an Anthropic cofounder downloaded Books3 in early 2021, that Anthropic downloaded at least five million books from LibGen in June 2021, and at least two million from the Pirate Library Mirror in July 2022. The court treated building a central library from pirated books differently from using copies to train models. The case settled, with final approval in July 2026; the settlement does not resolve other companies’ conduct. 2025 order; final approval order.

Kadrey v. Meta (N.D. Cal.). The June 2025 opinion describes Meta downloading LibGen in October 2022 and using Books3 in training data, and granted Meta summary judgment on the training claims on that record; the judge said the ruling “stands only for the proposition that these plaintiffs made the wrong arguments and failed to develop a record.” Distribution through BitTorrent was left for separate proof. The opinion also describes Meta downloading Anna’s Archive, a compilation that includes LibGen and Z-Library, in early 2024, and says it is undisputed that Meta torrented LibGen and Anna’s Archive (Dkt. 598 at 11–13). In March 2026 the court granted leave to add a contributory infringement claim based on BitTorrent uploading. In its July 13, 2026 Answer in Elsevier v. Meta (S.D.N.Y. 1:26-cv-03689, Dkt. 78), Meta admits obtaining data from LibGen, Anna’s Archive and Sci-Hub to train some Llama models, and obtaining Anna’s Archive data in 2024 (¶¶82, 97); that is a company statement in a pleading, not a court finding. 2025 opinion; March 2026 order.

NVIDIA. A court allowed claims to proceed past dismissal on allegations connecting Books3, The Pile and a model-training dataset; that was not a finding that NVIDIA copied any particular book. The amended complaint alleges that NVIDIA sought high-speed access to Anna’s Archive in August 2023 and that management approved within a week, and that the offer included millions of Internet Archive lending books (ECF 235 ¶¶53–57). NVIDIA withdrew its motion to dismiss as to Anna’s Archive, Z-Library, LibGen and Sci-Hub (ECF 271, March 6, 2026), and its June 2026 Answer admits downloading data for AI research from annas-archive.org/datasets and libgen.rs, without dates, while denying that it improperly copied the plaintiffs’ works (ECF 321 ¶¶58, 77). Order.

DeepSeek. DeepSeek’s own DeepSeek-VL paper (arXiv 2403.05525, March 8, 2024, §2.1) says it cleaned 860,000 English and 180,000 Chinese e-books from Anna’s Archive for its document-OCR training data. That is a company statement; we know of no U.S. court ruling on it. Paper.

OpenAI. A November 2025 discovery order described as undisputed that an OpenAI employee downloaded books from LibGen in 2018; other details about derived datasets came from testimony and party assertions. Order.

Where a company’s own statements or a court record show it copied an entire collection, what an author then needs is the link between the work and that collection: the work, its registration, and the specific entry or file name in a dated source. That is the link a PayAuthors report is built to document. What the company copied, and when, has to come from the company’s own records, statements or court findings.

Added in this build (v6)

Z-Library records after PiLiMi, the Internet Archive lending catalog and the journal-article DOI lists. The two scimag files were read as a stream and not stored; their checksums are the Internet Archive’s.

Every source file behind the current dataset build. LibGen items link to their Internet Archive pages; “SHA-1 (archive)” is the checksum the Internet Archive publishes for the file (for the Wayback Machine capture, the capture’s recorded payload digest), which we matched before use, and SHA-256 is our own. Only catalog metadata was read from these files. The file IDs and MD5 checksums shown on report lines were collected in a second pass over the same files, each downloaded again and matched to the same SHA-1 and SHA-256 (the February 8, 2020 fiction dump, added later, in a single pass).

This page is general information, not legal advice. Court records and source datasets may change or be interpreted differently as litigation proceeds.