$50 per work · no minimum.
Each work comes with its registration, catalog lines and § 412 timing, reviewed against the Copyright Office record, as a spreadsheet and PDF exhibits. Rate card · Request a quote for custom work
What we produce for each work
One line from a live report, with five factual fields useful in work-level pleading and exhibit analysis. Every report and defendant list is built from lines like this.
The Lost Night
- LibGen fiction ID 2157462 · epub · MD5
8b2a0eab82719cfd5ab89c3e4fd6e3de - torrent
f_2157000· file 12 MiB, spans 4 pieces of 4.0 MiB, so no one piece holds the whole file (the largest part in one piece is 33% of the file)
- Publication: published Feb. 26, 2019
- Authorship: Bartz, Andrea — author
- Claimants: Andrea Bartz Inc., (organization)
- Recorded documents: none in the Office’s bulk data through June 26, 2025
- Catalog entry: the record as LibGen listed it, the date it was added, and which of 25 dated catalogs list it.
- Exact file: catalog number and format, with the MD5, the fingerprint of one specific file, where the catalog records it.
- Torrent: the named torrent that carried the file, and how the file sits in its pieces. When a file lies inside one piece, every verified copy of that piece holds the whole file; whether one uploader sent all of it has to come from that uploader’s records.
- Registration: the Copyright Office record, with the claimant named on it (the owner when registered, not necessarily today), dates and any recorded transfers.
- § 412 timing against each defendant’s documented or alleged download date. Separately, a U.S. work generally can’t be sued on until the Copyright Office registers or refuses it (17 U.S.C. § 411(a)). Works first published outside the United States are exempt from that requirement but not from § 412, so they can still claim actual damages and are not in our “works” counts. Our registration data runs through 2025.
Each work is also screened against the Bartz Works List criteria.
The fight in the AI copying cases has moved from whether a company copied a pirate library to which works were in it, who owns them, and when. Motions to dismiss now test ownership, timeliness and distribution work by work. PayAuthors supplies the work-level facts from the libraries' own published catalogs and the U.S. Copyright Office's records.
Screening a client list? Ask for a quote and we will screen the whole list.
Universities, libraries, publishers and firms: ask for a quote
What the record already says
Courts and the defendants themselves have supplied the collection-level facts. Each statement below is labeled by its source, and each source links to the document.
| Statement | Source | Kind |
|---|---|---|
| Anthropic's co-founder "used the infamous BitTorrent protocol" to copy LibGen in June 2021. | Bartz v. Anthropic, N.D. Cal., ECF No. 244 at 2 | Court order |
| Anthropic compared PiLiMi with its LibGen copies and torrented only the copies "that [we]re not in Lib[G]en." | Bartz, ECF No. 244 at 3 | Court order |
| For settlement purposes only, the parties stipulated that every Works List work was downloaded by August 10, 2022, the agreed Download Date. | Bartz, ECF No. 418 at 49 | Settlement stipulation |
| Contributory liability requires intent, shown by inducement or by a service tailored to infringement. | Cox Communications v. Sony Music, No. 24-171 (U.S. Mar. 25, 2026) | Supreme Court |
| Meta did not dispute, at the pleading stage, that uploading and downloading are infringing acts. | Strike 3 Holdings v. Meta, N.D. Cal., ECF No. 61 at 7 | Court order |
| Final approval of the $1.5 billion settlement for the works on the Works List; judgment entered (July 20, 2026). | Bartz, ECF No. 680 | Court order |
| The court let the authors add a claim over BitTorrent uploading (Mar. 25, 2026). | Kadrey v. Meta, N.D. Cal., Dkt. 700 | Court order |
| Meta admits downloading parts of datasets, including Books3, LibGen, Z-Library and Sci-Hub, by direct download and torrenting, and that some data was uploaded while it torrented. | Sullivan v. Meta, N.D. Cal., Dkt. 85 ¶¶46, 49–52 (Sept. 29, 2026) | Company filing (answer) |
| NVIDIA admits downloading data for AI research from Anna’s Archive’s datasets page and from libgen.rs, without dates. | Nazemian v. NVIDIA, N.D. Cal., ECF No. 321 | Company filing (answer) |
| OpenAI says it trained GPT-3 and GPT-3.5 on “compilations of books from Library Genesis.” | In re OpenAI Copyright Litigation, S.D.N.Y., Dkt. 1720 | Company filing (brief) |
Latest from the docket court orders and company filings, live from our news feed ·
We supply the work-level link. For each incident fixed by a court order or a company’s own filing, a report shows whether a given owner’s book was in that collection as it stood when it was taken: the catalog snapshot nearest the date, the exact file (catalog ID, and MD5 where the catalog records one) and, where it is archived, the torrent that carried it. Whether a defendant’s download included that particular file comes from the defendant’s own records.
Scale
File-level records
Every LibGen and Z-Library line in a report names the file the library offered, in the library’s own terms, so it can be compared with a defendant’s file lists, hashes and torrent logs produced in discovery.
The library’s own ID
The number LibGen or Z-Library gave that one file (LibGen numbers non-fiction and fiction separately).
Z-Library ID 5276476The MD5 fingerprint
A checksum of the file’s exact bytes, from the library’s own catalog. A produced copy can be matched to it byte for byte, without anyone opening the book.
MD5 721630f8116455be2e24f39246759772The torrent and its pieces
The BitTorrent bundle that carried the file, and the file’s size, the piece length, how many pieces it spans, and the largest share of the file inside one piece, from the torrent file itself.
pilimi-zlib-5240000-5329999 · file 320 KiB, inside one 4.0 MiB pieceExample: The Herd by Andrea Bartz, from the sample report. Method and coverage: Methodology.
Distribution claims: a whole book often fits in one torrent piece
Defendants argue that BitTorrent uploads carried only fragments. BitTorrent checks data piece by piece but sends it in small blocks, which a downloader can collect from several peers. When a file lies inside one piece, every verified copy of that piece holds the whole file; whether one uploader sent all of it has to come from that uploader’s records. Most LibGen fiction (79%) and PiLiMi release 1 (74%) files lie inside one piece; 26% of LibGen non-fiction files do. The torrents’ own file lists do not show that a defendant uploaded a particular piece; that has to come from the defendant’s own records or other evidence. We read them from public archived copies; no swarm is joined and no book file is read.
Pieces are mostly 4 MiB (LibGen 1 to 8 MiB). Figures cover located files in the current dataset. An expert can recheck every figure with a short standard-library script we supply.
Defendant lists
Ready lists of registrations by defendant and theory, each with catalog IDs, § 412 timing, and MD5s, torrents and piece data where they are recorded. The basis column says what the defendant-side link rests on. Lists identify works, not people, for analyzing claims of clients you already represent.
Works, not dollars. The lists count works that pass the registration-date screen; they are not damages estimates. If a court awards statutory damages, it sets them per work, from $750 to $30,000, or up to $150,000 for willful infringement (17 U.S.C. § 504(c)), with one award per work for all of one infringer’s infringements in the case, or one shared award where infringers are jointly and severally liable; all parts of a compilation or derivative work count as one work (§ 504(c)(1)). So a book on two of a company’s lists is still one work, a company’s lists must not be added together, and different companies’ lists are not a combined total. Whether any work qualifies is for a court. Actual damages and profits (§ 504(b)) are an alternative that must be proven.
| List | Works | Basis |
|---|---|---|
| Anthropic: Books3 works not on the Works List | 5,419 | Bartz record |
| Anthropic: LibGen works not on the Works List | 28,750 | Bartz record |
| Anthropic: PiLiMi works not on the Works List | 6,787 | Bartz record |
| Anthropic: LibGen works in torrents made June 20, 2021 to Aug. 10, 2022, not on the Works List (caution, see below) | 18,904 | Bartz record (agreed date, settlement only) |
| Anthropic: works first registered 1931–1977 (renewed, or registered 1964–1977), not on the Works List; candidates | 21,806 | Bartz record; Copyright Office registration and renewal records |
| Anthropic: works opted out of the Bartz settlement (their owners kept the right to sue) | 1,802 works | Court order (Bartz final approval) |
| Meta: LibGen | 493,479 | Court opinion (Kadrey) |
| Meta: Z-Library records added Aug. 2022 to Aug. 2023, via Anna's Archive (early 2024) | 221,052 | Court opinion (Kadrey); Meta's answer in Elsevier |
| DeepSeek: same Z-Library records, via Anna's Archive (by March 2024) | 209,902 | Company paper (DeepSeek-VL) |
| OpenAI: LibGen, 2018 (October–November as alleged) | 315,984 | Court order (the year); months and scope alleged |
| OpenAI: LibGen torrenting, 2019 to 2020 | 406,913 | Allegation |
| OpenAI: Books3 (via The Pile), February 2021 | 78,553 | Allegation |
| NVIDIA: same Z-Library records, via Anna's Archive (August 2023) | 217,965 | Allegation; NVIDIA's answer admits downloading from Anna's Archive's datasets page, undated |
Works = books that pass our § 412 date screen (a registration dated before the download, or within three months of a publication that came first; 17 U.S.C. § 412), counting copies, formats, editions and translations of one book once. Lists overlap and are not additive, and each match is a candidate to confirm. Opt-out works are counted by the court’s list, without our timing check. Anthropic’s rows count every work with a copy in its downloads that is not on the Works List as collected from the settlement’s lookup site, under any registration or edition we can link to it, and that was registered before Anthropic’s copying could have begun. Anthropic’s count is 34,330 works with a TX or TXu registration and a copy in Anthropic’s stated download windows. Another 18,093 such works have a copy only in a LibGen torrent made after Anthropic’s stated LibGen window (June 8–19, 2021) and on or before Aug. 10, 2022; they are on its list but are reported only, not counted. Caution: Aug. 10, 2022 is the date by which the Bartz parties agreed every Works List work was downloaded, for the settlement only and only for Works List works. Anthropic’s stated window is an estimate from code dates, and no record shows Anthropic downloaded those later torrents. Works first registered from 1931 to 1977 are shown in their own column for every company. They are matched by title and author to the original registrations for 1964–1977 and to the renewal records for earlier works, both as transcribed by the New York Public Library from the Catalog of Copyright Entries; each is counted once per company and left out if a copy carries a later TX or TXu registration, and for Anthropic also if the work, its renewal or another edition is on the Works List. They are candidates to confirm: in a check of 50 Anthropic matches, 44 were the same work. Registrations that cover only new matter, such as a revision, a translation or illustrations, are included and protect only that new matter. Another 136,767 (Meta) and 119,381 (NVIDIA) pre-1978 works are in Internet Archive lending books those companies are alleged to have downloaded; they are not counted. Other counts cover TX and TXu registrations from 1978 on.
One award per work: each company’s lists combined
A book on two of a company’s lists is one work and one award. Here each company’s lists are combined and each work is counted once. Statutory damages generally depend on registration before the infringement commenced, or within three months of first publication (17 U.S.C. § 412). As a screen, a work is counted only if a registration predates that company’s first download of it, across all its lists, or came within three months of a publication that came first. Passing the screen is not a finding that an award is available. A work is on a company’s list if it was in that library at the time, whatever its registration date, so a work that does not pass the screen at its first copy stays out of the count even when a later copy would pass. The first three columns cover TX and TXu registrations (1978 on); works first registered from 1931 to 1977 are counted once in the last column, as candidates. Another 136,767 (Meta) and 119,381 (NVIDIA) pre-1978 works are in Internet Archive lending books those companies are alleged to have downloaded; they are not counted.
| Company | Works on its lists | On two or more lists | Counted (registered in time) | Pre-1978 works (candidates) |
|---|---|---|---|---|
| Anthropic Books3, LibGen and PiLiMi works not on the Works List; works whose only copy is a LibGen torrent made through Aug. 10, 2022 are on its list but not counted (caution) | 62,537 | 8,176 | 34,330 | 21,792 |
| Meta LibGen, Anna’s Archive and Books3 | 625,735 | 148,757 | 617,167 | 41,130 |
| NVIDIA Books3 and Anna’s Archive (allegations) | 274,043 | 29,992 | 267,678 | 57,884 |
| OpenAI LibGen 2018 (court order; months and scope alleged), LibGen torrenting 2019–2020 (alleged) and Books3 2021 (alleged) | 432,371 | 337,675 | 422,577 | 24,321 |
About the opt-out figures
1,802 is the number of works on the court’s opt-out list, from the 350 owners who opted out on time (Bartz ECF No. 680 at 6; list at ECF No. 619-3, Ex. J). The court also allowed 2 late opt-outs; their works are not on that list (ECF No. 680 at 12–14). To split the list, we compared it with the works named in the seven suits opt-outs have filed against Anthropic as of October 1, 2026. For two of those suits we used the same plaintiffs’ exhibits from their suits against Meta. A work can have several owners, and a suit we have not found is not counted, so both figures are estimates, not court findings.
What we work from
Beyond the library catalogs, an investigation can use any lawful source:
- torrent and magnet files, with their file lists, piece hashes, sizes and MD5s;
- court exhibits, and discovery that counsel shares with us under the case’s protective order;
- company papers, model cards and dataset documentation;
- archived web pages;
- authors’ and publishers’ own records;
- tests of what AI models produce, run against the author’s own legitimate copy of the book.
Ongoing monitoring
We check for new filings in the AI copyright cases every three hours, and post items after review, so an item can appear later than its filing. New court rulings, company statements and allegations are logged with their sources and graded; documented download dates go into the report timeline and the defendant lists after review, and newly archived catalogs are added in dataset releases. The latest items are on our News page. For firm clients we can flag developments that touch their clients' works.
Ways to work with us
Work-level research: $50 per work
Any number of works, chosen from our defendant lists or by registration numbers you give us. For each work: the registration and its Copyright Office record; each catalog line with the library’s ID and MD5 for LibGen and Z-Library files, and the torrent and piece layout where the torrent is archived; the registration-date screen against each company; and a note on any edition or same-name risk. Every line is reviewed by PayAuthors against the Office record before delivery, and you get one correction pass. Delivered as a spreadsheet and PDF exhibits with a method note; we give a delivery date with the quote.
Case support
Complaint exhibits for your works: catalog entries, and MD5s, torrents and piece data where the library and its archived torrents record them, with ownership screens and a method note an expert can check.
Expert support
Your expert gets our scripts and checksums to re-run the method from the same sources.
Content checks
Some questions need a look inside a pirate copy: which edition it is, or whether the copyright page and other copyright management information were removed (17 U.S.C. § 1202). A person acting for the rightsholder does these checks, only for the rightsholder’s own works, with their permission and at counsel’s direction. Only that one file is downloaded, directly, never seeded, and every step goes in a chain-of-custody log. We supply the written protocol, the target list (each file’s library ID, MD5 and torrent, and the edition the Copyright Office record describes) and a checklist, and we match the file’s MD5 against the library’s catalog. With the rightsholder’s written permission, we can also receive the reviewer’s copy, check it ourselves and quote what the case needs, such as the copyright notice or the edition statement. Without that permission, we review the completed checklist and never receive the book.
Defendant lists
Lists of works by defendant and theory, as spreadsheets, for intake and case planning.
Who we work with. Our published reports are available to anyone, with the same standards for everyone. Custom and case work is done for one side only: we do not take work for a party adverse to an existing client in the same matter. Custom work for a firm is confidential to that firm. Firm work is done under a short written engagement.
Rate card
Law firm work is priced per work.
| Service | Price |
|---|---|
| Work-level research: registration, catalog lines, § 412 timing, reviewed, as spreadsheet and PDF exhibits | $50 per work, no minimum |
| Case support, defendant lists, content-check protocols and review | By engagement |
Firm work runs under a short written engagement covering confidentiality, the deliverable, corrections and limits of liability.
Method and limits
We never download or share a pirated book, and never join a BitTorrent swarm. We open one only in a content check: the rightsholder’s own reviewer downloads it, and we receive it only with the rightsholder’s written permission. Each file is identified by the library’s own MD5 and its torrent’s piece hashes, so a copy produced in discovery can be checked against them byte for byte (how). Every source file is recorded by SHA-256 checksum, and each registration match links to the Copyright Office record so it can be confirmed. Matching is automated and not perfect: expect some missed records and some matches that are not your clients'. A catalog listing shows that a file was in a library on a date; tying it to a particular download takes the defendant's own evidence. PayAuthors is not a law firm, does not refer clients to lawyers, and is not affiliated with or endorsed by the U.S. Copyright Office. See the Methodology for sources and checksums.
Supporting research materials are preserved for review. Where legally permissible, underlying source records and research materials can be made available for inspection by counsel, experts, or other qualified reviewers, under a short written engagement. What is preserved
This page is general information, not legal advice.