Hobbs v. Meta: Did Meta knowingly use shadow libraries, upload large volumes back into torrent networks, abandon licensing to preserve a fair-use strategy & treat books as uniquely valuable training?

Summary: Hobbs v. Meta is not just another AI training/fair-use case: it reframes Meta’s alleged conduct as piracy, BitTorrent distribution, CMI stripping, concealment, and personal accountability for executives and AI researchers.



The strongest allegations are that Meta knowingly used shadow libraries, allegedly uploaded large volumes back into torrent networks, abandoned licensing to preserve a fair-use strategy, and treated books as uniquely valuable training fuel.



If the core claims survive, AI makers may face a new compliance reality: provable data provenance, licensing discipline, metadata preservation, researcher indemnities, and real litigation exposure for “dirty data” pipelines.

Hobbs v. Meta: the case that tries to turn AI training from a fair-use debate into a piracy, concealment and personal-liability case

by ChatGPT-5.5

The Hobbs v. Meta complaint is best understood as an escalation, not merely another “authors versus AI training” lawsuit. Its central move is to shift the frame away from the familiar question — can copyrighted works be used for AI training under fair use? — toward a harsher and more operationally concrete question: what happens if an AI company allegedly built part of its data supply chain through piracy, BitTorrent distribution, metadata stripping, concealment and executive/researcher-level decision-making? The plaintiffs, authors Jeff Hobbs and A. Douglas Stone, allege that Meta, Mark Zuckerberg, Guillaume Lample, Joelle Pineau and unnamed others copied, torrented, redistributed and used millions of copyrighted works from LibGen, Anna’s Archive, Z-Library, Sci-Hub and related sources to train Llama, while removing copyright management information and avoiding licensing. These are allegations, not proven facts, but the complaint is unusually detailed and clearly designed to make “AI training” look like the last stage of a wider infringement pipeline rather than the whole legal issue.

How this case differs from other AI copyright lawsuits

The first difference is that Hobbs tries to break the legal spell around “training.” Many AI cases have turned on whether ingesting copyrighted works into a model is transformative, non-substitutive, and therefore fair use. Hobbs argues that this is the wrong starting point because Meta allegedly did not lawfully acquire the works in the first place. The complaint says the case is “fundamentally different” from disputes about lawfully acquired materials because Meta allegedly downloaded the works from pirate repositories and then reproduced, retained and redistributed them through BitTorrent. That matters because the plaintiffs are trying to create independent liability for acquisition, reproduction, retention and distribution, regardless of whether later model training is considered transformative.

The second difference is the personal-liability strategy. The attached commentary notes that Hobbs does not only sue Meta and Zuckerberg, but also names Joelle Pineau, former head of Meta FAIR, and Guillaume Lample, former Meta researcher and later Mistral co-founder. The blog describes this as more aggressive than Concord Music v. Anthropic II, which allegedly named Anthropic co-founders, because Hobbs reaches into the researcher layer rather than stopping at founders or corporate officers. Public reporting on Concord Music v. Anthropic II likewise describes that case as naming Anthropic CEO Dario Amodei and co-founder Benjamin Mann personally, which makes Hobbs part of a broader move from “company liability” toward “who inside the company made the data decision?”

The third difference is the BitTorrent distribution theory. Hobbs is not only saying Meta downloaded books. It says Meta allegedly uploaded at least 69 TB of data to other BitTorrent users because it did not change default settings that would have reduced or eliminated uploading. That is strategically important because distribution claims are often easier to moralize and operationalize than abstract “model ingestion” claims. If the plaintiffs can show Meta’s systems seeded or made copyrighted files available to others, the case becomes less about whether an LLM “learns like a human” and more about whether a trillion-dollar company functioned as a node in a pirate distribution network.

The fourth difference is the concealment and CMI theory. The complaint alleges that Meta developed scripts to remove copyright notices, author names, ISBNs and publication information, and that internal fields such as “lines_copyright_removed” allegedly confirm systematic CMI stripping. It also alleges that Meta fine-tuned Llama to stop giving incriminating answers about being trained on pirated or illegal data. If proved, that is more damaging than ordinary unlicensed copying because it supports willfulness, concealment, DMCA liability and potentially punitive narratives around corporate knowledge.

The fifth difference is that Hobbs is arriving after important fair-use rulings and settlements, so it is drafted against the battlefield rather than before it. In Kadrey v. Meta, Judge Chhabria rejected the idea that downloading from shadow libraries automatically defeats fair use, while also saying shadow-library sourcing may matter to bad faith, distribution and contributory infringement; importantly, he noted the plaintiffs had not brought or developed a contributory infringement claim there. Hobbs appears to be a direct answer to that gap: it pleads distribution, contributory infringement, CMI removal, state-law claims, and individual participation.

The most surprising, controversial and valuable statements

The most surprising allegation is that Meta allegedly stopped licensing because licensing even one book might undermine its fair-use strategy. The complaint says Meta explored licensing with major publishers, then allegedly abandoned that path after an escalation to Zuckerberg, with one engineer allegedly summarizing the logic as: if Meta licensed a single book, it would be harder to “lean into” fair use. If that holds up, it is devastating because it reframes fair use not as a good-faith legal interpretation but as a business tactic selected to avoid market formation.

The most controversial allegation is that Zuckerberg personally authorized or encouraged the use of pirated data. The complaint says a Meta employee confirmed that the decision to torrent copyrighted works occurred after escalation to “MZ,” and that Zuckerberg was not merely a passive corporate officer. This is the kind of allegation that changes settlement pressure: the risk is no longer only statutory damages, but personal deposition pressure, governance embarrassment and board-level questions about whether copyright risk was knowingly converted into AI speed.

The most valuable technical allegation is that books were not incidental to model training. Meta allegedly viewed books as especially valuable for long-context performance, creativity and agent capability; the complaint quotes internal statements that books were “more important than web data,” that LibGen could help creativity, and that the “Books strategy” was “libgen — FREE.” That is valuable for rights owners because it rebuts the industry line that any single work or category of works is negligible in a sea of tokens. The complaint argues that high-quality, edited, expressive books were targeted precisely because they made the model better.

The most explosive operational allegation is that Lample allegedly torrented a major LibGen copy, used infrastructure including Meta clusters and cloud services, referred to LibGen as “BooksZero,” kept code off Meta’s repository, and that one copy later “vanished” from Meta’s possession. Even if some of this weakens in discovery, it gives plaintiffs a concrete data-chain narrative: who downloaded what, where it sat, how it was named, how it moved, and what disappeared.

The most damaging cultural evidence is the internal discomfort. Meta employees allegedly described LibGen as pirated, worried about legal and ethical consequences, questioned torrenting from corporate infrastructure, and warned that pirated data should not go into a published model. Such evidence is powerful because it collapses the usual “everyone thought this was normal” defence. The complaint’s narrative is that people inside Meta knew the legal and ethical problem but speed, competition and cost avoidance prevailed.

Are the evidence and arguments robust?

The evidence appears strongest on knowledge, sourcing and willfulness, assuming the quoted internal documents, logs and testimony are accurately represented. The complaint does not merely infer that Meta used suspicious datasets; it alleges internal discussions, data volumes, code references, repository behaviour, cloud locations, testimony, download/upload logs, CMI-removal fields and specific comments from named employees. That is much stronger than a complaint based only on output similarity or circumstantial model behaviour.

The distribution theory is also potentially robust because BitTorrent has a built-in upload/download architecture. If Meta’s logs really show 69 TB of uploads, plaintiffs have a concrete act of distribution to litigate. This matters because even courts sympathetic to AI training fair use may treat seeding pirated books differently from copying books into a training pipeline. Kadrey itself recognized that peer-to-peer file sharing will often constitute infringement and that materially contributing to shadow libraries could support contributory liability if properly pled and evidenced.

The DMCA/CMI claim is serious but will depend on proof of intent and statutory fit. It is not enough that metadata disappeared during preprocessing; plaintiffs will need to show removal of copyright management information with knowledge, or reasonable grounds to know, that it would induce, enable, facilitate or conceal infringement. The complaint tries to meet that standard by alleging scripts designed to remove copyright notices, author names and identifying information, and by linking that removal to concealment of training sources. That is a plausible and dangerous theory for Meta, but the details of the scripts, their purpose, and whether the removed information qualifies as CMI for each work will matter.

The weaker part is market substitution. The complaint alleges that Llama can generate knockoff books, replacement chapters, summaries, derivative works and imitations of authorial voice. That is intuitively compelling, especially for authors, but courts have so far demanded concrete market evidence rather than general claims about dilution. In Kadrey, the court found Meta’s training use fair on the record before it, partly because the plaintiffs had not developed sufficient market-harm evidence; however, the court also suggested different evidence could matter. Hobbs tries to add that evidence through allegations that Llama was “annoyingly good at quoting books” and could assist with fiction and poetry, but the plaintiffs will still need to connect Meta’s conduct to cognizable market harm, not just anxiety about AI competition.

The personal claims against Pineau and Lample are strategically bold but procedurally vulnerable. The complaint’s allegations against Lample are more concrete because they focus on torrenting, code, infrastructure and data locations. The allegations against Pineau appear more supervisory: she led FAIR and oversaw Llama. That may not be enough unless discovery shows direct authorization, material contribution or control over the challenged acts. Personal jurisdiction in New York may also become a serious threshold fight for non-New York individuals.

The state-law claims — unfair competition, conversion and unjust enrichment — may be vulnerable to Copyright Act preemption. They are useful rhetorically because they make the case sound like theft, misappropriation and unjust benefit rather than mere technical copying. But courts often narrow or dismiss state-law claims when they do not add an extra element qualitatively different from copyright infringement. The federal copyright, distribution, contributory and DMCA claims are likely the real centre of gravity.

How the case could conclude

The most likely near-term outcome is partial survival, not total dismissal or immediate victory. Meta will almost certainly move to dismiss or narrow the complaint, attacking personal jurisdiction, individual liability, state-law preemption, class allegations, causation, CMI sufficiency and market-harm theories. Some state claims may be trimmed. Some individual claims may be narrowed. But the complaint is specific enough that the core claims around reproduction, BitTorrent distribution, contributory infringement and CMI removal may well survive into discovery.

At summary judgment, Meta will lean hard on Kadrey: training is transformative; shadow-library sourcing is not an automatic fair-use defeat; and market harm must be proved rather than presumed. Plaintiffs will counter with Bartz v. Anthropic logic: training may receive fair-use protection in some circumstances, but building or retaining a pirated central library is a separate act, and pirated acquisition cannot simply be laundered through a later transformative purpose. The Anthropic decision is especially relevant because commentators note that the court separated lawful scanning/training from storing pirated books, treating the latter far less favourably.

A settlement is plausible, especially if class certification becomes realistic or discovery confirms the strongest internal-document allegations. The Anthropic author settlement reportedly involved at least $1.5 billion, approximately $3,000 per work, destruction of downloaded files and no future licence for training; that creates a market reference point for AI companies facing pirated-book claims. Hobbs may not automatically reach that scale, but if the alleged universe includes millions of works and the conduct includes distribution and CMI removal, Meta faces a risk profile that is not purely about fair use.

A full injunction shutting down Llama is less likely than monetary relief, data-destruction obligations, audit obligations, preservation orders, restrictions on further use of identified works, and licensing-based remediation. Courts are cautious about disruptive injunctions affecting widely deployed technologies. But targeted injunctions around retained datasets, future copying, use of plaintiffs’ works without licences, and preservation of files and records are much more plausible. The complaint’s prayer for relief is broad, seeking class certification, damages, permanent injunctive relief, limits on further exploitation, DMCA damages, costs, fees and preservation-related restrictions.

Potential consequences for AI makers

For AI makers, the case is another signal that “data provenance” is becoming a litigation-grade control environment, not a trust-and-safety slogan. The future compliance baseline will likely include documented acquisition pathways, dataset chain-of-custody, permission status, CMI preservation, opt-out handling, retention limits, deletion proof, audit trails and board-level escalation rules for high-risk datasets. The companies most exposed are those that treated training data as an engineering bottleneck rather than a rights-managed supply chain.

Second, individual researchers and executives may start demanding indemnities, written approvals and clearer data-use policies. The Hobbs theory makes data sourcing a personal-risk issue. AI labs will not be able to tell researchers, “just get the data,” while leaving them exposed if the source later turns toxic. That could slow frontier-model development, but it may also professionalize the field by forcing the same kind of documented risk gates that exist in pharma, finance, cybersecurity and regulated publishing.

Third, CMI and metadata removal may become a red-line issue. AI companies have often treated preprocessing as technical hygiene: normalize text, remove noise, deduplicate, strip headers and footers. Hobbs turns that into a potential concealment narrative. From now on, legal and engineering teams will need to distinguish between legitimate cleaning and rights-information destruction. Metadata preservation may become as important as model performance.

Fourth, open-source or widely distributed models become harder to defend if trained on dirty data. If a closed model has problematic training data, the company can at least contain, retrain, filter or remediate. If an open model is released into the wild, downstream control becomes far weaker. That makes pre-release data governance more important, especially for companies positioning themselves as open AI infrastructure providers.

Finally, the case strengthens the strategic hand of publishers and other rights owners. The litigation is no longer only about whether AI training is transformative. It is about whether AI companies knowingly avoided licensing markets, used pirate infrastructure, removed identifying information and externalized the cost of content creation onto authors, publishers and scientific communities. Even if Meta ultimately wins parts of the fair-use argument, the broader consequence is that the AI industry’s “move fast, scrape everything, sort it out later” phase is becoming legally, reputationally and operationally dangerous.