Anthropic trained Claude on illegally obtained Stanford professors' books
Anthropic’s use of pirated copies of Stanford professors’ books to train its Claude AI models has reignited debates over copyright, fairness, and double standards in tech. Commenters note that a recent ruling treated the act of training on lawfully acquired books as fair use, while penalizing Anthropic $1.5 billion for downloading works from shadow libraries like LibGen. Many argue this exposes a system where large AI firms can break rules and pay settlements while individual actors face harsh prosecution, and broader questions emerge over whether copyright and “information wants to be free” ideals can coexist in the age of large-scale AI training.
Title, legality, and court findings
- Many note the original headline was misleading: the court found that training on legally acquired books is fair use; the illegality was acquiring pirated ebooks (e.g., from LibGen).
- The large settlement was for copyright infringement in acquisition/storage, not for the act of model training itself.
- Some see the title change (“illegally trained” → “illegally obtained”) as implying training on illegally obtained data might still be lawful; others find this unclear.
Piracy vs fair use in AI training
- One camp says the distinction is moot: pirating books “to study them” is still piracy and should taint the resulting models.
- Another camp stresses the legal line: fair use applies only to legally obtained copies; downloading from pirate libraries is separate wrongdoing.
- Discussion of current practice: buying large lots of used books and scanning them “appears to be legal.”
- Several note LLMs can sometimes reproduce training data verbatim, undermining claims that nothing is “copied.”
Comparisons to Aaron Swartz and CFAA
- Long subthread contrasts harsh criminal prosecution of an individual for bulk downloading academic PDFs with a civil settlement for massive corporate book piracy.
- Some emphasize legal differences: CFAA covers unauthorized system access, not simple downloading from pirate sites.
- Others focus on moral and structural inequality: individuals get decades of threatened prison, corporations write big checks.
- Debate over whether using Swartz-like theories to criminally prosecute AI execs would fix injustice or entrench bad precedent.
Copyright, morality, and “information wants to be free”
- Strong anti-copyright voices argue knowledge should be freely shared, professors are gatekeepers, and “intellectual property” is illegitimate.
- Others defend copyright as constitutionally grounded and necessary to compensate authors; they worry that AI training without payment erodes incentives to create.
- Claim that using authors’ works without pay is “slave labor” is widely criticized as hyperbolic or offensive, though some empathize with feeling loss of control over one’s work.
Economic and competitive impacts
- Concern that strict rules on acquisition but broad fair use for training create a moat for well-funded labs that can buy/scan vast corpora.
- Some fear LLMs will substitute for books, undermining authors’ livelihoods; others note people still buy books despite decades of piracy and free online information.
- A few see LLMs as “IP laundering machines”; others welcome them as democratizing access to knowledge.
Justice, enforcement, and desired remedies
- Recurrent theme: double standards in enforcement—“rules for the little guy, settlements for the rich.”
- Some argue this should trigger criminal/RICO cases, model takedowns, retraining under strict audits, or even nationalization.
- Others caution against overcriminalization, arguing we should fix bad laws rather than extend punitive precedents.