The Hacker News headline an OpenAI researcher worried about in 2019 arrived on September 27 with 543 points and 389 comments. Seven years earlier, an internal Slack discussion considered removing LibGen references from a paper because a headline saying OpenAI used copyrighted data from a questionable Russian site would be unwelcome on Hacker News, according to a newly unsealed statement filed by authors suing OpenAI and Microsoft. The feared post has become the post.
That symmetry explains the attention, though the Hacker News score is a measure of community interest rather than evidence for the lawsuit. The factual dispute sits in court filings, and the widely shared account came from the Authors Guild, a plaintiff and advocacy group, in its summary of the unsealed material. Read past the incendiary quotations and a narrower issue comes into view: whether acquiring books from a pirate library can be judged separately from the later work of training a model. For developers, that makes dataset provenance part of the product record, not administrative debris to clean up later.
A Slack message becomes an exhibit
The plaintiffs filed a motion for partial summary judgment on September 17 in the Southern District of New York. Their 49-page memorandum says an OpenAI employee directly downloaded about 117,500 books from LibGen in late 2018. It says two employees later torrented roughly 35 terabytes between September 2019 and January 2020, a volume the plaintiffs' experts matched to LibGen's fiction and nonfiction collections of about 4.6 million books. Those are the plaintiffs' factual assertions at the summary-judgment stage, not findings by the judge.
The corrected statement of facts attaches the awkward internal trail to that acquisition. In July 2019, researcher Sam McCandlish discussed removing LibGen from a prospective paper and wrote that a headline about "copyrighted data from sketchy russian website" appearing on Hacker News would be unfortunate. Dario Amodei, then OpenAI's research director, had called LibGen a sketchier training set. Another internal note in August said, "We trained GPT-3 on pirated stuff!" The public scaling-laws paper later referred to the source as "Internet Books," the plaintiffs' statement says.
The same filing describes a June 2022 cleanup called Project Clear. An internal channel originally named "excise-libgen" was renamed, and staff discussed deleting direct copies, derived datasets and material in GitHub storage. One message warned that fully removing LibGen would prevent OpenAI from reproducing GPT-3 or GPT-3.5. A deletion list ran for about 18 pages and contained more than 498 entries, according to the plaintiffs. That detail joins two concerns that engineering teams often handle separately: reproducibility and legal exposure can point at the same files.
OpenAI's own account confirms a smaller but material part of the history. Its September 4 summary-judgment brief says compilations of books from LibGen were used to train GPT-3 and GPT-3.5. It says no other language models at issue in this case used material downloaded from LibGen. For purposes of its motion, OpenAI also does not dispute that one or more of the authors' asserted works were used in pretraining some models. That limited concession does not accept infringement or the plaintiffs' account of intent.
OpenAI says training transformed the books
OpenAI asks the court to treat collecting and processing the data as part of a single transformative use. Its brief says pretraining extracts statistical patterns rather than delivering the books to users, and that the company applies deduplication and output controls to reduce memorization. In OpenAI's framing, the models learn relationships among tokens and produce new text through next-token prediction. The books and the resulting service serve different purposes.
The company also offers a striking counter-number. An expert searched 95 million outputs from 20 million ChatGPT conversations and found 14 instances of verbatim or near-verbatim text from the works at issue, according to OpenAI's filing. OpenAI calculates a 0.00007% rate. Twelve instances contained 32 words or fewer; the other two contained 67 and 122 words. The plaintiffs' expert used a different, partly overlapping sample and found an even lower rate, the company says. OpenAI argues that such limited output cannot substitute for the books or establish market harm.
That evidence matters to claims about model outputs and substitution. It does not automatically answer how the source copies were obtained. A system can reveal almost none of a book to users while still having made a complete copy upstream. The parties disagree over whether copyright law should examine that acquisition on its own or fold it into the model-training purpose.
The acquisition and the training may get different answers
A 2025 ruling in a separate case against Anthropic shows why the distinction has teeth. In Bartz v. Anthropic, Judge William Alsup found that using books to train specific language models was fair use. He reached a different result for pirated books kept in a central library, denying Anthropic summary judgment on those copies and sending damages questions toward trial. The ruling required a separate justification for each use. OpenAI cites Bartz for the transformative nature of training, while the authors cite its treatment of pirate-site acquisition. Both readings describe different parts of the same decision.
The US Copyright Office drew a similar analytical boundary in its May 2025 report on generative AI training. It said different uses during model development and deployment require separate consideration. Knowing use of a pirated or illegally accessed dataset should weigh against fair use, in the office's view, although it should not decide the outcome by itself. The report also said some training uses may qualify as fair while commercial training on unlawfully accessed works that compete in existing markets can fall outside established boundaries.
That is why the newly public Slack history may matter beyond embarrassing language. The messages appear to identify the source, the perceived rights problem, changes in public naming and a later deletion effort. Plaintiffs offer that sequence as evidence of knowledge and concealment. OpenAI can contest the characterization, the relevance of individual remarks and the legal effect of the cleanup. A judge still has to decide which facts are genuinely undisputed and how fair use applies to each alleged act of copying.
Provenance becomes part of the model record
The practical lesson for a model team is visible in Project Clear's collision with reproducibility. Removing a raw corpus and its derivatives may reduce continued possession, but it can also break the path back to a trained result. Keeping everything preserves that path while preserving every unresolved licensing problem with it. The plaintiffs' filing says OpenAI staff recognized both sides when they discussed whether deleting LibGen-derived data would make older runs impossible to reproduce.
A usable provenance record therefore has to connect acquisition to downstream use. A source name alone would not answer the questions raised here. Teams need to know when material was obtained, under what permission, which snapshots and derivatives contain it, which model runs consumed those derivatives, and what a deletion did to later reproducibility. Those fields are an engineering inference from the evidence, not a rule announced by the court. The OpenAI litigation shows what happens when years of chat messages, storage paths and paper terminology have to reconstruct the chain after the fact.
The headline that reached Hacker News is likely to remain the memorable artifact. The next development worth watching is less quotable: whether Judge Sidney Stein treats OpenAI's source acquisition and its training copies as one use or follows the split analysis seen in Bartz. The Authors Guild says more briefing is expected over the next several months, with a hearing in early 2027. Until the court rules, the 543-point post records attention. The provenance trail is what may carry legal weight.