OpenAI’s 2019 note: “We trained GPT-3 on pirated stuff”

Newly unsealed sections of the class plaintiffs’ summary judgment papers in the Authors Guild case against OpenAI and Microsoft quote company staff, by name and with dates, on pirated training data and on replacing human authors. Publishers Weekly reported the material on September 21. It surfaced less than two weeks after both sides asked the court to decide the case without a trial.

The quotes come from two filings. Docket 1982 is the plaintiffs’ memorandum supporting their motion for partial summary judgment. Docket 1987 is their corrected Rule 56.1 statement, the list of facts a party says the other side cannot dispute. The plaintiffs argue that executives knew the training was illegal and expected AI-generated books to push human-authored ones out of the market.

The earliest item is an OpenAI note from August 2019: “We trained GPT-3 on pirated stuff! No sharing that!” The brief says Microsoft knew about OpenAI’s use of LibGen, a pirate e-book library, as early as April 2019.

Jack Clark, then OpenAI’s policy director, testified in May 2020 that “the better we do on GPT-X, the more worried genre fiction authors will become about us substituting for them on Amazon.” He also said that “there will be a point where a bunch of artists express worry about what we’re doing here and we’ll likely ignore their concerns and release anyway.”

Most of the material is about Tarun Gogineni, whom OpenAI hired in 2022 to lead work on the writing quality of its models. The documents say he described his “research mission” as building a machine that would supplant human authors. The brief says he knew authors complained that the “datasets are stolen” and that they were losing work to AI-generated competition, but did not find those complaints “all that sympathetic” and saw them as “acceptable economic disruption.” He wrote that the world would soon see “the death of the reader” as “machines creat[ed] slop for more machines.” A tweet entered into testimony says that if George R.R. Martin dies early, “GPT can write the next [S]ong of [I]ce and [F]ire book.”

The Authors Guild’s own account describes the mission more narrowly: having GPT models write the last two books of Martin’s series. It also quotes Gogineni saying he had “an undue amount of power over the english language” and that the models depended on “the vast lineage of human mastery recorded in bytes, without which models could not do anything great.”

In 2022 OpenAI deleted its LibGen files under an effort called Project Clear. “Given how much OpenAI is in the news, now is the right time to excise Libgen from our systems and storage. What would be involved in that?” Bob McGraw, VP of Research, wrote in a memo to colleagues.

The quotes go to the factor the plaintiffs lean on most. Their brief calls harm to the market the “single most important element” in the fair use analysis and argues that “OpenAI’s models are capable of displacing human authors precisely because it exploited their works.” OpenAI and Microsoft filed summary judgment motions of their own. On October 27, 2025, Judge Sidney Stein refused to dismiss the claims over ChatGPT’s outputs, finding that a jury could consider summaries of the plaintiffs’ novels substantially similar to the books. He did not rule on fair use.

The filings came days after unsealed documents in the New York Times case described the use of millions of news articles to build LLMs as an “existential threat” to newspaper publishers.

Oppositions to the summary judgment motions were due October 9 and replies are due November 6. No hearing date has been set.