Two authors file class action against OpenAI for ‘copyright violation’ by ‘training’ ChatGPT

A pair of best-selling novelists have filed a lawsuit against ChatGPT’s parent company, claiming it had broken copyright law by training its chatbot to “take” their books without permission.

The lawsuit, filed in federal court on Wednesday, comes from authors Mona Awad and Paul Tremblay, who both live in Massachusetts.

They claim that ChatGPT has been trained in part by “taking” several of their novels – all without their consent.

A class action lawsuit, the first charge against the San Francisco company responsible for the polarizing copyright chatbot. The company, OpenAI, released it in November – and has since seen a surge in profits.

That said, ChatGPT’s underlying model has been trained with data publicly available on the internet – and attorneys for the authors say this includes copies of several of their clients’ books, which are copyrighted.

Referring to three titles in particular, the pack reads: ‘Instead of being programmed in the traditional way, [ChatGPT] is ‘trained’ by copying huge amounts of text and extracting expressive information from it.

“This text is called the training dataset,” it continues. “Once a large language model copies and incorporates the text into its training dataset, it is able to convincingly broadcast natural text output.

“Each time it compiles a text output, the model relies on the information it has extracted from its training dataset.”

It then goes on to mention how Tremblay, author of the award-winning horror novel The Cabin at the End of the World, and Awad, a horror writer whose 2019 book Bunny was selected by Time as Best Novel of the Year, said that ChatGPT provided “highly accurate” summaries of their books – indicating that they have appeared in the database.

In particular, lawyers said ChatGPT, when requested, sent extremely detailed summaries of The Cabin at the End of the World — which won the Horror Writers Association’s Bram Stoker Award for Best Novel in 2019 — and Awad’s Bunny and 13 Ways of Looking at a Fat Girl.

Both authors argue that this is proof enough that their novels have been mined to “train” the chatbot. It contains in-depth responses to clues regarding those novels as evidence.

Now seeking compensation, the authors claim that OpenAI has “unfairly” benefited from what they say is “stolen writing.” They added that OpenAI is primarily “increasingly secretive” about how it collected the data during the bot’s “training phase.”

OpenAI, meanwhile, has not commented on those claims, while maintaining the data used to power its new technology comes only from sources publicly available on the internet.

That said, in papers released alongside early iterations of ChatGPT – which has the uncanny ability to mimic human writing – OpenAI gave some hints about the size of the “Internet-based book corpora” it used as training materials.

In June 2018, OpenAI revealed that it had trained its prototype GPT-1 with BookCorpus, a large collection of free new books written by unpublished authors that contains 11,038 books.

However, the suit described this dataset as “controversial,” claiming that artificial intelligence researchers who collected it in 2015 “copied the books from a website called Smashwords.com that hosts unpublished novels that are available to readers for free.”

“Those novels,” wrote the writers’ lawyers, “are largely copyrighted.”

The alleged copyright infringements worsened with subsequent iterations, the lawsuit states.

It cites how the company revealed in a July 2020 paper introducing GPT-3 – the bot’s third prototype – that 15 percent of its training dataset came from “two internet-based book corpora” billed only as “Books1” and ‘Books2’.

The lawyers then deduced that, based on numbers revealed in OpenAI’s paper on the prototype, Books1 would contain approximately 63,000 titles and Books2 would contain approximately 294,000 titles.

Due to their size, the lawyers said the books could not come from unpublished authors and had to come from shadow libraries such as Library Genesis (LibGen) and Z-Library – other sites where books can be massively secured via torrents.

The suit goes on to state that while a variety of materials may have been used to train the large language models, books “must have been a key ingredient in training datasets,” given the chatbot’s detailed responses regarding their ” long writing style’.

Andres Guadamuz, a reader in intellectual property law at the University of Sussex. told The protector that the new lawsuit will certainly explore the uncertain “boundaries of legality” of actions within the generative AI space as OpenAI comes under scrutiny.

Lilian Edwards, professor of law, innovation and society at Newcastle University, told the paper the case “is likely to rest on whether courts consider using copyrighted material in this way as ‘fair use’, or as simply unauthorized to copy.’

Tremblay’s book The Cabin at the End of the World was recently adapted into the M. Night Shyamalan movie “Knock at the Cabin,” which came out in February.

Awad, who also works as a professor of creative writing at NYU’s College of Arts and Sciences, was recently touted as her “heir apparent to the literary throne” by the legendary Margaret Atwood.

The dystopian writer of Oryx and Crake and the Handmaiden’s Tale said in May: “I’ve been an admirer of Mona’s novel ‘Bunny’ for some time now.

“It’s a form of gothic satire, and she puts it in writing school. It’s very funny, a little gruesome and pretty much out of the ordinary. You think, “She’s not going there… yes, she will.”‘

The suit is currently making its way through the right channels.