ChatGPT and its successor GPT-4 appear to have memorised details from vast numbers of copyrighted books, posing questions about the legality of how these large language models (LLMs) are created.
Both artificial intelligences were developed by private firm OpenAI and trained on huge amounts of data, but exactly which texts make up this training data is unknown. To find out more, at the University of California, Berkeley, and his colleagues looked at whether the AIs were able to fill in missing details from a selection of almost 600 fiction books, drawn from sources such as nominees for the Pulitzer prize between 1924 and 2020, and The New York Times’s bestsellers lists over the same time period.
The team picked out 100 passages from each book that contained a single, named character. The researchers then blanked out the name and asked the AI to fill it in.
Advertisement
In general, LLMs like ChatGPT and GPT-4 work by predicting the most likely next word in a sentence, based on statistical data learned during training, but this task was designed to expose whether the AIs could return the exact right answer. “It really requires knowledge of the underlying material in order to be able to get the name right,” says Bamman.
Both AIs completed the task with high accuracy – as much as 98 per cent for passages from Lewis Carroll’s 1865 book Alice’s Adventures in Wonderland – which is out of copyright – and 76 per cent for J. K. Rowling’s Harry Potter and the Philosopher’s Stone, which is not.
The researchers say this suggests the AIs were trained on significant proportions of both books. They also scored highly when tested on other copyrighted works such as Ray Bradbury’s Fahrenheit 451, and George R. R. Martin’s novel, A Game of Thrones.
OpenAI didn’t respond to a request for comment on this story or say whether it had used copyrighted material when training its AIs. Whether the company would be subject to legal action if it had fed books into its LLMs is uncertain.
“The legal issues are a bit complicated,” says at the University of Sussex, UK. “OpenAI is training GPT with online works that can include large numbers of legitimate quotes from all over the internet, as well as possible pirated copies.”
But these AIs don’t produce an exact duplicate of a text in the same way as a photocopier, which is a clearer example of copyright infringement. “ChatGPT can recite parts of a book because it has seen it thousands of times,” says Guadamuz. “The model consists of statistical frequency of words. It’s not reproduction in the copyright sense.”
“The use of copyright works without permission, in training datasets for large language or image models, has already emerged as one of the most pressing legal challenges to this novel industry,” says at Newcastle University, UK.
Bamman says that, ultimately, the legal system in each country will have to determine whether LLMs are infringing copyrights. “I think that’s an open question that a lot of court cases are going to decide for us in the coming months,” he says.
Regulation is also likely to play a key role: the European Union’s Artificial Intelligence Act, which has been two years in the making, will include a requirement that companies making generative AI tools will have to disclose any copyrighted material used to train their models. That requirement was a late change, added to the draft law in April, .
But there are more questions than simply the legal ones to be answered. Bamman says that one goal of the experiment was to find out what kinds of books each model knows well. “What we found was that it does really well on texts that show up a lot on the internet,” he says. “There was a lot of science fiction.”
The AIs performed less well when asked to reproduce books by Black authors, such as Toni Morrison. That suggests – as with many algorithms – there is a lack of representation across the training data that is then reflected in a model’s output, says Bamman.
“The results of this paper confirm that ChatGPT and GPT-4 are more familiar with books that more frequently appear online and that this may have implications in terms of how literary taste reflects and reinforces social inequalities,” says at Queen Mary University of London. “I would love to see a deeper reflection on the political and ethical implications of this.”
arXiv
