art New study It appears to lend the confidence of allegations that opuke has trained some of the AI ​​models of copyright content.

The Openi was injured in suits brought by the authors, programmers, and other owners of the employees using their works – books, the condes, and the development of its models without permission. Openi has long been admitted to a equal use Defense, but the accusers in these cases argue without a carve-out in the US Copyright Law for training data.

The study, written by the researchers of the University of Washington, the University of Copenhagen, and Stanford, suggested a new method for identifying training data “like the backs of the backs of the API, like OpenII’s.

Models are insect machines. Trained in large data, they learn patterns – how they can make essays, pictures, and more. Most outputs are not copies of training data, but because of the “learn,” models are unavoidable. Image models are found Regurite screenshots from movies they trainedWhile language models are observed Effective wearing news articles.

The study method depends on the words called by co-authors “tall-surprise” – that is, words that are not usually in the context of a larger body at work. For example, the word “radar” in the passage “Jack and I sit fully sitting in the same radar form” which is considered a more “machine” or “radio”. ”

Co-authors examined many models in Openia, including GPT-4 and GPT-3.5, for signs of memorizing words satisfaction from snippets of fiction books and models tried to guess correctly, they most likely to have the snippet during training, concludes with co-authors.

According to the results of the tests, GPT-4 shows signs of memorizing the great fiction books, including books in a datas with samples called booksmia. The results also suggested that the model has memorized parts of the New York Times articles, however at a relatively low rate.

Abhilasha Ravicherner, a student at the University of Washington and a study co-author, told the TechCrunch that the found models of “steady data” can be trained in.

“To have more language models, there should be models that we can check and examine and examine scientists,” Robichander said. “Our job is intending to give a tool to check for many language models, but there is a real need for more than the mostly transparency of the entire ecosystem.”

Openi has long been suggested Looser restrictions In progress models using copyright data. While the company has some licensing deals in the area and offer opt-out mechanisms that allow the copyright owners to be unavailable in training purposes, this logged in many governments To codify the “fair use” rules around AI training methods.