OpenAI’s models ‘memorized’ copyrighted content, new study suggests

Ay To study new Openly seems to have trained at least some AI models in its copyright content, it seems to be credible to the allegations.

Opena is involved in the suits brought by the author, programmer and other rights who complain to the company to use their works, codebase and more to use its models without permission. OpenAI has long claimed Fair use Defense, however, argued that the plaintiff in the case argued that there was no engraving in US copyright law for training.

The survey was co-authored by researchers at Washington University, Copenhagen University and Stanford to offer a new approach to detecting “memorized” training data by an API like Openai.

Models are predicted engines. Trained in lots of data, they learn the patterns – this is how they are able to create essays, photos and more. Most outputs are not training data verbatim copies, but the models are somewhat inevitably because of “learn”. The models of the image have been found Rebuild the screenshots from their trained moviesLanguage models have been observed Effectively steal news articlesThe

The method of study depends on the words that co-authors say “high-protected”-these are the words that stand as abnormal in the context of the larger work body. For example, the word “radar” in the phrase “Jack and I am sitting perfectly with radar humming” will be considered high-end because it is statistically less likely than “Humming” than “engine” or “radio”.

Searched several OpenAI models, including co-authors GPT -4 And GPT -1.5, fiction book and the New York Times pieces removed from the snipts of the pieces to remove high -protected words and “guess” which words were masked. If the models are able to guess properly, they probably memorized the snippet during training, reached the conclusion of co-authors.

According to the test results, GPT -4 showed signs of keeping the memorable parts of the books of popular fiction, which contains books in a datasate of ebooks called bookmia. The results further suggested that the model memorized the parts of the New York Times articles at relatively low rates.

Doctoral student of the University of Washington and co-authors of this study, Abhihilasha, told Rabicandar TechCrunch that these searches focused on “controversial information” models could be trained.

“To keep the credible big language models, we need to have a model that we can scientifically investigate and monitor and examine,” Ravicandar said. “The goal of our job is to provide a large language model to the investigation, but the entire ecosystem needs the true need for transparency.”

Opena has long been supported Loose restrictions In the development of models using copyrighted data. Although the organization has the specified content licensing deals and provides the opt-out process that lets copyright owners flag the content that they prefer to use for the purpose of training, it contains Several governments have planned To coding the “fair use” rule around the AI ​​training method.