Researchers suggest OpenAI trained AI models on paywalled O’Reilly books

Openai has been defendant by too AI provides permission of AI training parties on copyrighted content. Now new paper The AI ​​Watchdog provides a serious accusation for the company’s increasingly reliability because the company has not licensed to develop more complex AI models.

AI models are essentially complex prediction engines. Over many information – books, movies, television shows, etc. They learn patterns and novels to extrapolate for a simple desire. A model “writes” a Greek tragedy or “draws”, “writes ghibling style pictures,” it simply draws an approximate knowledge. It doesn’t come to something new.

AI laboratories, including Openai, a number of AI’s laboratories reflect a small number of real world data from the EU’s real world sources (mainly public website). Most likely, training on purely synthetic information comes with risks such as worrying a model.

The new paper, the AI ​​statements project, Media Mogul Tim O’reilly and Non-Commercial, Openai, founded in 294 by Economist Snake Strauss, draws a conclusion GPT-4O O’Reilly Models in Paywalled Books from Media. (O’Reilly is the CEO of O’Reilly Media.)

In ChatgptGPT-4O is a standard model. There is no licensed agreement with O’Reilly Openai, says paper.

“GPT-4O, Openai’s more recent and skillful model, payawalled o’reilly demonstrates strong recognition of book content […] “The previous model of Openai compared to the GPT-3.5 Turbo,” wrote the co-authors of the paper. ” Instead, GPT-3.5 Turbo shows more relative recognition of O’Reilly book samples, which are open to the public. “

Used a method called paper To the childFirst of all, in an academic document in 2024, it was designed to detect copyrighted content in language models training information. In addition, it is also recognized as a “membership result”, a model tests that the author’s texts are paraphrasized and the EU can reliably distinguish from the generated version of the EU created AI. If he succeeds, the model offers not to know the text in advance about the text of the text.

The co-authors of the paper – O’Reilly, Strauss and AI researcher Sruly Rosenblat – say they tested GPT-4O, GPT-3.5 TURBOand the knowledge of O’Reilly media books published before and after the training of other Openai models. They used 13,962 paragraph from 34 o’reilly books to assess the possibility of a certain part to be entered into the training database.

According to the results of the paper, GPT-4O “recognized”, GPT-3.5, including GPT-3.5 Turbo, more PayWalled O’Reilly Book content from Openai’s old models. It also said that the authors of the authors and the authors of the authors of the new models were made as progress in the ability to understand the text of the text for potential mixed factors.

“GPT-4O [likely] It is recognized and therefore there is in advance about those who are not very openly open to those published before the history of pre-warning, “said co-authors.

Not a smoking weapon, and the co-authors are careful not to draw attention. They are acknowledged that experienced methods are not foolish and the PayWalled book of Openai can copy the PayWalled Book to Chatgep and paste them.

The waters did not evaluate the collection of Openai, which includes the GPT-4.5 and “providers” models such as co-authors O3-mini and O1. It is possible that these models are not taught to PayWalled O’Reilly Book data or are taught less than GPT-4O.

It is said that Openai, Advocated Openai Looser Restrictions Using copyrighted data, looking for better quality training information around developing models. The company went so much Hire journalists to delicate the performances of their models. It has a trend from the extensive industry: AI companies hire experts in domains such as science and physics Effectively these experts have their knowledge of AI systems.

It should be noted that Openai pays at least some training information. The company has licensing deals with publishers, social networks, stock media libraries and others. Openai also offers opt-out mechanisms – Although the flawless ones – This allows copyright owners to prefer the content they prefer not to use for the company’s training purposes.

Still, like Openai battles, teaching information in training information and the treatment and treatment of copyright laws in US courts, O’Reilly paper is not the most flattering view.

Openai did not respond to a request for comment.