OpenAI Trained AI Models by Memorizing Copyrighted Content, New Research Proves
Apr 05, 2025 - digitalinformationworld.com
A recent study highlights concerns about OpenAI's training of AI models using copyrighted material without consent, leading to lawsuits from authors and programmers. The study, co-authored by experts from the University of Washington, Stanford, and the University of Copenhagen, reveals that OpenAI's models, such as GPT-4 and 3.5, may memorize parts of copyrighted texts, including fiction books and New York Times articles. The research introduces a method to identify memorized training data, focusing on unusual terms, or "high-surprisal" words, within a larger body of work. This method suggests that models might memorize snippets during training, raising questions about the legality and ethics of using copyrighted data for AI development.
OpenAI maintains that their model development falls under fair use, but the plaintiffs argue that U.S. copyright law does not provide exceptions for training data. Despite having content licensing deals, OpenAI continues to advocate for fewer restrictions on using copyrighted data in AI training. The study underscores the need for transparency and scientific investigation into AI models to assess their reliability and ethical implications. As AI models increasingly rely on vast amounts of data, the debate over the use of copyrighted material in training processes remains contentious.
Key takeaways:
OpenAI is facing lawsuits for allegedly using copyrighted material without consent to train its AI models.
The study highlights how AI models, like GPT-4, may memorize parts of copyrighted texts, raising concerns about fair use.
The research introduces a method to identify memorized data in AI models using high-surprisal terms.
There is a call for greater transparency and scientific investigation into AI training practices to ensure reliability and legality.
