The debate over using copyrighted content to train generative AI models has entered a new phase. New documents in a legal case show Brent Hecht, director of applied sciences at Microsoft, suggested in statements filed in the case that people around the world might view training large language models using Internet content as a kind of massive theft of human labor.
According to ITNA citing Techspot, these remarks come as Microsoft and OpenAI argue in a lawsuit related to The New York Times that using copyrighted material to train AI models can fall under the concept of "fair use." They also believe that AI model outputs sometimes differ sufficiently from the original content and are transformative in nature.
However, internal documents submitted to the court also raise examples of AI models' ability to reproduce published content. Greg Brockman, an OpenAI executive, acknowledged in statements that ChatGPT can predict and complete some sentences from The New York Times when working with its material. According to the case documents, Hecht also said defending such a situation could weaken the concept of fair use.
OpenAI has previously stated that training practical AI models without access to copyrighted content is very difficult or impossible. Meanwhile, various tech companies continue to use large-scale data to train their models; a matter that has raised legal and ethical debates about user data, news articles, books, and other Internet content.
Another part of The New York Times' dispute with AI companies concerns the impact of this technology on the news industry. AI-generated summaries in search engines can present information contained in articles without requiring users to visit publishers' websites directly. In recent documents, one OpenAI executive also described the publishing and media industry as facing serious threats from the impact of AI.