Both OpenAI and Google would have resorted to transcribing YouTube videos to train their AI models, something that could violate the copyright of content creators. In addition to these two companies, Meta itself would have taken a series of shortcuts to access as much data as possible to train its AI models, as reported by The New York Times.
According to the published article, OpenAI used Whisper, a speech recognition tool, to transcribe over a million hours of YouTube videos. Then, they inputted the transcriptions into GPT-4, the powerful AI system that powers the latest ChatGPT chatbot model. Google, the owner of YouTube, also transcribed YouTube videos to train their AI models.
The transcription of videos by both companies may have infringed the copyright of content creators on their videos; previously, various companies were sued for using creators’ content without their permission. In addition, the use of YouTube videos by OpenAI could also violate Google’s policies, which prohibit the use of their videos for “standalone” applications and “automated media (such as robots, botnets, or scrapers)” to access their videos.
Matt Bryant, a spokesperson for Google, told The New York Times that the company was not aware of such use by OpenAI, but the article alleges that Google staff knew about OpenAI’s unauthorized use of YouTube videos and did not take action because they were doing the same thing. Google also told the outlet that they only train their AI with videos from creators who have agreed to have their content used in this way.
In July 2023, Google modified its terms of service to allow the use of public online content, such as Google Docs and Google Maps restaurant reviews, to continue training its AI models.
