The information was confirmed by Nick Clegg - President of Global Affairs of Meta to Reuters in an interview this week. Accordingly, the data for training "multiple personalities" AI (which the company introduced on September 27) is taken from posts (written text, images) that users post publicly on Facebook, Instagram. Private content or content limited to friends, family, and chats are not used.
Clegg said Meta takes technical measures to control what data is mined by AI, but declined to elaborate on how the model works. "We exclude datasets that contain a lot of personal information, and we don't use data from sites like LinkedIn to train AI due to privacy concerns," Meta's leader stressed.
Public posts on Facebook, Instagram used to train Meta's AI
Meta CEO Mark Zuckerberg describes the Meta AI assistant as being able to “converse like a human” thanks to its large language model called Llama 2, combined with the Emu text-to-image model. The company’s assistant is capable of generating text, audio, images, and has access to real-time information thanks to a partnership with Microsoft’s Bing search engine – a platform that is also using the latest generation of ChatGPT.
Meta AI consists of 28 chatbots (automatic response programs) based on the stereotype of famous people. On the first day of testing, this tool encountered many controversies when it showed signs of toxicity, overreacting due to its... individual personality. Among them, some chatbots gave information that tended to be racist, delving too deeply into users' private lives. The problem was more serious when Meta employees themselves announced that they would not test this AI due to concerns about the unusual content given by the company's artificial intelligence.
Meta's use of user data posted on social networks to "feed" artificial intelligence also raises concerns about copyright issues. When asked whether Meta complies with steps to avoid copying copyrighted content, the company representative only mentioned the clause prohibiting users from creating content that violates privacy and intellectual property rights on the platform.
Source link
Comment (0)