Meta used public Instagram and Facebook posts to train parts of its new AI assistants, but the company stopped short of including private posts with family and friends and private chats on its messaging service, Reuters reports.
"We've tried to exclude datasets that have a heavy preponderance of personal information,” Meta President of Global Affairs Nick Clegg told Reuters at this week's Meta Connect conference. Clegg adds that the “vast majority” of the data Meta used to train its AI assistants is publicly available.
Earlier this week, Meta announced a line of AI chatbots across its messaging platforms including WhatsApp, Instagram, and Messenger. The chatbot can help with things like answering questions, planning trips, and generating images from text prompts. The assistant will also be able to provide real-time results from Microsoft’s Bing search.
Meta’s model was based partially on its LLaMA 2 language model, which was released in July.
In addition to its general assistant, the company is rolling out 28 AI-powered characters across its messaging platforms, including a few modeled after celebrities such as Snoop Dogg, Kendall Jenner, and Tom Brady. The chatbots launched in limited beta on Wednesday, but have yet to roll out to the general public.
Many companies have recently come under fire for using web content to train their AI without the permission of the original author or artist. That includes Elon Musk's xAI, which plans to train its AI on tweets, prompting artists who share their content on Twitter to threaten to leave the site.