Meta’s New AI Assistant: Trained with Public Data

At Connect 2023, Meta rolled out a slew of AI editing tools and features, introducing AI stickers and image editing capabilities like restyle and backdrop. The showstopper, however, was the unveiling of Meta’s new AI assistant, set to join WhatsApp, Messenger, and Instagram in the coming days. But, the method used to train this assistant might not sit well with everyone.

According to reports, Meta used public Facebook and Instagram posts to train portions of its new Meta AI virtual assistant. The company assured that private posts shared exclusively with family and friends were excluded to respect consumers’ privacy.

Meta’s President of Global Affairs highlighted that private chats on messaging services were also not included in the training data. Meta took steps to filter out private details from public datasets used for training.

The majority of the data used for training was publicly available, with privacy concerns leading to the deliberate omission of certain platforms such as LinkedIn.

Meta’s development of the assistant involved a custom model based on the Llama 2 large language model and a new model named Emu for generating images in response to text prompts. Real-time information will be accessed through a partnership with Microsoft’s Bing search engine.

Public Facebook and Instagram posts played a role in training Meta AI, with safety restrictions in place to prevent the creation of photo-realistic images of public figures.

Meta acknowledges potential litigation regarding copyrighted materials, particularly in terms of whether creative content falls under the existing fair use doctrine. The company anticipates legal battles in this regard.