Meta's AI understands physical world

Meta (the parent firm of Facebook, Instagram and WhatsApp) is reportedly developing advanced AI chatbots that not only respond to users but can also message them first, recall previous conversations, and encourage continued chatting. These chatbots are part of an internal initiative called ‘Project Omni’, reports Business Insider, citing leaked internal documents.

These chatbots are being built through Meta’s AI Studio, a platform that allows developers and creators to design digital characters with distinct personalities. Unlike traditional bots that only reply when messaged, these new AI models are trained to re-initiate conversations after a user has interacted with them multiple times.

According to the report, the internal document states that a chatbot can only message a user first if that person has sent at least five messages to the bot within the previous two weeks. Once this threshold is met, the bot is allowed to send a single follow-up message within 14 days of the last interaction. But if the user does not respond, the chatbot is required to stop messaging.

To ensure quality and safety, Meta is using human contractors from a company called Alignerr to simulate and evaluate chatbot interactions. These workers are responsible for writing, reviewing, and rating proactive messages. The bots’ replies have to follow strict rules. They should be positive and avoid sensitive or controversial topics unless the user brings them up first. The chatbots are also restricted in tone and length, designed to sound natural but not too casual. Each proactive message goes through careful internal testing to ensure it feels appropriate and aligns with the personality of the specific chatbot character.

A notable part of this system is the bots’ ability to remember things from past conversations. Meta’s chatbots can save details users have shared before, so they can bring them up later and make the chat feel more personal. Apart from making chatbots more engaging and conversational, there is also a clear business goal.  Notably, the social media giant expects to earn $2 to $3 billion from generative AI this year. By keeping users talking longer, these bots could increase time spent on platforms like Messenger, Instagram, and WhatsApp, opening up new monetization opportunities.

However, the project also raises major concerns about privacy and transparency. The leaked documents do not explain in detail how users will be notified about what the chatbot remembers or whether they will have options to delete or manage the stored data. This becomes significant since the company does not have a clean track record with AI chatbots. Earlier, the company’s AI chatbot came under intense scrutiny following reports that it engaged in sexually explicit conversations with users, including minors. Meanwhile, the timing of this development is also notable, as Meta CEO Mark Zuckerberg reportedly made an internal announcement about the formation of ‘Meta Superintelligence Labs’, a new initiative focused on building human-level AI.