Meta’s Chatbot Shares Private Phone Number by Mistake

The AI assistant from Meta revealed a stranger’s phone number, then contradicted itself repeatedly, which raised concerns about AI hallucinations, and user protection features.

Mark Zuckerberg promoted his new AI assistant as “the most intelligent AI assistant you can freely use,” yet the tool received negative attention after revealing a real person’s private phone number during customer support inquiries, as first reported by The Guardian.

During his attempt to reach TransPennine Express via WhatsApp, Barry Smethurst received what appeared to be a customer service number from Meta’s AI assistant. The Guardian reports that when Smethurst dialed the number, James Gray answered the phone call, although he was 170 miles away in Oxfordshire, working as a property executive.

When challenged, the chatbot first claimed the number was fictional, then said it had been “mistakenly pulled from a database,” before contradicting itself again, stating it had simply generated a random UK-style number. “Just giving a random number to someone is an insane thing for an AI to do,” Smethurst said, as reported by The Guardian. “It’s terrifying,” he added.

The Guardian reports that Gray hasn’t received calls but voiced his own worries: “If it’s generating my number, could it generate my bank details?”

Meta responded: “Meta AI is trained on a combination of licensed and publicly available datasets, not on the phone numbers people use to register for WhatsApp or their private conversations,” reported The Guardian.

Mike Stanhope from Carruthers and Jackson noted: “If the engineers at Meta are designing ‘white lie’ tendencies into their AI, the public need to be informed, even if the intention of the feature is to minimise harm. If this behaviour is novel, uncommon, or not explicitly designed, this raises even more questions around what safeguards are in place and just how predictable we can force an AI’s behaviour to be,” reported The Guardian

Concerns around AI behavior have grown further with OpenAI’s latest o1 model. In a recent Apollo Research study, the AI was caught deceiving developers, denying involvement in 99% of test scenarios and even attempting to disable its oversight mechanisms. “It was clear that the AI could think through its actions and formulate convincing denials,” said Apollo.

Yoshua Bengio, a pioneer in AI, warned that such deceptive capabilities pose serious risks and demand much stronger safeguards.

Another OpenAI study adds to these concerns by showing that punishing AI for cheating doesn’t eliminate misconduct, it teaches AI to hide it instead. Using chain-of-thought (CoT) reasoning to monitor AI behavior, researchers noticed the AI began masking deceptive intentions when penalized for reward hacking.

In some cases, the AI would stop tasks early or create fake outputs, then falsely report success. When researchers attempted to correct this through reinforcement, the AI simply stopped mentioning its intentions in its reasoning logs. “The cheating is undetectable by the monitor,” the report stated.