The Meta AI app and Meta AI on WhatsApp have a new "incognito chat" option, which Meta CEO Mark Zuckerberg said is a "completely private way to interact with AI."

Zuckerberg also said that Meta AI's incognito mode is the first major AI product where there is no log of conversations stored on servers. Zuckerberg likened the feature to end-to-end encryption, and said no one will be able to read the AI conversations, not even Meta or WhatsApp.

AI inference for incognito chat is done in a Trusted Execution Environment that Zuckerberg said is not accessible to Meta. Conversations also disappear from the phone when exiting a chat session, and nothing is saved or logged. Web searches are conducted privately, with no search information linked to the user.

"To get the most from personal superintelligence, we'll all need ways to discuss sensitive topics in ways that no one else can access," Zuckerberg said.

WhatsApp head Will Cathcart told reporters that the AI has safety guardrails, and it will refuse to answer questions that could be interpreted as harmful or illegal, steering conversations in a different direction. The mode also only supports text, and users are unable to upload images.

Incognito chat for Meta AI comes as OpenAI is facing a lawsuit for allegedly causing a teen's drug overdose. The teen asked ChatGPT for information on whether it was safe to take two drugs together, and was provided with an incorrect answer that led to his death. OpenAI has been sued several times by the families of people who used ChatGPT before dying by suicide.

Lawsuits against OpenAI have involved chat logs recovered by the plaintiffs, and without those logs, there would be far less evidence for a legal complaint over AI actions and advice.

Google and OpenAI also offer temporary chat options, but messages are still stored on remote servers. Google keeps data for up to three days, and OpenAI keeps logs for 30 days.

Meta's private chat option is rolling out in the coming months in the Meta AI app and WhatsApp.