WhatsApp Launches Incognito Chat Feature With Meta AI

WhatsApp announced on Wednesday it is launching Incognito Chat, a feature that runs AI inference inside a Trusted Execution Environment so that Meta cannot access users' conversations with its AI chatbot. The system relies on the company's Private Processing scheme, introduced a year ago, which already supports features such as message summarization.

Meta said Incognito Chat is ephemeral by default and conversations disappear once a user exits the session. WhatsApp head Will Cathcart told WIRED the company may develop an option to retain conversation history for users who want it. The feature is currently text-only, with image processing and voice recognition capabilities in development.

Cathcart described the engineering challenge as building something in a data center that has the security properties of a giant phone without giving Meta the passcode. Reducing latency through optimized routing was critical to making the experience usable given the added demands of the secure environment.

Meta CEO Mark Zuckerberg wrote in a post that he is proud Meta Superintelligence Labs is the first lab to deliver private AI. The company is also adding Incognito Chat as an option in the Meta AI app. A related feature called Side Chat with Meta AI lets users ask the chatbot about an ongoing conversation with friends or a group without sharing that context directly.

Meta invites third-party audits and vulnerability reports on Private Processing. The company said Incognito Chat itself will undergo expert oversight to verify the durability of its code. " With more than 3 billion users worldwide, Incognito Chat could give many people their first experience interacting with an AI chatbot.

Most generative AI platforms offer some form of incognito mode. Zuckerberg highlighted that conversations on the phone disappear when the session ends. The rollout comes as Meta eliminated opt-in end-to-end encryption from Instagram Direct Messages this week after years of promising default encryption.

Separately, Anthropic head of product Cat Wu discussed the company's approach to AI development in an interview at the Code with Claude conference in San Francisco. Wu said the team focuses on staying at the frontier of improving models rather than reacting to competitors, noting that competitor focus can leave a company perpetually behind.

Anthropic released at least six models last year and has already released nearly as many this year. Wu expressed hope that this pace continues, though deployments may vary for safety reasons, as seen with the limited release of its cybersecurity model Mythos to select partners including Amazon, Apple, CrowdStrike and Microsoft.

Wu said managing AI agents resembles managing people, requiring domain expertise to debug mistakes and clarify instructions. She anticipates AI handling tedious tasks such as email responses, allowing humans to focus on more creative work and potentially reducing the need for some entry-level roles while enabling greater overall productivity.

The company is particularly excited about proactivity in the next six months, where Claude would understand a user's work and automatically set up automations rather than requiring synchronous prompts.