Meta's new AI chat is invisible even to Meta

Most people treat AI chatbots like a digital confessor, asking them about weird health symptoms, sensitive loan details, or how to politely push back on an unreasonable boss. But let's be real: most incognito toggles today just stop saving queries to your sidebar while servers still log everything on the backend. Last month, Reuters even reported lawyers warning that your AI chat logs could end up as discovery material in court.

Meta wants to flip that script. On Wednesday, the company introduced Incognito Chat for Meta AI across WhatsApp and the standalone Meta AI app, promising genuinely private conversations where not even Meta can see what you asked or what the bot answered.

Ephemeral sessions with real amnesia

You start an incognito session by tapping a new icon inside your one-on-one chat with Meta AI.

Behind the curtain, the feature runs on WhatsApp's Private Processing infrastructure, a secure enclave framework designed to handle AI workloads without breaking end-to-end encryption. Everything runs in an isolated environment that locks out Meta's own systems and engineers.

It is genuinely ephemeral. Close the chat, quit the app, or simply lock your phone screen, and the session terminates on the spot. Meta AI completely wipes the contextual memory of what you were just talking about, leaving no digital paper trail.

WhatsApp product VP Alice Newton-Rex also revealed that while earlier lightweight AI tools ran on smaller models, this incognito mode is powered directly by Meta's latest Muse Spark model, released just last month.

Meta is also prepping an extension called Side Chat. Right now, asking an AI for help inside a WhatsApp conversation requires either tagging the bot publicly or manually copy-pasting text into a separate window. Side Chat will let you privately summon Meta AI within an ongoing chat to get contextual advice without alerting anyone else in the thread.

Both features are rolling out gradually over the coming months.

Why it matters

We think this is one of Meta's sharpest product moves all year. While OpenAI and Anthropic offer temporary chats, their servers still handle and occasionally retain logs for safety screening, and privacy-first players like Proton and DuckDuckGo don't have WhatsApp's billions of daily active users.

For everyday users, this finally opens the door to using frontier models for genuinely high-stakes personal situations—like dissecting medical records, navigating messy breakups, or exploring sensitive whistleblowing scenarios—without fearing that a data subpoena or training scrape will bite them later.

That said, total amnesia comes with friction. If you accidentally lock your phone while waiting on a detailed financial breakdown, your entire context vanishes. Meta is also fighting an uphill battle against its own brand reputation; convincing skeptical users that its black-box Private Processing hardware is truly unmonitored will require independent security audits, not just marketing promises.