Signal's Creator Is Bringing End-to-End Encryption to Meta AI
Ten years ago, Moxie Marlinspike brought end-to-end encryption to billions of WhatsApp users by integrating the Signal Protocol. Now he's doing it again — this time for AI chat. Marlinspike announced this week that Confer, his encrypted AI chatbot, will integrate its privacy technology into Meta AI.
Why This Matters
Right now, every AI chat conversation — across ChatGPT, Gemini, Claude, and Meta AI — flows through servers where the AI company can read, analyze, and store your data. As Marlinspike puts it, AI chat apps have become "some of the largest centralized data lakes in history, containing more sensitive data than anything ever before."
Think about what people share with AI chatbots: medical questions, financial information, relationship issues, half-formed thoughts they'd never share publicly. It's the unfiltered version of our inner lives, sent to an API endpoint specifically designed for extracting meaning and context.
Confer was built from the ground up so that nobody has access to your conversations but you — not even Marlinspike himself. Now that technology is coming to Meta AI at scale.
The WhatsApp Playbook, Redux
The parallel to WhatsApp is striking and intentional. In 2016, Marlinspike worked with Meta (then Facebook) to deploy the Signal Protocol across WhatsApp, enabling end-to-end encryption for over a billion users essentially overnight. It was one of the most significant privacy upgrades in internet history.
The Meta AI integration follows the same logic: take proven encryption technology and deploy it at scale through a platform that already has billions of users. Confer will continue operating as an independent entity while Marlinspike works on the integration.
"The world is changing quickly; we need to prepare for a future that is coming fast. I want that future to be safe, private, and accessible — which requires acting now, at scale."
Can You Actually Encrypt AI?
Encrypting AI conversations is harder than encrypting messaging. In messaging, the encryption is between two endpoints that both have keys. With AI, the model needs to process your text to generate a response — you can't encrypt it from the model itself. Confer's approach uses techniques like private inference to solve this, but the technical details of how it works with Meta's frontier models at scale remain to be seen.
Key Takeaways
- Moxie Marlinspike's Confer will bring encryption to Meta AI
- Repeats the WhatsApp end-to-end encryption playbook from 2016
- AI chats contain deeply sensitive personal data currently unprotected
- Confer continues as independent entity alongside the Meta integration
Our Take
If Marlinspike can actually pull this off at scale, it could be the most important privacy development in AI since... well, since he encrypted WhatsApp. The challenge is immense — AI inference is fundamentally different from message relay — but if anyone has the track record and the technical chops to solve it, it's the person who already did it once for billions of users. The fact that Meta is working with him rather than building their own solution also suggests they understand the credibility gap they face on privacy.
