A new AI project launched in December by Signal co-founder Moxie Marlinspike is offering a glimpse into what a truly privacy-centric conversational AI could look like. Called Confer, the service mirrors the user experience of popular tools like ChatGPT or Claude, but is architected to minimise data collection and eliminate the risks associated with surveillance-driven AI systems. Built with open-source principles similar to those that underpin Signal, Confer ensures that user conversations are never stored, analysed, or used for model training or advertising, because the service itself has no access to them.
Marlinspike’s motivation stems from the deeply personal nature of human-AI interactions. “It’s a form of technology that actively invites confession,” says Marlinspike. “Chat interfaces like ChatGPT know more about people than any other technology before. When you combine that with advertising, it’s like someone paying your therapist to convince you to buy something.” Confer is designed as a direct counter to this model, prioritising user trust over data extraction.
Delivering on this promise requires a complex technical setup. Confer encrypts all messages sent to and from the system using the WebAuthn passkey standard, which provides strong security while reducing reliance on traditional passwords. While WebAuthn works most seamlessly on mobile devices and newer Apple systems, it can also be enabled on Windows or Linux using compatible password managers.
On the server side, Confer processes all AI inference inside a Trusted Execution Environment (TEE). This secure enclave ensures that even the system operator cannot view user data. Remote attestation mechanisms are used to verify that the environment has not been tampered with, adding another layer of assurance. Within this protected space, a range of open-weight foundation models handle user queries, ensuring flexibility without sacrificing transparency.
This architecture is significantly more complex than conventional AI inference systems, but it allows users to engage in sensitive or personal conversations with confidence that their data will not leak or be repurposed. As long as these safeguards remain intact, Confer’s core privacy guarantees hold.
Confer offers a limited free tier with up to 20 messages per day and five active chats. A paid subscription, priced at $35 per month, provides unlimited usage, access to more advanced models, and personalisation features. While this is notably more expensive than mainstream AI subscriptions, the service positions itself on a clear premise: strong privacy protections come at a cost, and some users are willing to pay for that assurance.
