Confer and the Case for Confidential AI

As AI increasingly becomes the digital mirror reflecting our doubts, ideas and even secrets, Confer’s value proposition provides an unsettling reminder that perhaps we are giving away more than we get in return.

Because Confer is not just another conversational AI assistant: it is a critical response to the direction LLMs are travelling in; an attempt to to return to us absolute ownership of our data and thinking. Created by Moxie Marlinspike, who rewrote the rules of messaging with Signal, the goal is to do the same for AI he did for private communication. It’s not a full application, but the name of the encrypted AI engine that powers it.

Unlike ChatGPT, Gemini or other platforms that have turned user conversations into raw material to train AI, optimize products or even resist judicial orders to retain complete interaction logs, Confer’s model is based on the principle that what is discussed with AI is a private conversation, not a “piece of data” to be exploited.

Confer is built on several technical pillars: every user input and every response is encrypted from source to destination so that only the user himself possesses the keys needed to decrypt them. That principle, end-to-end encryption, popularized by Signal and now applied to the domain of conversational AI, ensures that not even the platform itself can access content. When the data is not technically accessible, it cannot be handed over to third parties, whether that’s data brokers or the police.

Confer goes beyond simple message cryptography. To execute inference (the moment your question reaches a language model and a response is generated), each user input is encrypted locally…, and is only decrypted within the Trusted Execution Environment (TEE), isolated even from the provider. Thanks to remote attestation, the user can verify that the code that processes their data is correct and has not been altered.

This combines end-to-end encryption, modern passkey-based key synchronization, confidential execution and cryptographic transparency. It is not a legal promise of privacy in the form of some kind of clause in fine print, but a guarantee based on cryptographic design and verifiability.

Why is Confer worth trying? Because we’re in a phase of AI adoption where few products question the logic of “data for service,” and even fewer offer a usable experience without sacrificing confidentiality. The illusion of privacy offered by most AI assistants comes with policies that allow the provider to analyze, use, or reuse your conversations, which Confer explicitly doesn’t do.

For security-conscious users, organizations that handle sensitive information or simply people who don’t want their thoughts to be a training token for a business model, this proposal is a game-changer. And why else is it worth trying? Because this is the best personal dialogue experience with an AI chatbot I’ve had since I started testing them, and that was some time ago. Let’s be clear what this is: a personal dialogue, without the possibility, for the moment, of uploading documents or sharing links. But it does what it does phenomenally well.

Then there’s the Moxie Marlinspike factor. The co-author of the encryption protocol that drives Signal and underpins the security of services such as WhatsApp or Google Messages. His background as a cryptographer and digital privacy advocate gives him technical and moral authority over how communications should be protected in the digital age. His critical response to the current state of AI is practical. On his blog, Marlinspike argues that the conversational format of LLMs tends to induce people to reveal their deepest thoughts, not just keywords. When we converse with AI, we not only transmit data, but also reasoning structures, uncertainties and private mental patterns. He believes that act of “digital confession” deserves the same guarantees that we take for granted in a personal conversation, rather than being raw material for Big Tech.

Some will debate the technical complexity, usability challenges or even the economic limitations of an AI model that prioritizes privacy over massive data collection. What does a model like this intend to live on? A freemium model: a free account gives access to a basic model, twenty messages per day and up to five active chats. A monthly fee of $34.99 gives unlimited access to a more advanced model that can be customized.

Confer’s value proposition isn’t a perfect product from day one, it’s an alternative that breaks with the dominant logic and forces us to rethink what we mean by “AI service”. It’s not just another chatbot: it’s a statement of principles about who controls what we think, what we ask and what we store in a machine’s memory. And previous conversations in ChatGPT or Claude can be imported.

At at time when privacy is measured in opaque clauses and opt-outs, Confer proposes a renaissance of confidentiality, not as an optional extra, but as the very foundation of interaction with artificial intelligence. And that, precisely, is what makes it an idea worth exploring.

—

This post was previously published on Enrique Dans’ blog.

***