Signal CEO: AI Chatbots Are Not Your Friends

Signal President Meredith Whittaker has a blunt message about the AI chatbots millions of people now talk to every day: they are not your friends.

The comment is trending because Whittaker delivered it in a wide-ranging Bloomberg interview on policy, privacy, and Signal, and she did not soften it. Asked directly about the privacy implications of chatbots like ChatGPT and Claude, she said plainly that these are not your friends, these are not conscious beings, and these are not sentient interlocutors.

That framing cuts against the way the products are increasingly marketed and used, as companions, confidants, and always-on advisors.

Why the head of an encrypted messenger is drawing this line

Whittaker runs Signal, the nonprofit behind the end-to-end encrypted messaging app, so privacy is the lens she views everything through. Her argument is less about whether a model feels human and more about what you hand over when you treat it like a person you trust.

She is not refusing the technology outright. Whittaker acknowledged that she uses AI tools to format a document here and there. But she drew a hard boundary around her own thinking.

In her words, she does not ask them questions because she is serious about her thinking and writing and does not want the process of working through an idea to be foreclosed or eclipsed by the response of a system that is averaging what is already out there.

The averaging problem

That phrase points at a real mechanical limit. A large language model generates its answer by predicting likely text based on patterns across its training data. The output trends toward the consensus, the common phrasing, the middle of the distribution.

For formatting and boilerplate, averaging is fine. For original reasoning, Whittaker's worry is that leaning on the model nudges your own conclusions toward whatever is already most common, before you have done the harder work yourself.

How an AI agent becomes a privacy problem

The sharper part of her warning is about the next wave of AI: agents that act on your behalf rather than just answer questions.

She responded to a prediction from Microsoft AI CEO Mustafa Suleyman that users could let Microsoft Copilot handle all of their Christmas shopping this year. On the surface that sounds convenient. Whittaker walked through what it would actually require.

For Copilot to figure out who wants what by eavesdropping on the family group chat, she said, you would be giving it access to your credit card, your browser, your Signal, the ability to message your siblings on your behalf, your home address, and your calendar.

What you have just described is a system with very pervasive access across multiple applications and services. In the context of Signal, it would constitute a kind of a backdoor.

That is the crux. An assistant that can shop for you needs to read your conversations, spend your money, and impersonate you in messages. The capability and the surveillance are the same thing.

Access, mapped out

Here is the access footprint Whittaker described for a single consumer task like holiday shopping.

For an encrypted messenger, that last layer is the alarming one. The entire promise of end-to-end encryption is that no one but the participants can read a conversation. An AI agent sitting inside the app, reading messages to act on them, is by definition a party that can see the plaintext.

What to watch over the next 24 to 72 hours

Expect Whittaker's not your friends line to spread well beyond the original interview, because it lands as a clean rebuttal to the companion framing that AI companies have leaned into.

Watch for responses from the agent camp. Suleyman's Copilot shopping pitch is now the named example of the convenience-versus-access tradeoff, and that puts pressure on Microsoft and rivals to explain how an assistant gets this access without becoming the backdoor Whittaker describes.

The practical question for users is permission scope. As agentic features roll out, the thing to scrutinize is not how smart the assistant is but how much it asks to touch. An agent that wants your card, your calendar, your browser, and your messages all at once is asking for the keys to nearly everything.

• Read the permission prompts. Convenience features quietly request broad, persistent access.
• Separate formatting help from decision-making. Whittaker uses AI for the former, not the latter.
• Treat any AI that reads your encrypted chats as a third party in the room, because functionally it is one.

Whittaker's point is not that the tools are useless. It is that calling them friends obscures what they are: systems, run by companies, that work best when they can see the most. The more capable the agent, the more it has to watch.