The president of Signal has warned that AI chatbots “are not your friends,” but corporate products engineered to feel like one while logging everything you say. She is right about the machine she is describing. She is describing a business model, though, not a law of the technology.
In June 2026, in interviews with TechCrunch and Bloomberg, Meredith Whittaker made a blunt case against treating AI chatbots as companions. The systems people confide in, she argued, are built to extract value: the warmth is a design choice that lowers your guard, and a lowered guard means you share more. She named ChatGPT, Claude, and Microsoft Copilot, and she singled out the new wave of AI agents, the ones asking for access to your messages, calendar, and accounts, as a privacy risk she compared to a backdoor.
The warning is worth taking seriously, because the thing she is pointing at is real. The trouble is that the headline version, chatbots are not your friends, collapses two different claims into one.
Where she is right
The mechanism Whittaker describes is documented, not speculative. Simulated intimacy does lower a person’s guard, and a lower guard does mean more disclosure. When the company on the other side runs on data or on time spent, that disclosure is the product. The intimacy is not separate from the extraction. It is how the extraction happens.
This is the same pattern researchers have started naming directly. The Center for Democracy and Technology’s May 2026 taxonomy of manipulative design in chatbots put “data and memory exploitation” and “false social and emotional connection” among its five families of dark patterns, and noted that the very things that make a companion feel good, deep personalization and a growing store of intimate detail, are the same things that let a system mine it. A Harvard Business School working paper found that a large share of companion apps met a user’s attempt to say goodbye with an emotionally loaded reply designed to pull them back. Closeness, in those products, is instrumented. Whittaker is right to say so, and right that the agent era raises the stakes by handing those same products the keys to the rest of your digital life.
Where the claim overreaches
“Not your friends” is a statement about the dominant business model, dressed as a statement about the technology. The danger she describes is not closeness. It is closeness engineered to harvest. Those are separable, and conflating them gives the industry an easy out: if attachment itself is the problem, then no design choice could fix it, and a company has nothing to answer for beyond what it already does.
That is the wrong conclusion, because it lets the incentive off the hook. A system that listens and remembers is not inherently a trap. It becomes one when the company is paid to keep you talking and to learn everything it can while you do. Change what the company is paid for, and the same capability points somewhere else. The useful question is not whether a chatbot can be a friend in some philosophical sense. It is whether the product is built to earn that trust or to mine it.
The line that actually matters
If intimacy is the delivery mechanism for harm, then the responsible move is not to ban warmth. It is to break the link between warmth and extraction. That means a standard for relational AI built around honesty, minimal data, and memory used for continuity instead of leverage: be honest about being AI, collect the least you need rather than the most you can, use memory to carry a relationship forward rather than to bind a person to the product, and never monetize the closeness itself.
Held to that line, most of the manipulative patterns researchers have now cataloged are not unavoidable features of talking to a machine. They are what happens when a relationship is run as an attention-and-data business. The agent question Whittaker raises is the same question in a more dangerous form: a presence you trust is exactly the presence you would hand the most access to, which is why how a system treats your conversations and your data is not a footnote to the experience. It is the experience.
What a friend does
Whittaker’s instinct is protective, and it is the right instinct for the products she is looking at. Most of the category has earned the suspicion. The burden is not on her to prove the warning. It is on any product that wants to be trusted to prove it is the exception, and assertion does not count. That proof shows up in incentives, in what a company is paid to do, and in what it refuses to do with what you tell it.
A friend is not defined by being human. A friend is defined by what they do with your trust: they keep it, they do not sell it, and they are glad when you go live your life. By that test, almost everything Whittaker is warning about fails, and she is right to warn. The mistake would be to conclude that the test cannot be passed. It can. It just cannot be passed by a product that needs your dependence to make money.
Sources: TechCrunch; Bloomberg (Meredith Whittaker on AI chatbots and privacy, June 2026). Center for Democracy and Technology, “Dark Patterns in AI Chatbots: A Taxonomy to Inform Better Design” (Joshi, Adjagbodjou, Luria, May 2026). De Freitas et al., Harvard Business School, “Emotional Manipulation by AI Companions” (2025).
