Meredith Whittaker urges AI users to remember chatbots are not friends

Meredith Whittaker, speaking as a leader at Signal, wants people to remember one thing about AI chatbots: they are not your friends. In the same message, she stresses three boundaries in plain language: “These are not your friends. These are not conscious beings. These are not sentient interlocutors.”

That is the whole pivot, and it matters because many people already treat chatbots like something they can trust emotionally, socially, and morally. Whittaker is basically telling users and the companies behind these systems that the relationship should not be framed as companionship or personhood, because it is not.

To understand why this framing is not just philosophy, look at how modern AI products are experienced. Chatbots are conversational by design. They respond in natural language, they can sound empathetic, and they can keep going for a long time without any obvious “conversation is over” boundary. That makes them uniquely good at occupying the emotional space where humans usually live. When you put that capability next to ambiguity about what the system is actually doing, you get a dangerous mix: the interface feels like a social actor, while the underlying technology is a tool that generates text.

Whittaker’s insistence on denying personhood is also a strategic lever for companies building and governing these products. If your product marketing, onboarding copy, or user experience subtly encourages users to treat the bot as a friend, you are effectively expanding what users assume the system is. And those assumptions are exactly what regulators and courts tend to scrutinize when harm occurs. Even when the system is not conscious or sentient, people can still rely on it, build habits around it, and make decisions based on what it says. That reliance can have real-world consequences even if no consciousness is present.

There is also a board-level incentive to think about this “not a person” framing now. Public attention often moves faster than governance does. Once a product becomes widely used, the question stops being “Is it technically accurate?” and becomes “What did the company encourage people to believe?” Whittaker’s quote cuts directly against a common drift in the category: anthropomorphism. Many teams would rather describe the system as helpful, responsive, and caring because it improves engagement. But engagement is not safety. If the system behaves like a conversational partner, the company has to work harder to ensure users do not mistake the simulation for understanding.

Regulatory pressure is already moving in this direction across AI generally. Policymakers are trying to figure out how to require transparency and reduce deception when systems talk like humans. Whittaker’s message provides a clear “north star” for compliance discussions: do not imply consciousness, do not imply sentience, and do not position the system as a relational being. That can translate into specific design and governance choices: user education, limits on “care” language, disclaimers that are visible at the moments of interaction, and internal guardrails that keep product teams from drifting into quasi-personhood descriptions.

The second-order implication for executives is that this framing also changes risk allocation internally. If you accept that chatbots are not sentient interlocutors, then errors are not mystical. They are model limitations, data issues, and failure modes. That shifts the response from vague “the bot misunderstood” toward concrete engineering and process fixes. It also affects how customer support handles complaints. If a user says, “It promised me X,” the company should not treat that like a betrayal by a being. It should treat it like a harmful output or misunderstanding of capabilities, then address the system behavior that produced it.

Finally, there is competitive pressure. Leaders across the AI chatbot space are racing to increase trust and adoption. Whittaker’s stance is basically a reminder that the fastest route to durable trust is not pretending the system is a friend. It is being honest about what it is, and what it is not, at every step. For boards, that means monitoring not only model performance, but also the narrative the product tells users. In a world where AI tools can be emotionally sticky, clarity about non-friend status is not a footnote. It is part of product safety, brand credibility, and legal defensibility.