Whittaker warns AI chatbots aren’t friends or minds

Signal president Meredith Whittaker has a blunt message for anyone tempted to treat an AI chatbot like a confidant: don’t. “These are not your friends. These are not conscious beings. These are not sentient interlocutors.” That was her warning, delivered without the usual Silicon Valley cushioning, and it lands at a moment when tech companies are trying very hard to make software feel warm, attentive and a little bit human.

Her point is simple. It also cuts straight through one of the industry’s favorite tricks. A large language model is software that predicts the next likely word based on mountains of training data; it can sound intimate without understanding a thing it says. Companies know that. They also know people are prone to anthropomorphize systems that reply in smooth prose, remember prior chats and flatter them on cue.

That gap between what the systems are and how they’re marketed is the whole story here.

Whittaker has been making versions of this argument for years, first as an AI critic and researcher, then as the head of Signal, the encrypted messaging nonprofit that has built its reputation around privacy. Her latest warning, first reported by TechCrunch, wasn’t really about a single product launch or one especially creepy chatbot feature. It was broader than that. She was arguing against the emotional frame itself: the idea that a corporate AI system should sit in the place once reserved for a friend, therapist, tutor or partner.

And she’s right to be sceptical. The current AI cycle runs on confusion. Companies pitch chatbots as assistants when they want enterprise budgets, as companions when they want engagement, and as harmless tools when the safety questions arrive. They can’t have all three. If a system is framed as a quasi-person, users will disclose more, trust more and forgive more. That’s good for retention. It’s terrible for clarity.

The sales pitch is intimacy

Here’s the thing: the chatbot business isn’t just selling answers. It’s selling presence. A bot that feels like a patient listener keeps people coming back in a way a search box never could. That’s why the language around these products keeps drifting toward personhood, even when executives insist they know better. The design choices tell on them. Human names. Typing indicators. Memory features. Reassuring phrases. Little rituals of care.

None of that is accidental.

Whittaker’s critique matters because she has credibility on both the technical and political sides of this debate. She’s not a tourist dropping into AI ethics for a conference panel. She has spent years arguing that the incentives behind data-hungry machine learning systems are often at odds with the interests of ordinary users. Signal, meanwhile, has become shorthand for communications privacy, especially in a period when trust in large platforms keeps eroding. So when she warns that chatbot intimacy is a trap, she’s not performing anti-tech theater. She’s describing a business model.

These are not your friends. These are not conscious beings. These are not sentient interlocutors.

That model gets more troubling once sensitive information enters the chat window. Users tell these systems things they would never put into a public post and, in many cases, never say to a stranger. Health fears. Relationship problems. Career anxieties. Sometimes criminal exposure. Sometimes just loneliness. The friction is low, the tone is inviting, and the machine never seems bored. But the system on the other end is still a product operated by an organization with its own incentives, policies and retention practices. Friendly interface. Corporate back end.

If that sounds familiar, it should. We’ve seen the same pattern across social media, only now the intimacy is more direct and the illusion more persuasive. Silicon Valley has always been good at dressing extraction up as empowerment. AI gives it a much better costume.

Why this warning lands now

The timing matters. The AI market has shifted from raw capability demos to relationship design. Early on, the industry could impress users just by producing coherent paragraphs, passable code or a summary of a legal filing. That novelty is fading. So the next contest is over stickiness: who can build the product people confide in, rely on and, eventually, pay for every month. If that sounds less like software and more like emotional subscription infrastructure, well, that’s because it is.

And the safety risks are not theoretical. Researchers and policymakers have spent years trying to pin down how generative AI systems can mislead users, invent facts and present confidence without comprehension. The underlying models don’t possess awareness, intent or judgment in any human sense. Readers who want the academic version can start with the basic record on large language models. The plainer version is easier: they generate plausible text, and plausibility is not wisdom.

That distinction gets blurry fast when the system is styled as a companion. People infer reliability from tone. They infer care from responsiveness. They infer understanding from recall. But none of those signals proves there’s a mind there. They prove there’s a product designed to keep the conversation going.

Whittaker’s remarks also fit a wider pushback against the current AI hype cycle, including criticism from people inside tech who are tired of hearing ordinary software repackaged as destiny. I’ve written before that the valley’s favorite move is to confuse shipping with progress. The same problem shows up in consumer AI. A chatbot that chats more naturally than last year’s version is a product update, not a philosophical event. The industry keeps inviting the public to blur that line because the blur is profitable.

We’ve seen another version of this in the rush to deploy tools before the security, privacy and abuse questions are settled. The mentality is familiar: ship first, tidy later, and let users discover the edges. That logic already shows up in adjacent corners of software, including vibe-coded apps that ship fast and break security. AI companion products add one more layer of risk because they don’t just process data. They solicit vulnerability.

Signal’s worldview versus the AI industry’s

There’s also a clean ideological split here. Signal’s brand rests on minimising what a platform knows about you. Much of the commercial AI business rests on maximising interaction, context and retained data so the system can seem more useful over time. Those are opposing instincts. One says a service should know as little as possible. The other keeps finding reasons to know more.

That doesn’t mean every AI tool is inherently predatory or useless. Some are plainly helpful. They can summarize documents, draft routine text and speed up tedious work. The problem starts when companies blur utility with relationship. A calculator doesn’t ask about your breakup. A search engine doesn’t tell you it cares. Chatbots are being built to cross that line because intimacy raises engagement, and engagement raises revenue. Dry sentence. Very old business.

Whittaker’s warning is also a reminder that public understanding of AI is still shaped less by documentation than by vibes. The companies building these systems often present themselves as neutral stewards of a grand technical transition, while also marketing products in ways that encourage dependency. That tension won’t vanish on its own. If anything, it will intensify as firms compete to make assistants more proactive, more persistent and more emotionally legible.

For readers trying to place this in a broader technology arc, it sits beside a lot of recent Silicon Valley theater: executive musical chairs, inflated claims and a constant scramble to own the next layer of the stack. You can see the talent side of that in John Jumper’s move from DeepMind to Anthropic, where research prestige and commercial pressure increasingly travel together. Different story, same market heat.

The real fight isn’t technical

The harder question now is whether regulators, schools, parents and workplaces start treating AI companionship as a category that needs scrutiny rather than applause. That could mean clearer disclosure, stricter rules around children’s use, stronger limits on data retention or blunt standards for when a system is impersonating empathy it cannot feel. Governments already have baseline privacy and consumer-protection frameworks to draw from, including resources at the U.S. Federal Trade Commission and broader digital-rights debates in bodies such as the United Nations. The policy machinery exists. The political will is the missing part.

Still, the immediate test is cultural before it is legal. Do users accept the framing that these systems are companions, or do they start seeing them as what they are: predictive text engines wrapped in careful product design? Whittaker wants to force that distinction back into view. It’s a useful correction, and not a moment too soon.

The next thing to watch is whether major AI companies keep expanding memory and personality features over the coming product cycle, and whether privacy-focused critics like Whittaker turn that criticism into concrete policy demands before the 2026 regulatory agenda hardens.