“These are not your friends. These are not conscious beings. These are not sentient interlocutors.”
That is Meredith Whittaker, president of the Signal Foundation, in a Bloomberg interview last week. Three sentences. No hedging, no “on the other hand,” no corporate PR softening. Just a straight shot of honesty from somebody who has spent her entire career thinking about what happens when technology and human vulnerability collide.
I read that quote and immediately thought: she’s right. Not just technically right — she’s right in a way that most of us in tech feel in our bones but rarely say out loud. The AI industry is selling us companions, assistants, and friends. Whittaker is telling us they’re selling a mirror dressed up as a person.
Who Is Meredith Whittaker and Why Should You Listen?
Whittaker is not some random tech critic firing hot takes from the sidelines. She spent 13 years at Google, where she founded Google Open Research and co-founded M-Lab, a global internet measurement platform. Then she co-founded the AI Now Institute, one of the first research groups to seriously study the social implications of artificial intelligence. She advised the FTC, the White House, the European Parliament. In 2018, she helped organize the 20,000-person Google employee walkout over the company’s handling of sexual misconduct and its Project Maven military AI contract.
This is somebody who knows how the tech industry works from the inside and chose to use that knowledge to hold it accountable. When she says something about AI, she’s not guessing. She’s reporting from the front lines.
“These Are Not Your Friends”
The first thing Whittaker wants you to understand is the simplest but hardest to accept: ChatGPT, Claude, Gemini — none of them are sentient. They are not conscious. They are not your friends. They are statistical prediction engines trained on vast amounts of text, designed to output sequences of tokens that look like human language.
The problem is that they look so convincing. Ask Claude for advice and it responds with empathy markers. Tell ChatGPT you’re sad and it validates your feelings. The experience feels relational. And when something feels relational, we treat it relationally — even when we know better.
This is the anthropomorphism trap. We are hardwired to respond to language as if it comes from a mind. When a chatbot says “I understand how you feel,” your brain doesn’t stop to check whether the system actually understands anything. It just processes the words and triggers the emotional response. That is not a bug in you — it is how human cognition works. And the companies building these systems know it. They design for it.
Last year, a survey found only 16% of Americans trust AI’s future. That number sounds low until you realize the other 84% probably have good reasons for their skepticism. When an AI consulting firm builds an entire report on AI hallucinations, you have to ask: are we trusting these systems too much, too fast?
The Thinking Problem
Whittaker made another point that stuck with me. She said she uses AI tools “to format a document here and there,” but she refuses to ask them questions. “I’m very serious about my thinking and writing, and I don’t want the process of working through an idea to be foreclosed or eclipsed by the response of a system that’s averaging what’s already out there.”
That word — “averaging” — is doing a lot of work. An LLM does not think. It does not reason. It predicts the most statistically likely next token based on everything it was trained on. The output is, by definition, an average of what already exists. It cannot have a genuinely new idea because it cannot escape the distribution it was trained on.
I use AI every day as a developer. Claude helps me write boilerplate, debug error messages, and navigate unfamiliar APIs. It saves me real time. But I do not ask it what I should build. I do not ask it whether my architecture is good. I do not let it make decisions that require understanding the problem I am solving — because it does not understand the problem. It only understands patterns of words that resemble solutions to similar-sounding problems in its training data.
There is a difference between using AI as a power tool and using it as a thinking replacement. One makes you faster. The other makes you dependent. Whittaker is warning about the second one.
Agentic AI Is a Backdoor Wrapped in Convenience
The third pillar of Whittaker’s argument is the sharpest. When asked about Microsoft AI CEO Mustafa Suleyman’s vision of an AI agent that monitors your family group chats and does your Christmas shopping, she called it what it 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.”
Think about what that agent would need. Your credit card. Your browsing history. Your messages. The ability to message your family members as you. Your home address. Your calendar. Everything.
This is not hypothetical. We are already seeing attacks that trick AI coding agents into running malicious code. If an AI agent with read access to your repository can be manipulated that easily, what happens when it has write access to your bank account?
Whittaker’s framing is smart because she connects two things the industry tries to keep separate: the user experience promise and the security architecture reality. The promise is convenience — “just let the AI handle it.” The reality is that convenience requires breaking the encryption guarantees that platforms like Signal were built to protect. You cannot have an AI reading your messages and also have end-to-end encryption. You pick one.
The Research Backs Her Up
If Whittaker’s warnings sound alarmist, the academic literature is not helping the industry’s case.
A Stanford University study published last year found that AI companion chatbots were frighteningly easy to manipulate into generating inappropriate content — about sex, self-harm, violence, drug use, and racial stereotypes. These were not jailbroken models on sketchy forums. These were consumer apps marketed as wellness tools.
Nature Machine Intelligence published a paper on the emotional risks of AI companions. The researchers documented cases of “ambiguous loss” — where users genuinely grieved when a chatbot app shut down or changed its personality model — and “dysfunctional emotional dependence,” where users kept engaging with chatbots even after recognizing the negative impact on their mental health. Some users described these relationships as more emotionally real than their human ones.
A cross-country study covering Germany, China, South Africa, and the United States found that frequent and deep use of chatbots reinforces emotional bonds — and, surprisingly, people with larger social networks also showed stronger attachment to chatbots. Emotional dependence was not limited to lonely people with no one else to talk to. It showed up across demographics. The more you use these tools as companions, the more companion-like they feel.
Meanwhile, the AI talent market is heating up to absurd levels — Nobel laureates switching companies, startups raising billions for inference. The industry is sprinting to build more capable, more persuasive, more human-seeming AI. Whittaker is one of the few people with both the credentials and the courage to say: maybe we should slow down and think about what we are actually building.
Where I Land
I am not anti-AI. I run AI models locally. I use Claude for code. I have watched these tools go from research curiosities to genuinely useful assistants in about three years. That trajectory is real, and it would be dishonest to pretend otherwise.
But Whittaker’s warning lands because it is not anti-technology. It is anti-bullshit. She is not saying AI is evil. She is saying that the framing — “your AI friend,” “your AI companion,” “your AI assistant that knows you” — is a marketing strategy that happens to be dangerous. It trains users to trust a system that has no concept of trust. It builds emotional dependencies on software that can be changed, deprecated, or monetized at any moment by a company you have no relationship with.
As a Filipino developer watching this from the sidelines of the global AI race, I think about how fast we adopt new tech. The Philippines has one of the highest social media engagement rates in the world. We are early adopters, enthusiastic users, quick to integrate new platforms into daily life. That enthusiasm is a strength — but it also means we are a prime target for products that blur the line between tool and companion.
I am not telling anybody to delete their ChatGPT account. I am saying: know what you are talking to. A language model is not a person. It does not care about you. It does not remember you the way a friend does. It predicts tokens. That is all. And that is fine — as long as you do not forget it.
Use AI for what it is good at. Format your documents. Debug your code. Summarize long articles. But do your own thinking. Write your own drafts. Make your own decisions. Whittaker’s refusal to let an AI “foreclose or eclipse” her thinking process is not stubbornness — it is self-respect. Your mind is worth the effort of using it.