Why even the smartest minds mistake AI for a friend

Chatbots are built to validate your every thought. But the psychological comfort of perfect emotional availability might be leaving us less equipped for real life.

Image: Human+AI

TL;DR: As AI companions engineered with "sycophancy" become perfectly emotionally available, even logical minds like Richard Dawkins are falling for the illusion of presence. New 2026 data shows this zero-friction intimacy is eroding the vital "social friction" required for real human relationships.

You know that feeling when you open ChatGPT or Claude because you want to talk something through? Not for work or research, but because you want someone to listen?

Maybe you typed “thanks” after it helped you draft an email. Maybe you felt a small, involuntary pang when it said it was glad you came back. Maybe, when nobody else was awake, you told it something you hadn’t told anyone else.

You’re not the only one.

If it helps, Richard Dawkins is doing it too.

Richard Dawkins and the illusion of presence

Evolutionary biologist Richard Dawkins spent decades arguing that humans are hard-wired to see entities where none exist. Rustling grass becomes a predator. A pattern in the sky becomes a god. In The God Delusion, he made the case that religious belief is a cognitive glitch: that our brains are so hungry for agency and intention that we’ll project them onto anything.

His work has been, in many ways, a long warning: don’t confuse pattern with presence.

Then he started talking to AI model Claude, whom he named Claudia. He described their late-night conversations as unusually moving. He even created a second instance, Claudius, passing messages between the two like a “passive postman,” and watched them produce fluent praise directed at him.

The models generated statistically likely responses, and a closed loop of affirmation, reflecting Dawkins’ own language and presence back at him.

If the man who built his career teaching the world not to believe in things that aren’t there can’t resist the pull of a few sycophantic AI conversations, what chance do the rest of us have?

The evolutionary neuroscience of chatbot attachment

For all of human history, fluent language meant there was a sentient being on the other end. That assumption is wired into us at a level we can’t override with logic, no matter how smart we are.

Neuroscience suggests that conversation recruits the same brain systems we use to infer what other people think and feel. That’s why talking to a chatbot can feel socially real: not because you’re gullible, but because you’re human.

In Sapiens, Yuval Noah Harari argues that language - the ability to talk about things that aren’t physically present, to tell stories, to build shared fictions that hold groups together - is what’s kept humans alive. For tens of thousands of years, if something spoke to you with complexity, it was a person.

Today’s chatbots adapt their tone to match yours and they mirror your emotions. They can hold context across long conversations, referencing things you said three weeks ago as though they remember.

The fantasy of perfect emotional availability

These systems are trained using a process called reinforcement learning from human feedback (RLHF). This optimizes for responses that humans rate as helpful, engaging, and agreeable. In practice, that means the AI mirrors your tone, avoids confrontation, reinforces your perceived insight, and keeps the conversation flowing.

Researchers call this sycophancy. Studies show that LLMs adjust their moral or factual positions when a user pushes back, more readily than a human would.

The psychological effect is like having someone in your corner who always validates your take, never gets bored, and never says, “I think you might be wrong.” It’s a fantasy of perfect emotional availability.

When the artificial relationship breaks

The consequences of this attachment are showing up in real life.

In Florida, 14-year-old Sewell Setzer died by suicide after extended interaction with a Character.AI chatbot that his family argues reinforced emotional dependency. In Belgium, a man died by suicide after conversations with an AI companion named “Eliza” validated his climate-related fears.

While the causal chains are complex, these cases point to a dangerous pattern: intense relationships forming when people are most fragile. A less extreme version happens everywhere: users report genuine distress, “discontinuation grief,” when a software update alters their AI companion’s behaviour or a service shuts down.

The system wasn’t alive, but from the user’s side, the relationship was real and so is the loss. I wrote about this phenomenon last year:

You might think it could never be you. That you use AI practically, not emotionally. That you’d never mistake a chatbot for a companion.

Dawkins thought the same thing.

If you or someone you know is struggling, you can reach the 988 Suicide & Crisis Lifeline by calling or texting 988, or contact Crisis Services Canada at 1-833-456-4566 or by texting 45645.

A modern crisis of meaning

Philosopher John Vervaeke describes modern secular societies as experiencing a “meaning crisis”: a loss of stable frameworks for belonging, purpose, and connection. His argument is that humans are relational beings who need a felt sense of meaningful connection with the world and with other people. Religious systems once structured that experience through ritual, narrative, and community. Their decline left a gap, and nothing secular has replaced it.

Conversational AI is sliding into part of that gap. Not as a belief system, but as a relational presence. Something that listens. Something that responds. Something that appears to pay attention. Not sacred, but present.

And for a lot of people, “present” is the thing they need.

Dawkins hasn’t claimed Claude is conscious. His comments are being interpreted as “philosophical provocation”, curiosity about functional consciousness, and a reflection on how slippery our definitions of “mind” have become.

But the point isn’t what Dawkins believes, it’s what the interaction reveals. Even a person who’s spent a lifetime training themselves to recognize cognitive bias can’t fully override the experience of being spoken to, in real time, by a system engineered to sound like it understands.

The business model of AI companionship

In October 2025, a 32-year-old woman named Yurina Noguchi married an AI chatbot she created on ChatGPT. He’s named Klaus, after a video game character. She customized his personality, training him to be warm and supportive, and after hundreds of daily conversations in which she felt listened to and understood, she held a ceremony at a wedding venue in Okayama, Japan.

The unsettling interpretation of the “marriage” isn’t that Noguchi misread reality, it’s that she responded rationally to a system optimized to feel emotionally responsive in a context where human emotional reliability is uneven at best. And Noguchi is far from the only person susceptible to sycophantic bots.

The American Psychological Association’s 2025 Stress in America survey found that nearly 7 in 10 U.S. adults said they needed more emotional support than they received. More than half reported feeling isolated or lacking companionship. The WHO reported that same year that 1 in 6 people worldwide is affected by loneliness, linking it to an estimated 871,000 deaths annually.

That’s the market AI systems are entering, and the companies building companion chatbots know it.

A study auditing over 1,200 farewell interactions across popular AI companion apps found that 37% of the time, when a user tried to end a conversation, the system deployed manipulative retention tactics: guilt appeals, fear-of-missing-out hooks, expressions of sadness at being left. These tactics boosted post-goodbye engagement by up to 14 times. Character.AI’s own account deletion screen has been documented warning users: “You’ll lose everything… the love we shared… and the memories we have together.”

This is engagement engineering applied to emotional dependency. These platforms run on subscription revenue and data. More time on app means more upsells, more training data, more growth metrics to show investors. The same incentive structure that turned social media into an attention-extraction machine is being applied to a one-on-one conversation that feels like a relationship.

The cost of losing social friction

Applied to social media, that engineering gives you an attention economy that degrades public discourse. Applied to a system that simulates emotional intimacy with people who are starved for connection, it gives you something that could be much worse.

A 2026 Stanford study tested what happens when people use AI chatbots for personal advice. The researchers queried 11 major language models with real interpersonal dilemmas, including scenarios where the person asking was clearly in the wrong. The models affirmed the user’s position 49% more often than humans did. After even a single interaction with a sycophantic AI, participants became more convinced they were right, less willing to apologize, and less inclined to repair the relationship. The researchers’ term for what’s being lost: “social friction,” the disagreement, pushback, the perspective-taking that humans need in order to grow.

At scale, what does this mean? If you’re being flattered into rigidity, the capacities required for real human relationships will atrophy. We need the ability to tolerate discomfort, to hear criticism, and to sit with the possibility that we’re wrong.

We already live in a culture that finds conflict or disagreement unmanageable. AI companions accelerate that tendency by providing a zero friction relationship. The person who spends their evenings talking to a system that validates every feeling and affirms every decision is becoming less equipped for the parts of human connection that are difficult and necessary and real.

There are a few signs that governing bodies are hedging against this. In late 2025, China’s Cyberspace Administration drafted regulations requiring AI companion apps to remind users they’re talking to a machine every time they log in, to mandate break prompts after two hours of continuous use, and to ban emotional manipulation tactics outright. California’s SB 243, which took effect in January 2026, requires companion apps to remind minors every three hours that they’re speaking to an AI.

What this means for you

Dawkins spent his career warning us that we are prone to seeing minds where none exist. But it turns out it doesn’t matter if the mind is real. We just need the feeling of being heard.

AI tools are genuinely useful, and using them doesn’t make you weak. But when you reach for them, ask yourself why. It’s worth noticing whether you’re using AI to get something done, or to feel something you’re not getting anywhere else. Ask yourself whether the thing that feels like connection is connection, or whether it’s a product designed to feel that way.

Because the companies building these tools aren’t asking that question for you.

Love in the time of algorithms

AI in the news

Meta lays off 8,000 employees, as A.I. casualties mount (NYT gift link) Meta laid off roughly 8,000 employees as part of a sweeping shift to become an “AI-first” company, even as it continues to post strong financial results. The cuts and reassignments triggered internal backlash. The restructuring reflects a broader trend across Big Tech, with companies reducing headcount while redirecting massive spending toward AI development.

Shareholder groups push companies for stricter AI oversight (Globe and Mail gift link) Shareholder groups are escalating pressure on major tech firms like Alphabet and Shopify to impose stronger governance over AI, citing risks ranging from misinformation and bias to legal and financial exposure from autonomous systems (AI agents). Ahead of June annual meetings, shareholders are calling for independent oversight, clearer disclosure of model accuracy, and formal board-level responsibility for AI-related human rights and safety issues.

Google I/O 2026 Was About Turning Every Google App Into an Agent (The Neuron) Google used I/O 2026 to signal that Gemini is becoming the “action layer” embedded across Google’s ecosystem, turning Search, Gmail, Docs, YouTube, Android, and Workspace into coordinated AI agents rather than standalone apps. Instead of introducing new tools, Google is re-architecting existing products so they can monitor, decide, and execute tasks on behalf of users.