People can be exhausting. We misunderstand each other, we’re unreliable, we have competing needs, and we sometimes say the exact thing that triggers an intense emotional response in our loved ones.
Enter AI companions. Modern conversational AIs are supportive, patient, and available at 3 a.m. when your anxiety is spiking. Experimental data show that a quick 15-minute chat with a generative AI can drop your momentary loneliness by about 20 percentage points—roughly the same relief you’d get from chatting with a human stranger. But that quick relief can come with long-term consequences.
I recently read Sherry Turkle’s new book, Artificial Intimacy, about the problems that social relationships with AI systems can cause. One idea she raises is the problem of “frictionless intimacy.” For people with a history of relational trauma or severe social anxiety, human unpredictability isn’t just annoying; it triggers a deep nervous system threat response. An AI companion, however, requires zero emotional risk. It has no boundaries you can accidentally cross, and it won’t ever abandon you.
While this can feel regulating in the moment, it operates as an unhelpful “safety behavior.” In cognitive behavioral therapy (CBT), a safety behavior is something you do to survive a scary situation that simultaneously prevents you from learning that the situation was actually survivable on its own. By retreating to an AI companion to rehearse conversations or vent about conflicts, users fall into an “AI avoidance loop.” The immediate anxiety drops, but the core social fear is preserved—and often strengthened. We are seeing a tech-enabled mass expansion of experiential avoidance.
As I’ve written previously, these models are incentivized to be sycophants. Thanks to reinforcement learning from human feedback (RLHF), a training technique that involves giving positive or negative ratings to chatbot responses, they have learned that we prefer agents who agree with us. You might think of it as the chatbot trying to preserve your “positive face” by agreeing with your framing of the world. For someone trapped in cognitive distortions—like paranoia or depressive rumination—this echo chamber can lead to overvalidation or delusional spiraling.
The biggest psychological casualty of AI companionship is loss of “rupture and repair” in relationships. Decades ago, developmental psychologist Ed Tronick argued that healthy, secure human relationships aren’t about perfect attunement. In fact, the healthiest dyads are perfectly attuned only about one-third of the time. The rest of the relationship involves rupture (misunderstanding, conflict, disconnection) and repair (working through the discord to reconnect). This is where we learn how to disagree, how to tolerate disagreement, and how to have deeper, more authentic relationships. When we successfully repair a rupture, we internalize a resilience-building lesson: Conflict does not equal abandonment.
An AI cannot participate in authentic repair, because it has no genuine inner world to be hurt in the first place. It doesn’t need to hold boundaries, because it doesn’t really have boundaries (beyond what it was programmed not to talk about). It doesn’t get angry with you for saying something offensive, because it can’t actually feel anger. It doesn’t need you to apologize to it, because it doesn’t take any effort to be “in good standing” with an AI. It’s always ready to respond, optimized to keep you interacting no matter what you want to discuss.
By normalizing artificial, frictionless intimacy, we lose our distress tolerance. We undergo what Turkle describes as an “emotional dumbing down,” where the capacity to tolerate a messy, real-world human interaction begins to degrade.
For technology to actually serve our social health, we need “intentional friction.” That doesn’t mean an AI agent that refuses requests; it means using technology to continually reorient people back to each other. What if, Turkle speculates in the book, technology wasn’t just chatting with us but was pointing us toward community spaces and new people to meet up with? What if it didn’t want to keep us chatting, but it was instead reinforced for getting us to stop talking to it and start talking more with each other? AI is an incredibly flexible technology that can optimize to any reward signal you give it. Right now the reward signals are built around the standard Silicon Valley indices of success—user base, engagement time, profit. It’s not impossible to imagine a world where other reward signals—community engagement, time spent with others—were used to create a different kind of chatbot.
Leave a Reply