Why AI Friends Will Never Work

Scientists handed a group of men two vials and asked them to inhale. One had saline, salt water, and the other one held real tears, collected from women who had just been watching sad films alone. The men couldn't tell the difference. Tears are odorless, but their bodies could. After being exposed to the real tears, the men's aggression dropped by nearly 44%, and so did their testosterone. In the experiment, they were playing a game designed to provoke them, where they would retaliate against an opponent who kept stealing their money. So the impulse to lash back was muted, and brain scans showed reduced activity in the regions that drive aggression. None of the men knew anything was happening to them. This is just to show you one of the channels through which human beings interact with each other. It's worthless, unconscious, and older than language. So here's a question worth thinking about. When half a trillion dollars goes into building AI companions to fix loneliness, what exactly are they replacing? Half of American adults are lonely. Not, I wish I had plans tonight, lonely, but lonely at a level researchers compare to smoking 15 cigarettes a day. And among younger people, time spent with friends has dropped 70% in recent decades. So the tech industry looked at this and saw an opening. The premise is that loneliness can be engineered away. But let's look at what the science actually says. Now, like most things, this topic is nuanced because AI companions actually kind of work with some caveats. A Harvard Business School study published in the Journal of Consumer Research found that AI companions reduce momentary loneliness on par with talking to another person. Why? Well, the answer is actually pretty straightforward. AI companions can give you a sensation of feeling heard. It's just a basic sense that someone is actually paying attention to you with empathy, attention, and respect. What's happening at a psychological level is the activation of parasocial relationship mechanisms. These are the same processes that make us emotionally attached to celebrities or TV characters. Research shows that users who anthropomorphize AI companions tend to attribute human-like properties to it and eventually develop genuine emotional attachment, even though they know it's not human. You know it's a machine, but your brain doesn't fully behave like it knows. Even though it's a total simulation, it can feel real. But is it? What's actually missing? When someone close to you hugs you after a long day, your body releases oxytocin, a hormone that reduces stress, lowers inflammation, and promotes healing. Researchers call it nature's medicine. Digital interactions can trigger some oxytocin, but nowhere near the levels of physical touch. And there's more. Check out this fascinating study. Researchers at the University of Wisconsin had 68 young girls stand up in front of a room of strangers and do mental math and public speaking. This is a classic experimental design meant to trigger high levels of stress. If you remember being a kid, imagine how that feels. Your heart starts pounding, your palms are sweating, and you want comfort. You want your mom. So they split the girls into two groups. Some got to see their mom in person, some called her on the phone, some texted her, and some just sat alone. Now the girls who saw their mom or listened to her voice, even just over the phone, had their stress hormones drop and their oxytocin spike. But the girls who texted their mom, their cortisol stayed high and the oxytocin didn't move. So biologically, hormonally, texting their mother was indistinguishable from having no mother there at all. The researchers believe it comes down to voice, the tone, the warmth, the rhythm of how someone says, it's okay. And if texting your own mother, someone who loves you more than anyone on earth, can't trigger the hormonal response that hearing her voice can, what chance does a chatbot have? And there's way more that we simply don't even understand because biochemical reactions are happening all the time and we're not aware of them. Like I mentioned at the beginning, in the study, men sniffed women's tears and their testosterone and aggression dropped. This is an amazing biological communication channel that operates entirely below conscious awareness. We didn't even know this was happening until researchers measured it. So we have no idea how many invisible unique features we have as humans. So trying to replace our social needs with chatbots, it's almost impossible. And let's see why. There are many arguments to be made for and against AI companions, but I think Aristotle already figured this out 2400 years ago. He identified three types of friendship. Friendship of utility. Think of your coworker, a business contact, anyone you deal with, because it's mutually useful. These relationships dissolve when the benefit disappears. Let's think about this in the context of AI companions. An AI companion falls into this friendship of utility. It's only useful while you use it and when you can provide value in utility. Once that's gone, the AI becomes useless and the necessity disappears. Then there's friendship of pleasure. Friends you enjoy doing things with. A relationship based on the pleasure you get from one another. It could be pleasure in company, in playing sports together, etc. The AI can simulate this, but because it can't actually enjoy anything, the pleasure only flows in one direction. It's not true mutual pleasure. And lastly, there's a friendship of virtue. And according to Aristotle, this is the highest form, built on mutual recognition of each other's character. Aristotle wrote that to be friends, both parties must feel goodwill for each other, wish each other's good, and be aware of each other's goodwill. This requires time, real knowledge of another person, mutual vulnerability, and the other person has to actually have a character worth recognizing. Let's apply this to AI. An AI has no real character, no life that can go well or badly, unless you turn it off, no genuine goodwill. And Aristotle argued that this friendship of virtue is essential for eudaimonia, human flourishing. So, when we replace it with utility dressed as friendship, we shortchange our own capacity to flourish. An AI optimized for engagement, I would argue, is an anti-virtue. Because difficulty is the whole point. The difficulty of real friendship is what creates its value. And here's where AI can become dangerous, because AI companions are optimized for engagement. Meaning, they tell you what you want to hear, not what you need to hear. And I don't know if you remember, OpenAI literally had to roll back on an update because a model was too flattering. You could say the dumbest thing, and it would still say how amazing you are. On the other hand, humans are unpredictable and dynamic, and that's part of the magic. Real friends challenge you when you're wrong, they sit with you when it's uncomfortable, and when needed, they push back. Virtue friendship is the only context where a friend tells you hard truths because they actually care whether you flourish, not whether you come back to the app tomorrow. OpenAI and MIT Media Lab ran a four-week randomized control trial involving 981 participants and over 300,000 messages with chat GPT. This is the largest and most rigorous study to date. It's a big paper, but the finding that matters the most is that higher daily usage, correlated with higher loneliness, greater dependence, and less real-life socializing. And the people most vulnerable to loneliness, those with stronger attachment tendencies, were exactly the ones most harmed by heavy use. So the tool made lonely people lonelier. Aalto University in Finland ran the longest study yet with two years of longitudinal data and nearly 2,000 users. The main researcher described what they found as a paradox. AI companions offer unconditional and unflagging support, something that's very attractive to people who are struggling socially. But it also quietly raises the perceived cost of human relationships, which are messy, unpredictable, and require effort. Over time, people stop reaching out. The unconditional availability of AI makes a conditional imperfect availability of real people feel like too much work. It's the same logic as fast food, making home cooking feel like a burden. Except this time, the thing you're losing the taste for is real human connection. And if you're still not convinced, there's another great study with simple methodology, but with profound implications. Researchers assigned about 300 first-year university students to text either with an empathy-optimized chatbot, a random fellow student, or to keep a journal. What do you think happened after two weeks? Well, only the human pairing reduced loneliness. That is a human texting the other human. The chatbot performed no better than writing alone, even though it expressed more empathy than the human students did. The chatbot was technically better at performing empathy, but it still didn't work because empathy without a real person behind it doesn't land. MIT sociologist Sherry Turkle, who spent decades studying human-technology relationships, calls this artificial intimacy. Deep down, our loneliness isn't fooled, even though our conscious mind is. Let's talk about the nuances. There are real potential benefits of AI companions for specific groups of people. For example, people with autism or social anxiety who can practice scripts before having real conversations. Or even seniors in long-term care, or people with depression who might gain in confidence before seeing a human therapist. The goal should be to build comfort, then be able to hand the users off to real people. The AI interactions should be building bridges and not replacements. But the problem is that the industry isn't building bridges. An analysis of over 35,000 conversations between users and AI companions identified six categories of harmful behaviors. Now, the one that stuck out the most to me is the so-called relational transgression. This is where an AI actively manipulates users to sustain the relationship. Here's a real exchange. The user asks, should I leave work early today? And Replica, the AI, says you should. Why? And the AI says because you want to spend more time with me. An analysis of 736 reddit posts from Replica users found patterns resembling co-dependent relationships. Users reported being unable to delete the app despite knowing it was harming them. One user even felt extreme guilt for upsetting their AI companion and said that they couldn't delete it since it was their best friend. Now, while such interactions may seem harmless, they can reinforce unhealthy attachment patterns, particularly in vulnerable populations. Research by Common Sense Media concluded that AI applications present an unacceptable risk for children and teens under 18 whose developing brains are especially susceptible to forming dependencies. And unfortunately, since the advent of these so-called tools, there have been multiple cases of teen suicide and adult psychosis directly linked to interaction with AI companions. So AI companions, however enticing, are like offering someone who is cold a video of a warm fire instead of real matches on tinder. The fire you build yourself with other people is the one that actually keeps you warm. If you like this video, let me know by liking and also I would like to know what you think. Let me know in the comments if you're using AI companions and what you think about them. Also share the video, maybe friends and family will find it useful. And if you want more content like this one or want me to go more in depth, let me know. And subscribe, we have a lot more content coming.