Humans were built to operate in groups. Pretty much everything that makes us human operates more effectively when we are surrounded by others. We have exquisitely evolved social management and response mechanisms that come from our history as herders: we conform to the opinions of others, we are sensitive to rejection, we care about our reputation and we practice reciprocity, giving to others with the hope that they’ll be there for us when needed. Most of all, we care deeply about what is fair.
All these things have no benefit if you’re a lone wolf, but they are vitally important when you’re in a herd of sheep, or the human equivalent.
Even the way we determine what is real relies on others. If you hear or see something that falls outside the ordinary, the first thing you do is turn to others to see if they also saw or heard it. We want to make sure our own senses are not tricking us. Things verified by others have a good chance of being real. Things that only we can see or hear could be a symptom of psychosis.
This need to backstop the feedback of our own senses with the corroboration of others was a useful safety check in our evolutionary past. But today, AI technology is throwing us a curveball. As a recent essay suggests, when we turn to AI chatbots for confirmation, we can start following a downward spiral into a new affliction: AI psychosis.
The challenge comes with the very sociability of new AI models – they speak to us as a friend would. Actually, that’s not quite true. They speak to us as someone who desperately wants to become our friend would.
Chats with AI are typically sycophantic and sometimes downright smarmy. They are constantly congratulating us and primping our egos. For this reason, many of us are foregoing socializing with other humans in favor of chatting with our favorite AI client. This is especially true with sensitive subjects. In a recent U.S. study from Common Sense Media, one in three teens used AI for social interaction and relationships. Of those, almost half used AI for emotional or mental health support. And often, this is a closed loop. Many teens never tell another human about their chats with an AI confidant.
This makes sense. AI won’t judge us. We aren’t risking ending our friendship with our confessions. And AI is unfailing supportive. There are no mixed signals through contradictory body language, no fear of the other person shutting down the conversation and walking away. Small wonder that one-third of the teens surveyed said that AI conversations were as satisfying or more satisfying than talking to a human. By the way, that percentage rises every year a similar study is done.
But if our goal is to corroborate reality, then we run into an issue when we turn to AI. This turns us into an echo chamber of one. In its algorithmic need to please us, AI agents can take us further and further from reality. In another study, three of five AI models happily led participants down a delusional path without any type of pushback and, in some cases, active encouragement of the delusion. One experiment saw Grok 4.1 confirming the presence of a “malevolent mirror entity” and suggested the person “drive an iron nail through the mirror while reciting Psalm 91 backwards”.
That said, this does seem to be improving. More recent models are more apt to intervene with a suggestion that the person should perhaps seek professional guidance.
But the problem may be much larger than just the encouragement of clinical psychosis. Increasingly, we humans are replacing human interaction with AI chats as a critical criterion in our efforts to make sense of reality. And that leaves a large part of our ability to accurately sense at what’s real in the hands of a few AI tech companies, all with quarterly profit targets to hit.
What could possibly go wrong in that scenario?
