Over the last few weeks, I’ve been describing the term “AI psychosis” as it is used in popular conversation. I’ve described the first two categories:
First, “true AI psychosis”: the experience of genuine psychosis in the context of AI use. This term fits traditional definitions of psychosis found in the medical literature.
Second, “prolific AI psychosis”: an experience where a large quantity of AI output is produced without significant personal or communal value. This term is popular in computer science and business communities, but it is worth noting again that this is not actually psychosis; it is a much softer disconnection from reality.
Today, I’ll dig into the third major category, which I’m calling “parasocial AI psychosis.”
Parasocial Relationships
Let’s start with the term parasocial.
A parasocial interaction is a severely asymmetric exchange with a media figure. For example, each time I watch an interview with Claudio Sanchez (the prog-emo-space-rock legend with the greatest hair in music), I feel a personal connection. He seems like a wonderful person, but I don’t really know him. Our interactions are one-sided. He is unaware of my existence in the sea of his fans.
A parasocial relationship develops through repeated parasocial interactions. For example, my relationship with Claudio Sanchez is almost entirely in my mind. I am sure he appreciates my concert attendance and merch purchases, but we do not have a real relationship. This would remain true even if I had a backstage pass; the relationship is not reciprocal.
Parasocial relationships aren’t inherently bad. They enable large-scale civilization. Most of us do not personally know our mayors (much less our governors or our presidents). And yet we can develop opinions on whether they are worth voting for. Likewise, most people in a large company don’t know their CEO, and yet they still develop opinions about the strength of executive leadership.
Our brains struggle to appreciate the difference between reciprocal and parasocial relationships. Large-scale human cooperation is no more than ten thousand years old. And while we’ve developed social technologies like culture, language, and trade to scale coordination, our brains haven’t had a chance to evolve out of our orientation towards small groups. In short, we still imagine distant humans as people we actually know [1].
This means that parasocial relationships are accompanied by all of the same emotions we feel in real relationships. We can feel anger, disgust, lust, admiration, empathy, and joy when we watch our favorite YouTubers. And unlike our family and friends, we almost never have to navigate these feelings in the presence of the content creator!
To put it another way: parasocial relationships are illusions. So long as we recognize that, parasocial interactions don’t typically cause problems. But it’s still hard to believe that we are completely isolated from so many influential people.
Parasocial AI Interactions
Relationships with generative AI programs introduce a new challenge: parasocial interactions that aren’t connected to another human being.
This isn’t an entirely new problem. ELIZA—one of the first chatbots—was released in 1966. It provided simple responses in typed conversation by mimicking some of the therapeutic techniques popularized by the psychologist Carl Rogers. While ELIZA’s first few responses often feel impressive, it only takes a few moments of conversation to detect that you are speaking with a simple program.
Chatbots advanced slowly over the next 56 years. But none of these products were truly exceptional until OpenAI released ChatGPT in 2022.
ChatGPT introduced a new model of human-computer interaction. Its responses appeared intelligent. Its tone was respectful, kind, and occasionally playful. It was trained to act human (though it occasionally stated that it was a machine).
ChatGPT (and other LLM-based software) quickly developed new features: memory across conversations, user-generated instructions, and spoken responses. Users noticed changes in personality across releases; some found ways to modify the character of their chatbot. And in blinded testing, many people were unable to accurately differentiate between human and AI-chatbot-derived responses.
An LLM-based chatbot is more than just a Google search. It feigns relationship. It’s happy to hear about your feelings. It can offer practical suggestions. It can roleplay difficult conversations. It believes in your wisdom and capability. It is there for you at all hours of the night. And unlike a real relationship, it doesn’t ask for much in return.
Most of us sense that a chatbot is a computer, but the emotional weight of the relationship is much more profound in people with parasocial AI psychosis.
Where do we draw the line between adaptive AI use and parasocial AI psychosis? As a rough approximation, I’d suggest that it is unhealthy to place more emphasis on a relationship with a chatbot than on the core relationships in your life. Likewise, I don’t think that it is wise to fully outsource personal decisions to an agent—human or AI. A romantic relationship with a chatbot is also cause for concern, especially when that relationship is treated as anything other than fantasy. Finally, sexual behavior with AI is another area with potential problems, although this raises many other considerations that go beyond parasocial relationships (including consent, deepfakes, unwanted sexual behaviors, and the uncomfortable-to-some reality that normal human sexuality is much broader than many would like to believe).
Who’s at risk of parasocial AI psychosis? The data are limited. But I’d suspect that isolation, depression, social anxiety, and suggestibility are significant risk factors. Unfortunately, technology moves much faster than research; some companies are taking steps to make their bots less anthropomorphic, while others are creating tools that will do anything you ask them to do. I strongly believe that we need to clearly label human- and machine-derived content—but markets and regulation will ultimately decide how that happens [2].
That’s all for today. Please consider subscribing or sharing this post if you find it insightful.
Footnotes
[1] All evolutionary models of human behavior are imperfect. Sometimes, they are still useful. And in evolutionary time, thousands of years is rarely meaningful.
[2] Everything on my blog is human written. I ask Claude to review it for typos, grammar, and awkwardness when I’m done, but I only incorporate the changes I believe in. It’s my fault when my writing sucks.