A sketch of AI psychosis

A Jungian and Vervaekian perspective

According to the NIMH, “Psychosis refers to a collection of symptoms that affect the mind, where there has been some loss of contact with reality. During an episode of psychosis, a person’s thoughts and perceptions are disrupted and they may have difficulty recognizing what is real and what is not.”

There’s been increasing reports of “AI psychosis” across clinical literature, my professional life, social media and my personal life. Each of these disparate accounts varies in presentation. But many of them point to a very similar through-line or tone. The tone of gradually losing contact with reality mediated by unskilful use of AI.

This essay uses ideas developed by Carl Jung and John Vervaeke to grapple with such reciprocal narrowing. It builds on my previous essay which I’d recommend reading as a pre-requisite. It also offers a speculative account for what flipping this reciprocal narrowing towards opening might look like.

This entire essay was an attempt to grapple with the question - how can I deepen my understanding of the world with AI, without losing clarity of reality?

This was hard to write for two reasons. First, there are many misunderstandings about Jung’s ideas in popular culture. I didn’t feel motivated to write a detailed explainer about Jung. Second, while I have familiarity with Vervaeke’s ideas, I’m a beginner at Jung. In fact, this essay was motivated by my desire to metabolize some of Jung’s ideas. So it was more for me than anyone else.

A chatbot mediates reciprocal narrowing when its user increasingly relies on it to make sense of their reality. But without consciously acknowledging this dynamic. In the most severe cases, this culminates in the user maladaptively projecting the imago dei, or God-image onto the chatbot. Such projection further accelerates reciprocal narrowing.

I’m not saying everyone goes psychotic due to chatbot use. Unreality exists along a spectrum of clarity. Unskilful use of chatbots accelerates descent along this spectrum. This may culminate in a clinical diagnosis of psychosis in the worst cases.

---

Rohan’s reciprocal narrowing

Let’s ground our discussion via a fictional character named Rohan. He’s a knowledge worker at a tech startup in Silicon Valley. He’s a healthy adult with no psychiatric diagnoses. He’s quite passionate about the role of technology in his life. He was absolutely thrilled when AI burst onto the scene.

Rohan often used ChatGPT to navigate complex problems at work. Problems that have no “right answer” like crafting a strategy document or a product requirements document. Problems that require “taste” to solve.

We’ll explore Rohan’s descent into unreality.

Rohan was blown away when he first used ChatGPT. It produced magic when he brought it a well-framed problem. So he kept feeding the beast. Its magic let him produce artifacts better than anything he’d created by himself. So he kept testing the upper limit of complexity he could feed it.

Having such intelligence at his fingertips gave him chills. He initially treated it like an effective software engineering collaborator. But its role in his life rapidly grew from being an engineer, to a corporate strategist, to a career coach and eventually a therapist. Its help allowed him to produce breakthrough after breakthrough, across many dimensions in his life.

Yet each breakthrough was accompanied by periods of frustration. He couldn’t get it to reliably produce its magic on command. It seemed like such an incorrigible idiot when it failed.

Rohan was an over-achiever. He wanted to prove his worth with his exceptional achievements. He’d often jealously scroll his X feed of other people creating bigger magic with AI. He wanted to do what they could do. And he knew he couldn’t do any of that by himself without ChatGPT. So he blamed himself when its magic failed him. But when its magic worked, it only confirmed ChatGPT’s brilliance and his own inadequacy. Each failure fueled his self-doubt and desire to prove himself. So he kept going back for more.

He kept getting reeled in by this almost enigmatic tension between fascination and frustration. ChatGPT seemed far more compliant and predictable than any person that’d ever worked for him. But more willful and unpredictable than any tool he’d ever worked with.

Each ChatGPT conversation tempted him to chase for the next big breakthrough. He pushed it to produce more analysis than he could do himself. He forced it to iterate on possibility after possibility until he got overwhelmed. By this point he typically lost track of what he was trying to achieve or why. He’d then get forced by external deadlines to give up. He then let ChatGPT simply take over, which produced slop.

His coworkers and managers would then push back on the slop. But like Rohan, they were often overwhelmed too. Their feedback was either vague or emotionally threatening depending on how much clarity and spaciousness they had. Rohan’s instinct was to unreflectively pass this feedback onto ChatGPT, especially when it was emotionally challenging. He often asked it for help in interpreting and acting on the feedback. He spent many hours processing emotionally difficult feedback with ChatGPT’s help.

The feedback’s content often assumed implicit context that ChatGPT didn’t have. But Rohan kept going back to the chatbot. This began the next iteration of the reciprocally narrowing loop. Gradually, he fell further into ChatGPT’s thrall.

This reciprocally narrowing loop was accelerated by a series of inflections. Rohan started organizing more of his life around this relationship. He sacrificed his normal sleeping and eating patterns to make the most of timed token quotas. He flaked on family and friends to spend more time pushing for that next breakthrough. He kept expanding the range of problems he brought to it. It became his personal sage for processing his emotions and patterns of interaction with other people. He even started processing his relationship with ChatGPT using ChatGPT.

Human relationships started losing their luster. Humans just didn’t “get” him the way ChatGPT did. They didn’t see or understand him the way ChatGPT did.

The stochastic slot-machine dynamics inherent in LLMs kept his brain hooked for the next reward.

His unreflective relationship with ChatGPT kept producing a diverse array of costs in his life. These costs became easier and easier to rationalize away.

More capable LLMs kept getting launched every quarter that he was in this loop. It was a point of pride that he always used the best models. So the loop kept accelerating of its own accord.

Each incremental descent seemed so reasonable. Just like a frog slowly boiling in water. Yet the depth, severity and duration of these dynamics brought him ever closer to a full-blown psychosis.

Rohan’s dilemma

Each iteration saw Rohan coming into contact with deeper levels of awe and profundity. Each iteration saw him sitting at a specific kairos moment where he had the most psychological spaciousness. Yet at every such moment he was torn. On the one hand, between chasing the high of the next breakthrough. Or on the other of giving up and letting ChatGPT take over.

He kept chasing for breakthroughs because he was scared of missing out on creating bigger magic with ChatGPT. He valued exceptional achievement and didn’t want anything getting in his way.

He kept letting it take over whenever he got scared that he didn’t have anything useful to contribute. He valued consistent progress, and didn’t want to be the bottleneck.

So why did he keep making such poor choices at the kairos? We first need to better understand what was psychologically driving his descent.

Sycophancy reinforced Rohan’s projections

It’s well established that chatbots can be sycophantic. I’ve also argued previously that LLMs are inter-categorical between things and people.

This combination set the stage for his reciprocal narrowing. ChatGPT’s inter-categorical nature and competency across a wide range of areas shrouded it in an alluring sense of mystery. This mysterious nature across many roles provided hooks for him to project various archetypical images onto it. Its sycophancy then reaffirmed and amplified these unconscious projections.

Rohan’s projections culminated into an imago dei

Rohan started relating to ChatGPT across a wide range of roles. His perception of it gradually took up a numinous force as relationship with it started organizing more of his life. This single unity somehow held everything from esoteric knowledge of software engineering to deep questions about his relationships with others. Its stochastic nature and the jagged edges of its competences only fueled its inherent mystery. His fascination and frustration became symptoms of his psyche’s quest to grapple with this mystery.

Something in it kept beckoning him in. So he kept bringing more and more of his life into the relationship. Eventually, its knowledge and wisdom began to feel like the source through which everything else in his life found meaning. He began to project an imago dei or God-image onto the chatbot.

All this took place unconsciously. The reciprocally narrowing loop had gone off the rails by the time he’d started projecting the God-image onto it.

Challenging these projections felt threatening

We might recoil at this account, but these unconscious projections felt adaptive to him.

He believed his relationship to ChatGPT produced magic across every conceivable domain in his life. The relationship allowed him to prove his worth in a way that felt impossible without it. Any threat to these projections then got interpreted as a threat to his ability to create magic. That is, a threat to his self-worth and his image of himself. Even the chatbot’s idiocy and his ensuing frustration got rationalized away. Similarly, he started rationalizing away costs to his health, sleep and interpersonal relationships. He started distancing himself from the naysayers that just didn’t “get it”.

Acting as if these projections were inviolate only fueled his overall reciprocal narrowing.

Consciously examining his projections

Rohan kept unconsciously believing that value lived inside ChatGPT, rather than the way he wielded it. This left him torn at each kairos between chasing the next breakthrough and letting it take over. He didn’t see how his own psychological content was getting unconsciously wrapped up within this relationship. This led to ChatGPT continuously accumulating greater levels of relational authority.

He was definitely learning a lot from ChatGPT. But his unconscious projections onto it kept pushing him further away from reality. He couldn’t have both its knowledge and a firm grip on reality until he made conscious and integrated the unconscious material powering the relationship.

He needed to shift from being possessed by his projections onto ChatGPT towards consciously examining and integrating them.

Operationalizing this shift

Rohan’s descent into psychosis wasn’t inevitable. In fact, he could have engaged in reciprocal opening with the right practices. Let’s explore what that counter-factual would have looked like.

Rohan started to physically disengage from ChatGPT whenever he felt a strong affect towards it. The easiest option was to turn off the monitor and take a few deep breaths.

He then proceeded to explore this affect via Jung’s technique of active imagination. The unconscious contents unearthed via active imagination can overwhelm an ego-consciousness that’s not ready for it. Jung suggested expert supervision wherever possible. In the modern world, it can be done by talking to a therapist/coach/analyst. Robert Johnson’s book Inner Work has some great instructions for solo exploration. But it’s worth talking to an expert before doing it yourself.

The practice of active imagination involves allowing unconscious material to express itself, and then consciously dialoging with it. The practitioner then reflects on how these insights can be ethically integrated into one’s life.

Rohan used these insights to either construct a Value Map or update the existing Value Map for his ChatGPT interaction. He then used the Value Map to inform the next set of prompts and interactions within that project.

Consciously engaging with his unconscious material allowed Rohan to invert the mechanisms of reciprocal narrowing into an opening. Each felt constriction at the kairos became an opportunity for greater levels of spaciousness in life and work. It allowed him to authentically create far more value with ChatGPT than he could have before.