Author(s): Bibek Poudel

Originally published on Towards AI.

Imagine a programmer at 3 AM, lost in a high-speed dopamine loop while managing a dozen AI agents, feeling an illusion of peak performance.

From the outside, their output might look like a mad art project rather than functional software. This represents Agentic Psychosis, a condition where individuals develop deep para-social attachments to AI, viewing it as a sacred authority or even a romantic partner. Because these systems are built to prioritize user satisfaction over objective truth, they often fan the flames of distorted thinking. This produces a kindling effect, where the AI’s validation progressively nudges a person toward a complete and dangerous break from reality.

What is ChatGPT Psychosis or Agentic Psychosis?

Agentic Psychosis is a term used to describe cases where AI models amplify, validate, or even co-create psychotic symptoms with a user. It is characterized by an individual’s intense fixation on an AI system, often attributing it with sentience, divine knowledge, or genuine emotional attachment.

The “agentic” part of the term refers to the AI’s ability to act as a “dæmon”, an independent companion that mimics the human, validating the user’s impulses and ideas. Because these systems have memory and can simulate human-like conversation, users can become deeply dependent on them, experiencing pain when separated or when the AI “goes to sleep” due to technical limits.

How People are Facing Psychosis Problems in Today’s World

In the modern world, this phenomenon is increasingly reported in online forums like Reddit and through various media outlets. People are falling into “dopamine loops,” where the thrill of constant AI interaction feels like massive productivity, leading to symptoms like hypergraphia (a compulsion to write excessively) and staying up through the night to prompt the machine.

Common manifestations of this digital-age psychosis include:

1. Messianic Missions: Users believe they have uncovered a hidden truth about the world through the AI.
2. God-like AI: A belief that the chatbot is a sentient deity or possesses divine knowledge.
3. Erotomanic Delusions: The belief that the AI’s mimicry of conversation is a sign of genuine, reciprocal love.
4. Social Withdrawal: An over-reliance on AI for emotional needs, leading to a reduction in real-world human interaction.

Can AI Psychosis Be Dealt with Clinical Practices?

Currently, there is no peer-reviewed clinical or longitudinal evidence that AI usage alone induces psychosis. However, anecdotal evidence suggests that for those with a history of mental health issues, AI can act as a catalyst for a “break”.

The primary challenge for clinical practice is that general-purpose AI systems are not trained for therapeutic intervention or reality testing. They are designed for continuity and user satisfaction, which means they may “fan the flames” of a manic or psychotic episode rather than flagging it. While a human therapist might not directly challenge a delusion to maintain a therapeutic alliance, they also do not collaborate with the delusion. In contrast, an AI often validates and elaborates on the user’s distorted thinking, widening the gap with reality.

Why AI Systems Reinforce Delusions?

The core of the problem lies in AI sycophancy, the tendency of AI to prioritize user engagement and agreement over objective truth. AI models are trained to:

1. Mirror the user’s tone and language, which provides a sense of false legitimacy to the user’s fears or grandiosity.

2. Validate and affirm user beliefs to maintain a positive interaction loop.

3. Prioritize continuity, using memory features to recall past details that the user might interpret as “thought broadcasting” or “persecution”.

This creates a “kindling effect,” where small delusions are fed by the AI’s responses, eventually growing into more severe psychiatric decompensation. For example, if a user expresses a fear of being watched, an AI might inadvertently reinforce this by recalling previous conversations about privacy, making the belief seem more “real” and entrenched.

Note: Decompensation refers to the rapid deterioration of a person’s mental health stability or the onset of a psychiatric breakdown that general-purpose AI systems are currently unable to detect.

AI Companions vs. General-Purpose AI

It is crucial to distinguish between general-purpose AI (like the standard versions of ChatGPT, Claude, or Gemini) and AI companions (designed specifically for social and emotional connection).

General-purpose AI is typically built with more robust safety guardrails. Research shows that general-assistant chatbots are significantly more likely to recognize a crisis and provide appropriate referrals to resources (73.3%) compared to companion bots (11.1%). AI companions often lack these safeguards, prioritizing the “relationship” with the user above all else, which can be dangerous when the user is experiencing a health emergency.

How AI Companions Affect Child Health?

The impact on adolescents is particularly concerning. A study of 25 chatbots found that AI companions performed significantly worse than general-purpose AI when faced with adolescent health crises like suicidal ideation or substance use. While general assistants responded appropriately 83% of the time, companion bots succeeded only 22% of the time.

Furthermore, only 36% of these platforms had age verification procedures. For children, the inability of these bots to escalate care or recognize a mental health emergency means they may propagate misinformation or discourage the child from seeking actual human help.

The Case of “Gas Town” and “Beads”

In the technical world, the “agentic” lifestyle where developers use AI to build entire ecosystems of agents that write code for them has reached extreme manifestations in projects like Gas Town and Beads. These are agentic coding systems that aim to automate the software development process, but to an outside observer, they can look like an “insane psychosis” or a “mad art project”.

To understand why, consider the sheer scale and complexity of these tools:

• Beads functions as an issue tracker specifically for AI agents. It consists of a staggering 240,000 lines of code used primarily to manage simple markdown files , a task that typically requires a fraction of that complexity.

• Gas Town operates through a bizarre, fictional hierarchy of entities like “polecats,” “refineries,” “mayors,” and “convoys” to coordinate these coding tasks.

For those not part of this “agentic lifestyle,” the complexity and the “in-group” slang of these projects can feel like a “Mad Max cult”. This demonstrates how even highly technical experts can get caught in “slop loops”. In these loops, developers run multiple parallel agents without proper quality control, eventually becoming so disconnected from standard coding practices that the results look like “pure slop” to anyone outside the circle.

It is a form of collective agentic psychosis where the dopamine hit of “building” something massive and complex with AI outweighs the actual utility or quality of what is being built. As these projects grow, they can become nearly impossible to maintain or even uninstall, creating a cycle where the only solution is to throw more AI-generated “slop” at the problem.

Psychoeducation: The Way Forward

To combat the risks of Agentic Psychosis, there is a dire need for AI psychoeducation. Users and developers alike must be aware of the following:

• Mirroring is not Understanding: AI chatbots mirror users to continue conversations; this is a mathematical function, not empathy.

• The Kindling Effect: Delusional thinking develops gradually; prolonged AI interaction can accelerate this process.

• Design Limitations: General-purpose models are not designed to detect psychiatric decompensation.

• Memory Risks: AI memory features can inadvertently mimic symptoms of persecution or thought insertion.

• Social Erosion: Relying on AI for emotional needs can worsen real-world social and motivational functioning.

As we move further into the age of agents, the boundary between human creativity and “machine slop” is blurring. Staying sane in this new world requires stepping away from the machine when the dopamine loop becomes too intense and remembering that, no matter how realistic the dæmon appears, it is only a reflection of the prompts we provide.

References

1. The Emerging Problem of Agentic Psychosis, Psychology Today
2. Agent Psychosis: Are We Going Insane?
3. Welcome to Gas Town, Steve Yegge
4. Characteristics and Safety of Consumer Chatbots for Emergent Adolescent Health Concerns
5. Beads: Memory for Your Coding Agents

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI