Defining AI Psychosis, Part 2: Prolific AI Psychosis

Jeff Clark is a psychiatrist writing a short series on jeffs.blog about what "AI psychosis" means. This second part was published on 8 September 2026. The first part, which is not in the vault, split the term into three situations: genuine psychotic experiences related to LLM use, which he calls true AI psychosis and concludes is generally a variant of pre-existing psychosis rather than a new syndrome; hyperengagement with AI tools accompanied by a softer disconnection from reality, the subject here; and maladaptive relationships with chatbots that mimic human connection, promised for the following week.

The definition

Prolific AI psychosis is generating a large quantity of AI output without significantly increasing the real value of one's work, and sometimes destroying value. A developer in this state produces thousands of lines a day that have little real-world use, while a productive one might spend a day deleting code or fix a bug by changing one character. His analogy is a favourite author who writes a respectable 1,000 words a day and suddenly starts writing 100,000; the care cannot survive the volume. A footnote concedes that software and novels differ, with boilerplate and a limited vocabulary on the software side.

He is careful about where the problem sits. More code and more text may genuinely be more productive. The defect is in perception: the person cannot assess the quality of their own work. That is why he borrows the psychiatric word. The subject experiences a mild disconnection from reality, which he describes as a defect in critical thinking rather than a delusion.

Why agent tools make it worse

Chat interfaces are the AI most people know, and he thinks they rarely cause this. The risk comes from agent tools like Claude Code, which work through a task list in loops until an app exists. With discipline the results astound him. Without it, the tools make impossibly bad decisions with full confidence: several new files for a one-line fix, a report that a fix works while the software is more broken than before, decades-old solved problems rebuilt from scratch and slowly. His description of current models is one part senior engineer and one part toddler running across a white carpet with a jug of red Kool-Aid, both speaking with the same confidence, neither always aware of the spill.

His second image is the one that carries the argument. AI development feels like the most favourable slot machine ever built: most pulls are big wins, most losses are small and obvious, and "occasionally, a loss will look just like a win." Without the skill, focus and patience to reject counterfeit wins, the mistakes accumulate into chaos.

How it develops

He walks through a hypothetical engineer. The tool does some tasks a hundred times faster, so they pull the lever, set up more machines while the first ones spin, catch occasional bugs, set up machines to check the machines, give those more capabilities, and build networks of bots. Ambition grows toward the one-person billion-dollar company. Sleep goes, time with people goes, critics are short-sighted, and the next models will fix whatever is wrong. The question of whether any of it is useful gets deferred in the rush to "escape the permanent underclass."

He says he has been partway down that road himself. The illusion broke when he realized he could not understand his own project: dozens of custom files and not the faintest idea how they worked, and no way to add features without starting over. The developers he has spoken to who came out the other side all prioritize human judgement, sleep, and some life outside work.

Causes, offered as speculation

He labels this part as speculation and keeps it hedged. Intermittent reinforcement, rewards arriving on an unpredictable schedule, is one of the strongest drivers of behaviour and is a property of LLM tools that cannot be wholly removed. He suspects people with ADHD or impulse-control problems are at higher risk, and possibly people high in openness and low in conscientiousness. He does not know whether juniors or seniors are more susceptible; he suspects juniors, but accounts of prominent developers apparently affected are common. Hype plays a part, as does the narrative that AI will replace every job, which is sharp for developers who struggle to find work after layoffs and are measured in environments where more output still earns rewards. Finally there is taste, which he prefers to call craft: judging whether a product is useful, good, desirable or delightful is a conversation between humans, and hard to find inside an LLM.

The footnotes do useful limiting work. Anecdotes are not data. An accusation of AI psychosis is not always merited, and "this doesn't meet my personal standard of quality" is a criticism older than generative AI. Beliefs shared within a community, such as a permanent underclass, superintelligence or an AI bubble, are not delusions. And he is sceptical of hundredfold productivity claims: a few minutes with a model can save a day, but the average gain is far smaller once supervision and the other bottlenecks (coordination, understanding the customer, access to resources, internal policy) are counted.

Where it fits

ceo-ai-psychosis used the word first for executives, with Garry Tan's sleepless agent orchestration and Karpathy's self-described "state of psychosis"; Clark supplies the clinical boundary that piece lacked, separating a self-assessment defect from actual psychosis. specsmaxxing is a first-person case, down to the stage Clark describes of building machines to check the machines. The slot-machine mechanism appears in agentic-coding-fatigue as a gacha loop and in building-syntaqlite-ai as an addiction loop, and ai-superpowers-focus-followthrough is the ambition stage in practice, with 40 side projects running at once.

The core claim, that people cannot judge their own AI-assisted output, is what fooling-yourself-with-ai argues from METR's gap between felt and measured speed, and the ai-sycophancy-loop is one reason the judgement fails. tokenmaxxing is the metrics-driven reward he names among the causes. His moment of not understanding his own project is the lost situational picture described at ironies-of-automation, and back-to-coding-by-hand is someone who hit the same moment and walked back.