Defining 'Prolific AI Psychosis': When High AI Output Destroys Value

Psychiatrist Jeff Clark, MD continues his series on "AI psychosis" by defining the second of three categories: prolific AI psychosis. His definition: a person generates a large quantity of AI output without significantly increasing the real value of their work — and in some cases the new AI workflow actively destroys value.

The core symptom: you can't judge your own output

Clark's example is a software engineer who can produce thousands of lines of code every day where the code has little real-world utility. Contrast that with a productive developer, who may write less code but produces code valuable to users and the organization. His framing:

• Some days are large additions to the codebase.
• Other days are spent deleting unnecessary code.
• Occasionally a software-breaking bug is fixed by changing a single character.

Lines of code correlate only loosely with productivity. His author analogy: if a writer you like produces a respectable 1,000 words per day, you'd rightly worry if they suddenly wrote 100,000 words per day — the care that goes into choosing 1,000 great words can't be maintained at 100x volume.

The key point: the problem isn't increased output. More code and more text may genuinely be more productive. The problem is the subject's perception of their output. They can't assess the quality of their own work. Clark says this mimics psychosis as a mild disconnection from reality — a defect in critical thinking.

Why agent harnesses raise the stakes

Clark distinguishes chat-style usage (ChatGPT conversations, Gemini summaries in Google search) from state-of-the-art tools wrapped in an agent harness — a program designed to work on complex tasks until a solution is found, making decisions, opening programs, accessing the internet, and writing software.

His concrete example: ten years ago an iPhone app meant dozens of hours of reading and writing code. Today you type a few sentences into Claude Code, answer a few questions, and an app pops out in minutes. Harnesses can also create loops that repeatedly call models, working through a software development todo list until a full app is built.

The failure modes he lists:

• Multiple new files created when a one-line fix would have done the job.
• Confident claims that a new solution works, while the actual software is more broken than ever.
• Code becoming a tangled mess that's hard to read and almost impossible to extend.
• Decades-old problems already solved by open-source libraries rebuilt from scratch, making the product painfully slow.

Clark's summary of current models: they're "one part senior engineer and one part toddler-running-across-white-carpet-with-a-jug-of-red-Kool-Aid." Your job is to determine which is which — and he notes both sides speak with complete confidence. Mitigations exist, but developing that judgment is a skill that takes time.

The other two categories

Part 1 covered true AI psychosis (genuine psychotic experiences related to LLM use, which he argues is generally a variant of pre-existing psychosis rather than a new syndrome). A third category, parasocial AI psychosis, covers maladaptive relationships with chatbots that mimic human connection.