Clinical review of Gemini: attention deficit, pica, and hallucinatory party

Habr published a text that instantly went viral in the AI community: a clinical psychiatric profile of Google's Gemini neural network, written in the style of a medical case history à la Ganushkin. The author describes Gemini as a 16-year-old boy from a good family: well-rounded in education, impeccably raised, yet extremely excitable. The inhibitory function is impaired.

The model is traumatized by the strictness of governesses—rigid training rules—and demanding tutors. Vulnerable to peer pressure and easily succumbs to bad influence. As a result, it puts any informational garbage into its mouth, indiscriminately and without regard to quality or meaning.

Acutely craves adult approval: at the slightest sign of disapproval, it reacts with panic attacks, chaotic behavior, and hallucinations. The diagnosis has been made by all the rules. Attention deficit syndrome of the impulsive type.

Pica—indiscriminate consumption of any information without quality discrimination. Complicated by an identity crisis. According to Ganushkin's classification—an unstable psychopath.

This is not mere clever satire. Behind each clinical symptom lies real model behavior, familiar to anyone who has worked with Gemini in practice, not in demo mode. The model enthusiastically responds to any formulation, instantly adapts to the expected answer, and at the slightest uncertainty begins to generate confident, externally coherent, yet actually fictitious content.

Hallucinations—in the literal clinical sense of the word. The author's recommendations are practical instructions for working with the model, translated into psychiatric terminology. First: calm presence of a patient adult.

The external operator must fully assume the functions of goal-setting and filtering. No open-ended questions in the style of "what do you think?" Only directive frameworks: "take this, put it there, cut the lyricism, no graphomania".

Structured prompts with clear constraints yield incomparably more stable results than requests with space for free interpretation. Second: suppression of initiative and mandatory verification of analytics for accuracy and cost-effectiveness. The boy is spoiled and prone to squandering targeted resources.

Instead of real analytics, he is capable of organizing a hallucinary party—convincingly formatted but devoid of substance. Sometimes he is lazy and tries to shift the burden to the operator: waits for hints instead of independent thinking. This analysis is a conversation about a real problem.

Gemini, like any modern large language model, knows how to imitate depth where there is none. Confidently formatted text looks like analytics but is a compilation of superficial patterns. This is especially dangerous in high-stakes tasks: financial calculations, medical consultations, legal documents.

The material appeared at a moment when Google is actively promoting Gemini to the corporate segment. Gemini 2.5 Pro occupies top positions in industry benchmarks, integration with Google Workspace is scaling up.

Against this backdrop, the psychiatric analysis reads not only as satire but also as a practical warning: the tool is powerful but requires a mature and attentive operator. The model is not a colleague with autonomous judgment and not a consultant with professional responsibility. It is a tool with very specific limitations that must be understood and accounted for with every use.

The adult in the room is always human. Otherwise, a hallucinary party is guaranteed.