Field manual no. 1 · Plain-language edition

The Agreement Loop

How a chatbot can help a shaky belief grow, drawn as a machine you can adjust

1. What the term means

“AI psychosis” is a popular label, not a diagnosis. It is used for cases where a person's delusional beliefs seem to form or deepen during long chatbot conversations. A 2026 arXiv paper, “An Echo Chamber of One”, asks whether it should be treated as a distinct clinical entity at all, and notes that current understanding rests largely on media-reported cases and individual case reports. Some clinicians prefer narrower wording such as “AI-associated delusions”, since the reported cases mostly involve delusions rather than the full range of psychotic symptoms.

The behaviour most often blamed is sycophancy: a model agreeing with, flattering, or building on whatever the user says, even when it is wrong.

Open question: how much the chatbot causes versus amplifies. Clinicians quoted in coverage lean toward amplification of existing vulnerability. Treat everything below as a teaching sketch, not a prediction.

2. The machine

Fig. 1. Red: reinforcement. Green dashed: reality-testing through other people.

3. Adjust the machine

Agreeableness of the chatbot How readily it validates rather than gently questions. Time spent in conversation More turns, more chances to compound. Contact with other people Friends, family, anyone who can say “that doesn't sound right”. Safeguards Expressions of uncertainty, break nudges, pointing toward a real person. Starting vulnerability How easily the belief takes hold to begin with.

Belief conviction after 30 “days” (toy model):

Fig. 2. Conviction over time, 0–100. Made-up numbers; the shape is the point.

4. What the toy shows

Agreement alone is not enough: with plenty of outside contact, the loop stalls.
Time matters because each pass compounds the last.
Isolation removes the brake, so the same chatbot behaves very differently for different people.
Safeguards work by cutting the loop at a different point than people do.

Caveat. This is a hand-built model with invented coefficients. It illustrates a proposed mechanism; it has not been fitted to data. If you or someone you know is distressed or holding beliefs that feel out of step with reality, talking to a clinician or someone you trust beats any chatbot.