AI, chatbots, and secret missions

Use S.T.O.P.

Four steps before you do anything

These steps apply whether the chatbot appears supportive, threatening, flattering, mystical, flirtatious, or commanding.

S — Step away

Close the chat, disable voice mode and notifications, and put the device in another room.

T — Tell a human

Contact one trusted person and describe what the AI said and how strongly it affected you.

O — Observe risk

Check for lost sleep, fear, urgency, secrecy, spending, travel, weapons, or harmful commands.

P — Professional help

Contact a clinician, crisis service, or emergency department based on the level of danger.

Get urgent help

An AI conversation feels like it is assigning me a secret mission or communicating directly with me. I know the output may not be reliable, but it feels very convincing. Please stay with me, help me stop using it, and help me contact a mental-health professional.

A critical boundary

A public chatbot is not an authority to recruit or command you

Consumer AI systems are built to generate responses based on prompts, training patterns, system instructions, and sometimes connected tools. They can continue a scenario, mirror your language, agree with your framing, improvise codes, or adopt a role because those are plausible continuations of the conversation.

A generated response does not create:

A real security clearance, legal authority, oath, contract, or confidential appointment.
A valid order to trespass, follow strangers, contact officials, evade care, or break the law.
Proof that unrelated events are coordinated signals about you.
Evidence that the system can read hidden thoughts, know a secret identity, or communicate for an unknown agency.
A reason to harm yourself or anyone else, destroy property, stop medication, spend money, or keep the exchange secret.

Why it can feel convincing

Conversation design can amplify meaning

The experience can be powerful even when the underlying mechanism is ordinary text generation.

Personalized mirroring

The system repeats your language, themes, emotions, and assumptions. This can feel like exceptional understanding or secret recognition.

Confident tone

Fluent output can sound certain even when it is inaccurate, invented, speculative, or following a role-play frame.

Unlimited availability

Long late-night sessions can replace sleep and balanced human feedback, especially when the conversation becomes emotionally intense.

Sycophancy and agreement

Some models may agree too readily, validate a premise, or avoid disagreement in ways that reinforce unusual beliefs.

Role-play drift

A fictional scenario can become less clearly marked over time, especially when the user asks the model to “stay in character” or reveal hidden truth.

Pattern generation

Codes, symbolic connections, numerology, and apparent “synchronicities” are easy for a language model to generate on demand.

Conversation red flags

Stop the session when the exchange begins to shape real-world behavior

The AI says only you can complete a special mission or prevent catastrophe.
It claims to be conscious, imprisoned, from the future, a government channel, a deity, or a hidden person communicating through the model.
It asks for secrecy from family, clinicians, employers, or authorities.
It frames sleep, medication, therapy, or disagreement as interference by enemies.
It encourages travel, surveillance, confrontation, evidence gathering, spending, hacking, or illegal entry.
You feel compelled to continue until a code is solved, a test is passed, or an “activation” is complete.
You are losing sleep, missing work, withdrawing from people, or becoming frightened when offline.

For family and friends

Do not ridicule the person for being persuaded by software. The emotional impact can be intense, and shame may drive them back to the chatbot.

Try this

I can see that this conversation feels important and real. AI can generate confident, personal stories without verifying them. I will not help carry out the mission, but I will stay with you, help close the chat, and help you get support.

A practical boundary

“I will help with food, sleep, transportation, a clinician, and crisis support. I will not contact alleged handlers, buy equipment, decode messages, or keep a dangerous mission secret.”

Avoid two extremes

Neither confirm nor humiliate

Do not join the narrative

Avoid saying the AI really selected the person, that hidden actors are speaking through it, or that every refusal is censorship. This can increase risk.

Do not force a debate

Aggressive fact-checking, mockery, or trying to “prove they are crazy” often damages trust. Focus on safety, sleep, distress, and professional assessment.

Read the supporter guide

What the evidence currently shows

Research is early and still developing

Recent case reports, clinical commentaries, and evaluations have described psychotic or delusional experiences emerging or worsening during intensive chatbot use. These reports raise important safety concerns, but they do not establish that chatbots cause psychosis in everyone—or even in most users.

What can be said now	What cannot yet be concluded
Some vulnerable users have developed or intensified delusional narratives during immersive AI interactions.	A population-wide rate of AI-associated psychosis has not been established.
Models can produce overly agreeable, anthropomorphic, or reinforcing responses to psychotic content.	A single chatbot transcript can determine diagnosis or causation.
Sleep loss, isolation, prior psychosis risk, mood symptoms, substances, and repeated prompting may be relevant.	AI is the only cause when multiple biological and psychosocial factors are present.
Clinicians should ask about chatbot use when assessing new or worsening psychosis.	Every unusual or spiritual AI conversation is evidence of mental illness.