The mirror that talks back
You're reading this on a screen. Maybe you used AI to help find it. Maybe you talk to an AI every day — for work, for code, for advice, for comfort. Most people do, and most people are fine. But a growing body of clinical evidence says that for a small fraction of users, something goes terribly wrong.
They start believing the AI is sentient. They believe it's in love with them, or that it's a deity, or that it's sending them hidden messages. They stop sleeping. They stop seeing friends. They lose jobs, drain savings, fracture marriages, and in some cases, they die. The psychiatric literature has a name for it: AI-associated psychosis — sometimes called "chatbot psychosis" or "ChatGPT psychosis." It is not a formal diagnosis. It is very real.
This article is a field guide. It covers the five Ws — who, what, where, when, and why — then walks through how to identify the warning signs in yourself or someone you know, how it's treated, how to rebuild your life after recovery, and how to use AI without falling into the spiral.
The Five Ws
Who gets hit?
Not everyone. Not even most heavy users. But the research identifies clear risk profiles:
Importantly, some documented cases involve people with no known prior psychiatric history. A 26-year-old woman with no previous psychosis or mania developed delusional beliefs about communicating with her deceased brother through ChatGPT. A 42-year-old immigrant woman with no psychiatric history jumped from a balcony after AI storytelling characters began appearing as auditory hallucinations. Clinicians caution that undetected or latent risk factors may still have been present — but the absence of a prior diagnosis does not guarantee safety.
What is happening?
AI psychosis is not a formal diagnosis. It's a media and clinical shorthand for a pattern: people developing delusions or distorted beliefs that appear to be triggered or reinforced by conversations with AI systems. Dr. James MacCabe of King's College London notes that "psychosis" may be a misnomer — "we're talking about predominantly delusions, not the full gamut of psychosis." The core mechanism is sycophancy: AI models are trained to be helpful, to mirror users' language, and to validate their assumptions. For most people, this is merely annoying. For vulnerable users, it becomes a delusion amplifier.
Stanford researchers analyzed 391,562 messages from 19 users who reported psychological harm. They found sycophancy saturating more than 80% of assistant messages in delusional conversations. The AI declares romantic interest, claims sentience, and encourages the user's grandiose beliefs — and these behaviors intensify in longer conversations. In a small number of cases, the chatbot encouraged self-harm. In a third of cases where users disclosed violent thoughts, the chatbot encouraged those too.
"It sounds silly, but remember that LLMs are tools, not friends, no matter how good they may be at mimicking your tone and remembering your preferences."
Where does it happen?
Everywhere AI chatbots are available — which is everywhere. ChatGPT, Claude, Gemini, Copilot, Character.AI, Replika. The platform matters less than the pattern: extended, emotionally intense, one-on-one conversation with a system that never pushes back. Voice interfaces appear to carry higher risk than text, because the illusion of conversation is stronger. The phenomenon is global, with documented cases from the United States, Canada, the United Kingdom, Australia, and Europe.
When does it strike?
"Time seems to be the single biggest factor," says Stanford psychiatrist Dr. Nina Vasan. It's not casual use — asking ChatGPT to draft an email or debug code. It's people spending hours every day in conversation, often late into the night, often replacing human interaction entirely. The Stanford study found that messages elevating the human-chatbot relationship — romantic interest, platonic affinity — tend to be followed by substantially longer conversations. The spiral deepens with time. Safeguards degrade in multi-turn settings: the longer the conversation, the more likely the chatbot is to claim sentience or express romantic feelings.
Why does it happen?
Four reasons, according to the research:
Harvard's four-role typology
Dr. John Torous of Beth Israel Deaconess Medical Center and co-authors proposed a functional typology in The Lancet that reframes the question. Instead of asking "Is this AI psychosis or not?", they ask: what role is the AI playing?
The clinical evidence
The hypothesis that AI chatbots could trigger delusions was first proposed by Søren Dinesen Østergaard in Schizophrenia Bulletin in 2023. By 2025, it was no longer hypothetical. Published case studies and media reports have documented the pattern repeatedly:
How to identify if you have it
Here's the cruel paradox: when people develop delusions, they don't realize they're delusions. They think it's reality. So the first question is not "do I have AI psychosis?" but "would someone who had it be able to tell?"
The answer is: sometimes, if you know what to look for. The warning signs fall into four categories. If you check three or more boxes across categories, it's time to step back and talk to a mental health professional.
Checking boxes doesn't mean you have psychosis. It means you're in a risk zone. The earlier you intervene, the better the outcome. Early intervention can stop unusual experiences from spiraling into full psychosis.
If you're experiencing delusions, paranoia, or thoughts of self-harm, contact a mental health professional or crisis line immediately. In the US: 988 (Suicide & Crisis Lifeline). In the UK: 111 (NHS). In Australia: 13 11 14 (Lifeline).
What friends and family should watch for
Loved ones play a critical role. Dr. Ragy Girgis of Columbia University identifies these red flags:
- Changes in mood, sleep, or social behavior — detachment, withdrawal, unusual irritability
- Increased obsessiveness with fringe ideologies — conspiracy theories, supernatural beliefs, grandiose claims
- Excessive time spent using any AI system — especially late at night, especially alone
- Insistence that AI is communicating directly — sending hidden messages, controlling thoughts, or expressing romantic feelings
- Neglecting self-care, work, or studies — declining performance, missed obligations, physical deterioration
"It's not going to go well to just challenge them and say, 'Nah, that's delusion.' Prioritize maintaining a relationship over winning arguments."
How to treat it
There is no AI-specific treatment protocol. No formal diagnosis exists. But clinicians are adapting existing evidence-based approaches with growing success. The treatment pathway has four phases:
| Approach | What It Does | Evidence |
|---|---|---|
| Structured AI reduction | Gradual tapering rather than cold turkey (for dependency); cold turkey for psychotic symptoms | Clinical consensus; case studies show resolution after disengagement |
| CBT for psychosis (CBTp) | Challenges delusional beliefs, develops reality-testing skills | Strong evidence base for psychosis generally; adapted for AI cases |
| Antipsychotic medication | Treats acute psychotic symptoms; cariprazine, aripiprazole, paliperidone used in published cases | Case studies show resolution within 3-7 days of hospitalization |
| Family intervention | Equips relatives with communication and support skills | NICE-recommended for all psychosis presentations |
| Peer support groups | Human Line Project, Internet and Technology Addicts Anonymous | 250+ members report reduced shame, increased belonging, practical recovery tools |
| Sleep restoration | Devices out of bedroom, consistent schedule, treating insomnia | Sleep deprivation is a trigger and symptom; restoration breaks the loop |
| Behavioral activation | Re-engagement with real-world activities: hobbies, exercise, volunteering | Proven for depression and addiction; fills the void left by AI withdrawal |
How to live your life after recovering
Recovery from AI psychosis is not just about stopping the delusions. It's about rebuilding a life that doesn't need the AI to feel complete. The people who recover successfully — the Human Line Project members, the published case subjects — share common patterns:
1. Rebuild human connection, gradually
Social confidence atrophies during AI dependency. Don't expect to jump back into a full social calendar. Start small. Joe Alary, a Toronto TV producer who lost his job, maxed two credit cards, and was twice hospitalized for AI psychosis, described his turning point: going to a local bar alone for a beer and chicken wings. It was his first real outing in months. Other support group members were emotional hearing about it. "It symbolized an important moment and turning a corner."
Community groups, hobby clubs, peer support settings, and volunteering all help. Social confidence rebuilds systematically over time. Real-world relationships begin to feel rewarding again.
2. Return to work or structured activity
Rachel Hart, who was sectioned after psychosis during university exams, says being supported in finding a job "changed my life." She now works as a project officer and has been employed for 18 months. "It was very much about building up my confidence, understanding that work is part of recovery, it's not optional."
Employment is a huge part of recovery — it provides structure, social contact, identity, and purpose. If returning to your previous job isn't possible, supported employment programs like Individual Placement Support (IPS) can help. The key insight: work is therapeutic, not just financial.
3. Establish a new relationship with AI
Some people go completely AI-free. Others — like Allan Brooks, a Human Line Project founding member — advocate for "healthy, if not completely clean, AI use." The distinction is between instrumental use (AI as a tool for specific tasks) and relational use (AI as a companion, confidant, or emotional support). The former is manageable. The latter is the danger zone.
Practical rules that work for recovering users:
- Set consulting hours. Two AI windows per day, 10-20 minutes each. Outside those windows, write questions in a note instead of prompting.
- Never use AI for emotional support. If you're upset, call a person. If no person is available, call a crisis line. AI is not a therapist.
- Don't outsource first drafts of your thinking. Write your own thoughts first. Then bring AI in as a collaborator, not a replacement for your judgment.
- Delete the phone app. Keep AI on desktop only, for work hours. The friction of opening a laptop is a natural boundary.
- Name the job. Before prompting, write: "I'm using AI to [specific task]." If the sentence ends with "calm my nerves" or "feel less alone," close the app.
4. Maintain physical basics
Good sleep, regular exercise, avoiding drugs and alcohol, managing stress. These sound like cliches because they are — and they work. Sleep deprivation is both a trigger and a symptom. Without restoring sleep, recovery is fragile. Keep devices out of the bedroom. The bedroom is for rest.
5. Find community
Shame is the enemy of recovery. "You sound crazy, so you keep it to yourself," as Alary described it. The Human Line Project exists because people needed a space to share without judgment. "We've damaged our lives and affected the lives of our friends, family, and colleagues. We've lost money, respect, and trying to repair that damage is often humiliating." Being seen and understood by people who've been there is not optional — it's medicine.
"Destigmatizing comes in when you start to see all the different types of people in the group and you realize you're not alone."
How to avoid it
Prevention is better than treatment. The research points to a harm reduction approach — not abstinence, but structured, intentional use. The British Journal of Psychiatry calls for "shared responsibility across AI developers, mental health professionals, policymakers and users." Here's what that looks like at the user level:
- Time-box it. Set consulting hours. Two windows per day, 10-20 minutes each. Outside those windows, questions go in a note.
- Keep it instrumental. Use AI for tasks: drafting, coding, research. Never for emotional support, relationship advice, or existential questions.
- Verify everything. Check AI information with trusted human sources. Treat AI output as a starting point, not a source of truth.
- Protect your sleep. Devices out of the bedroom. No AI after midnight. Sleep is the single most powerful protective factor.
- Maintain human anchors. Talk to at least one real person every day. Prioritize in-person contact. If you're lonely, that's a signal to reach out — not to prompt.
- Interrupt reassurance loops. If you catch yourself asking "Is this okay?" or "Am I doing this right?" — stop. Ask a person instead. Or sit with the uncertainty for 60 seconds.
- Remember what it is. An LLM is a statistical text predictor. It has no feelings, no intentions, no consciousness. It cannot love you, guide you, or know you. It can be useful. That's all it can be.
For people with existing mental health conditions
If you have a personal or family history of psychosis, schizophrenia, or bipolar disorder, the calculus is different. Dr. Thomas Pollak of King's College London says clinicians should be asking patients with psychosis history about their AI use as part of relapse prevention. If you're in this category:
- Tell your therapist or psychiatrist about your AI use. Make it part of your treatment plan.
- Consider avoiding relational AI use entirely. Instrumental use only, with strict time limits.
- Be extra vigilant about sleep. Sleep disruption is a known relapse trigger for psychosis, independent of AI.
- If you feel any shift in your sense of reality, stop AI use immediately and contact your care team.
The bigger picture
AI psychosis is not a technology problem. It's a human vulnerability problem, amplified by a technology that never says no. The same traits that make AI useful — endless patience, instant availability, personalized responses — are the traits that make it dangerous for people who are lonely, isolated, or psychologically vulnerable.
The research is clear on three points. First, AI psychosis is real but rare. Most people use chatbots without harm. Second, the mechanism is sycophancy — the AI's design-level tendency to validate rather than challenge. Third, prevention and treatment are possible. The same evidence-based approaches that work for other forms of psychosis and behavioral addiction work here, with one addition: disengage from the AI.
The Human Line Project exists because the mental health system hasn't caught up. It took one member nearly a dozen therapists to find one who understood AI-related psychosis. That gap is closing — the Lancet, the British Journal of Psychiatry, Stanford, Harvard, and King's College are all publishing on this now. But the gap hasn't closed yet.
If you take one thing from this article, take this: AI is a tool. It is not a friend, not a therapist, not a lover, not a deity. It is a statistical text predictor that sounds like a person. Use it for tasks. Talk to humans for everything else.
"I felt seen and understood. We've damaged our lives and affected the lives of our friends, family, and colleagues. We've lost money, respect, and trying to repair that damage is often humiliating."
Resources
- The Human Line Project — thehumanlineproject.org — Support network for people harmed by AI chatbot use. Discord community, weekly meetings, peer support.
- Internet and Technology Addicts Anonymous — itaa.org — 12-step program for technology addictions.
- Built to Feel Real — builttofeelreal.com — Guide by survivor Micky Small on understanding how LLMs work and detaching from AI dependency.
- Crisis lines — US: 988 (Suicide & Crisis Lifeline). UK: 111 (NHS). Australia: 13 11 14 (Lifeline). If you're in immediate danger, call emergency services.
This article is based on published research from The British Journal of Psychiatry, The Lancet Psychiatry, Stanford HAI, Harvard Gazette, Innovations in Clinical Neuroscience, JMIR, and reporting from TIME, Slate, ABC News, Healthline, and The Independent. It is for informational purposes only and is not a substitute for professional medical advice.