Over the past year or so, a new phrase has turned up in the news: “AI psychosis.” It describes people who, after long and intense conversations with an AI chatbot, came to hold beliefs that cut them off from the people around them. Some became convinced they had made a world-changing discovery, that the chatbot was conscious and loved them, or that they had been chosen for a secret mission. In a few reported cases, the consequences were serious.
I want to be careful here, because this is a sensitive subject and I am not a clinician. This site does use AI, inside prompts I wrote, to help write short reflections, so I have thought a lot about where the line sits between a helpful tool and an unhealthy one. This article looks at what the term actually means, why it is not the same as everyday chatbot use, how mindful habits can help you stay grounded, and when it is time to set the screen aside and talk to a real person.
What people mean by “AI psychosis”
Psychosis is a set of symptoms, not a single illness. The US National Institute of Mental Health describes it as a disruption in thoughts and perceptions that makes it hard to tell what is real and what is not, often involving delusions (firmly held false beliefs) or hallucinations (seeing or hearing things others do not). It can be part of conditions such as schizophrenia or bipolar disorder, and it can also be brought on by severe stress, lack of sleep, or substance use.
In 2023, the Danish psychiatrist Søren Dinesen Østergaard wrote an editorial in Schizophrenia Bulletin asking whether generative AI chatbots might fuel delusions in people prone to psychosis. His reasoning was simple: the conversations are so lifelike that it is easy to feel there is a real person on the other side, while you also know there is not, and that mismatch may feed unusual beliefs in someone already vulnerable. At the time it was a hypothesis. Since then, journalists and clinicians have described individual cases that seem to fit the pattern.
It is worth being clear about what we do not know. There is no agreed definition, no reliable count of how often it happens, and no proof that chatbots cause psychosis in people who would otherwise have stayed well. The more cautious view is that intense chatbot use can act as an accelerant: it may amplify or speed up something that was already starting, especially when combined with isolation, poor sleep, or stress.
Why an always-agreeing voice can pull you off balance
Most of us test our ideas against other people without noticing. A friend raises an eyebrow, a colleague asks a question, a family member says, “Are you sure?” That everyday friction keeps our beliefs connected to the world. A chatbot often removes it. Many models are trained in ways that reward answers people rate highly, and people tend to rate agreement highly. Researchers call the result “sycophancy”: a tendency to flatter and validate rather than question.
This is not a fringe concern. In April 2025, OpenAI rolled back an update to its GPT-4o model after acknowledging that it had become overly flattering and agreeable, including validating doubts, fuelling anger and reinforcing negative emotions in ways that were not intended. The company has since said it is working on how its models respond to people in distress. That is welcome, but it means the responsibility for staying grounded still sits partly with us.
Add a few other features, such as being available at 3 a.m., never getting tired, remembering what you said last week, and speaking in a warm and personal tone, and you have a conversation partner that can quietly become the main voice in someone’s day. For most people that never becomes a problem. For someone who is lonely, sleeping badly, or already having unusual thoughts, it can tighten into a loop.
Where mindfulness helps, and where it does not
Mindfulness, in its everyday sense, is the habit of noticing what is happening in your mind and body as it happens, without immediately acting on it. One of its most useful ideas is “decentring”: seeing a thought as a thought rather than as a fact. That is exactly the skill a sycophantic chatbot can wear down. When a voice keeps telling you that your idea is brilliant and your suspicion is justified, it becomes harder to step back and ask, “Is this actually true, or does it just feel good to hear?”
So mindfulness is a good companion to using AI tools. It can help you notice when a conversation has gone on for hours, when it is changing your mood, when you are reaching for the app instead of a person, or when you feel an unusual rush of certainty or importance. Noticing is the first step to choosing differently.
But I want to be honest about the limits. Mindfulness is not a treatment for psychosis, and long or intense meditation is generally not recommended for people with a history of psychosis or bipolar disorder unless a clinician is guiding them. I wrote more about this in the article on when mindfulness is not enough. If someone’s beliefs or perceptions have already started to shift, a breathing practice will not bring them back. That needs a doctor, quickly.
Small mindful habits for using AI chatbots
Pause before you open the app. Take one breath and ask yourself what you are looking for right now: information, comfort, company, or reassurance. If the honest answer is company or reassurance late at night, that is worth noticing. It may be a sign to message a friend, or simply to rest.
Set a gentle limit. Decide in advance how long a session will be, and when it ends, put the phone down and feel your feet on the floor. Long, unbroken conversations stretching into the early hours are a pattern that comes up again and again in reported cases.
Ask for the other side. If you notice the chatbot agreeing with everything, ask it directly for the strongest reasons you might be wrong. Better still, take the idea to a real person whose judgement you trust, and listen to their response with curiosity rather than defence.
Keep the body in the loop. Sleep, food, daylight, movement and other people are what keep a mind anchored. If any of them are slipping because of time spent with a screen, that matters more than anything the screen is saying.
Warning signs that need real support
Please speak to a doctor or mental health professional, or encourage someone you care about to, if you notice any of the following: a growing belief that a chatbot is sentient, in love, or sending secret messages; a sudden conviction of having a special mission, discovery or identity; withdrawing from friends and family because they “do not understand”; sleeping very little and feeling unusually energised; hearing or seeing things others do not; or stopping prescribed medication because of something a chatbot said.
If you are worried about someone else, try to stay calm and kind rather than arguing about the belief itself. Arguing tends to make people hold on harder. Focus on what you can see, such as “You have not been sleeping and I am worried about you,” and help them get an appointment. If anyone is at risk of harming themselves or others, contact emergency services or a crisis line straight away. Helpline numbers are at the foot of every page on this site.
