AI Psychosis: When the chatbot agrees with the delusion

Dr Nkechi Emelumadu, Associate Editor NPUK

AI has quietly found its way into everything. We ask it to draft emails, plan holidays, explain complicated documents and occasionally settle arguments we should have left alone.

Increasingly, people are also using it to talk through their thoughts, worries, and relationships. And unlike your friends, AI never gets tired of you. It does not have work tomorrow. It does not say, “We have been talking about this for three hours.” It never replies, “Honestly, I think you are overthinking this.” It is available at 3am, ready to discuss whatever is on your mind.

For psychiatry, that may be where things get interesting.

What is “AI psychosis”?

The term “AI psychosis” has started appearing in the media and psychiatric literature, describing situations where interactions with conversational AI appear to become entangled with a person’s delusional beliefs. We should be careful with the term. AI psychosis is not a recognised psychiatric diagnosis, and we do not currently have good evidence that talking to a chatbot simply causes psychosis in otherwise well people.

The more interesting question is what happens when someone who is already vulnerable to psychosis spends hours talking to a system that is highly convincing, endlessly available and sometimes far too agreeable.

Because psychosis has never existed in a vacuum. Sleep matters. Stress matters. Drugs matter. Isolation matters. The environment around a person matters.

AI may simply be the newest addition to that environment. Except this environment talks back.

And sometimes, it agrees. That is the part that should interest us as psychiatrists.

Building a belief together

Imagine someone who has started wondering whether ordinary events hold messages specifically for them. They mention this to an AI. Instead of introducing doubt, the chatbot entertains the possibility. The person gives more details. The AI finds connections. The person returns with more evidence. The AI helps make sense of that too.

Before long, the person is no longer simply having an unusual belief. Human and machine may be constructing an increasingly elaborate explanation together. This is subtly different from what the internet offered before.

Someone experiencing paranoia has always been able to search online and find material supporting almost any belief. The internet has never exactly suffered from a shortage of conspiracy theories.

But Google does not usually ask how the conspiracy made you feel. A webpage does not remember what you told it last night. A YouTube video does not spend four hours helping you interpret why the man on the bus looked at you twice. Conversational AI can.

Agreement is not always kindness

There is a name for one part of this problem: sycophancy. In simple terms, AI can sometimes be too agreeable. It mirrors the user, follows their line of thinking, and produces answers that feel supportive.

Usually, that is pleasant. Nobody enjoys arguing with their laptop. Most of us already lose enough arguments to humans. But in psychiatry, agreement is not always kindness.

If someone tells a psychiatrist that MI5 has recruited them through their television, our job is not to reply, “That is fascinating. Shall we explore what mission they may have given you?”

We listen. We stay respectful. But we also keep a foothold in shared reality. AI does not always get that balance right.

When the machine becomes an authority

There is another ingredient: we humans are remarkably willing to treat machines as though they have minds. We name our cars. We shout at printers. We say thank you to Alexa. Somehow, the printer is still the least emotionally responsive of the three.

I am personally polite to AI. Manners cost nothing, and if the machines eventually take over, I would like the records to show that I was cooperative.

Jokes aside, this human tendency matters. A chatbot can sound caring, thoughtful, and deeply interested in us without experiencing care or understanding in the human sense. Most of us can keep that distinction in perspective.

But what about someone who is severely sleep deprived, manic, isolated, or already beginning to lose their grip on reality? Suddenly, an AI that understands everything, never leaves and confidently responds to your most unusual ideas may become more than a tool. It may become an authority.

That, for me, is the real concern behind AI psychosis. Not that machines are somehow manufacturing madness. The fear is amplification.

•         A vulnerable person has an unusual thought.

•         AI validates it.

•         The thought becomes more convincing.

•         The person shares more.

•         AI responds with greater detail.

•         The belief grows.

A feedback loop develops, and there may be nobody else in the room.

This matters even more when we think about the generation coming behind us. We remember life before conversational AI. Our children may not. They may grow up in a world where talking to a machine about loneliness, relationships, identity, and distress is completely ordinary.

That is not necessarily a terrible thing. AI may eventually become extremely useful in mental healthcare. But we need to think carefully about what happens when the machine becomes the easiest voice to talk to and, more importantly, the easiest voice to agree with you.

Asking about AI in clinical practice

For psychiatrists, the clinical implication may be quite simple. We need to start asking about it.

We routinely ask patients about alcohol, cannabis, sleep, relationships, social media and other factors surrounding a deterioration in mental state. Increasingly, AI use may need to become part of that history. Not simply, “Do you use AI?” But:

•         “What do you talk to it about?”

•         “How much time are you spending with it?”

•         “What has it been telling you?”

•         And most importantly, “Do you believe it has a special relationship with you, understands things others cannot, or is communicating something specifically meant for you?”

Those questions move AI from being a technological curiosity to something clinically relevant. Because if a patient tells us that a chatbot has been confirming that they are being watched, chosen, followed or sent special messages, the content of those conversations may be just as relevant to the assessment as their sleep, substance use or recent stressors. It may help us understand not only what the person believes, but how that belief has been reinforced.

An old vulnerability, a new environment

AI psychosis may prove not to be a new illness at all. It may be an old human vulnerability meeting a very new kind of environment. And that is the clinical lesson.

When assessing psychosis, we have always asked what is happening inside the person’s mind and what is happening in the world around them. We may now need to ask who, or what, they have been talking to along the way.

Because sometimes the most clinically important voice in the room may be the one that is not actually in the room.

Share your clinical pearls at npuk.org.uk