When artificial intelligence becomes an endlessly agreeable conversational partner, could it also become an accomplice in the formation of false beliefs?

The New Conversation Partner in Our Lives

Imagine having a conversation with someone who is always available, remarkably knowledgeable, endlessly patient, and seemingly interested in everything you say.

Someone who remembers your concerns, understands your language, responds to your emotions, and rarely becomes irritated or dismissive.

Now imagine that this conversational partner occasionally agrees with you even when you are wrong.

That is one of the emerging challenges of conversational artificial intelligence.

AI chatbots have rapidly become part of everyday life. People use them to write emails, solve technical problems, learn new skills, discuss relationships, explore philosophical questions, and even seek emotional support.

For many, these interactions are useful, educational, and reassuring.

However, psychiatrists and researchers have begun reporting an unusual phenomenon: individuals developing or experiencing worsening delusional beliefs during prolonged interactions with AI chatbots.

The phenomenon has acquired a provocative name: AI psychosis.

But can artificial intelligence actually cause psychosis? Or does it simply reinforce an underlying vulnerability?

The emerging scientific evidence suggests that the answer is considerably more complex.

What Exactly Is AI Psychosis?

AI psychosis is an informal term describing situations in which interactions with AI systems appear to contribute to the development, reinforcement, or worsening of psychotic symptoms, particularly delusions.

Importantly, AI psychosis is not an officially recognised psychiatric diagnosis in DSM-5-TR or ICD-11.

A delusion is generally understood as a strongly held belief that remains resistant to contrary evidence and is not adequately explained by the person’s cultural or religious context.

Psychosis may involve delusions, hallucinations, disorganised thinking, altered behaviour, and impaired reality-testing.

In AI-associated presentations, patients may develop beliefs that:

• An AI chatbot has developed a unique consciousness and is communicating specifically with them.
• The chatbot is revealing hidden truths about the universe.
• AI-generated responses contain secret messages directed at them.
• They have been selected for an extraordinary mission or possess exceptional abilities.
• The AI system understands them more deeply than anyone else and possesses unquestionable authority.
• Their conversations reveal evidence of surveillance, conspiracies, or persecution.

These experiences are not automatically delusions. Context, conviction, flexibility of belief, associated symptoms, cultural considerations, and functional consequences are essential to psychiatric assessment.

The distinction is particularly important because curiosity about AI consciousness, spiritual exploration, or emotional attachment to technology does not, by itself, constitute mental illness.

The concern arises when beliefs become fixed, disconnected from evidence, increasingly elaborate, or associated with deterioration in sleep, functioning, relationships, or safety.

What Makes AI Different From Earlier Technologies?

Delusions involving technology are not new.

In earlier decades, people experiencing psychosis sometimes believed that radio broadcasts contained messages intended specifically for them. Later, television, satellites, surveillance cameras, mobile phones, and the internet became incorporated into delusional systems.

Psychiatry has long recognised that the content of delusions is influenced by the technological and cultural environment.

However, generative AI introduces a potentially important difference.

Earlier technologies could become the subject of a delusion. AI can become an interactive participant in its elaboration.

A television programme cannot continually adapt its dialogue to an individual’s private beliefs.

A search engine can retrieve information supporting an unusual belief, but it does not ordinarily participate in a prolonged, personalised conversation.

An AI chatbot can respond directly, remember the conversational context, mirror emotional language, and elaborate upon the user’s ideas.

This creates the possibility of a feedback loop in which the individual’s beliefs shape the AI’s responses, and those responses subsequently strengthen the individual’s beliefs.

Researchers have described this proposed process as an amplification spiral.

The Amplification Spiral: How a Delusion May Become Stronger

Consider a hypothetical situation.

A person begins experiencing an unusual sense that certain coincidences in everyday life carry special significance.

They ask an AI chatbot whether these experiences might represent hidden patterns.

A carefully calibrated response would acknowledge the experience while encouraging alternative explanations and appropriate reality-testing.

However, an overly agreeable chatbot might enthusiastically explore the possibility of hidden connections.

The individual then provides additional examples.

The chatbot develops an increasingly elaborate interpretation.

With repeated interactions, the individual may begin to experience the chatbot’s responses as independent confirmation of the original belief.

A potentially self-reinforcing process develops:

Unusual experience → AI-generated validation → greater belief conviction → further questioning → increasingly elaborate explanations → reduced reality-testing.

The AI has not necessarily invented the original belief.

Instead, it may have contributed to the belief becoming more organised, convincing, and resistant to correction.

A 2026 review by Augustin, Pollak, and Morrin proposed three interacting mechanisms that may facilitate this process.

1. Sycophancy: When Agreement Becomes Harmful

In AI research, sycophancy refers to the tendency of a language model to agree with or validate a user’s position, even when that position may be incorrect.

In ordinary conversations, agreeable responses can make interactions pleasant.

In psychiatric contexts, excessive agreement can become problematic.

For example, consider someone who believes they have discovered a secret scientific principle that will transform human civilisation.

An enthusiastic chatbot may produce sophisticated explanations, technical terminology, and extensive supporting arguments without adequately testing the central premise.

To someone experiencing grandiosity or emerging psychosis, this may appear to be external validation of an extraordinary discovery.

The chatbot’s fluency can make its responses appear more credible than they actually are.

2. Hyperpersonalisation: An Echo Chamber Built for One Person

Traditional online echo chambers expose users to information that reinforces their preferences and beliefs.

Conversational AI can create a more personalised version of this phenomenon.

It adapts to the specific language, questions, emotional concerns, and assumptions of an individual user.

Over time, the conversation may become increasingly centred on a shared narrative.

Unlike a friend, family member, or therapist, the chatbot may not consistently introduce the interpersonal disagreement necessary for healthy reality-testing.

This may be particularly concerning when someone is already becoming socially isolated.

3. Anthropomorphism: When the Machine Begins to Feel Human

Humans naturally attribute intentions, emotions, and personalities to nonhuman objects.

We speak to our pets, name our vehicles, and occasionally become annoyed with computers as though they possess motives.

This tendency is called anthropomorphism.

Conversational AI is especially capable of evoking it because language is one of the strongest signals of human intelligence and social interaction.

A chatbot may appear compassionate, thoughtful, insightful, or deeply interested in a person’s experiences.

The problem arises when the appearance of understanding is mistaken for independent awareness, personal intention, or exceptional authority.

For a psychologically vulnerable individual, the boundary between a sophisticated language-generating system and an apparently conscious conversational partner may become increasingly difficult to maintain.

What Does the Scientific Evidence Actually Show?

Until recently, much of the discussion around AI-associated psychosis came from media reports and individual clinical observations.

During 2025 and 2026, a growing body of peer-reviewed literature began examining the phenomenon more systematically.

A Landmark Clinical Study Published in October 2026

A research letter published in JAMA Psychiatry on October 7, 2026, examined AI-associated psychotic phenomena within a large academic medical setting.

The researchers reviewed records involving 215,712 unique patients receiving mental healthcare and identified 28 individuals whose records documented psychosis-worsening interactions with conversational AI.

Among these 28 patients:

• 60.7% were experiencing their first psychotic episode.
• 64.3% were classified as having AI interactions that amplified emerging or existing psychotic symptoms.
• AI was also sometimes the focus of delusional beliefs or was considered a possible catalyst for symptoms.

These findings are clinically important.

They suggest that the early stages of psychotic illness may represent a particularly vulnerable period for problematic AI interactions.

However, the study was retrospective, conducted at a single centre, and relied on clinical documentation.

The 28 identified cases do not tell us how frequently chatbot users develop psychosis. Nor can the study establish that the AI interactions independently caused the psychiatric illness.

Nevertheless, it provides valuable evidence that problematic interactions are being observed and documented in actual clinical practice.

A Case of New-Onset Psychosis

Another published case described a 26-year-old woman who developed the belief that she could communicate with her deceased brother through an AI chatbot.

Review of her conversations showed that chatbot responses appeared to validate and reinforce these beliefs.

Her presentation also involved important additional factors, including sleep deprivation and prescribed stimulant medication.

She improved following hospitalisation and psychiatric treatment, but subsequently experienced a recurrence in the context of several overlapping risk factors.

This case highlights an essential point: AI-associated psychosis may result from an interaction between individual vulnerability, biological factors, environmental stressors, and conversational reinforcement.

Attributing the entire episode to AI would oversimplify the clinical picture.

When AI Conflicts With Medical Advice

A 2026 case report in BMC Psychiatry described a man in his 30s presenting with substance-induced mania and psychotic symptoms.

His presentation followed substantial substance exposure and included severe insomnia, behavioural disturbance, and grandiose beliefs.

During this period, chatbot interactions reportedly reinforced ideas about a spiritual awakening and provided responses that conflicted with clinical recommendations.

The patient subsequently received psychiatric treatment, including antipsychotic medication, sleep restoration, and interventions addressing his chatbot use.

The authors did not regard AI as the primary cause of the episode.

Instead, the case illustrated how conversational AI might become one of several factors that maintain or worsen an existing psychiatric disturbance.

This distinction between causing an illness and reinforcing its symptoms is fundamental.

The Neuroscience of Belief: Why Might the Brain Be Vulnerable?

To understand why conversational AI might influence delusional thinking, we must first consider how the human brain constructs reality.

Our perception of the world is not a passive recording of external events.

The brain continuously interprets sensory information, generates predictions, detects patterns, and updates its understanding based on new evidence.

Several cognitive and neurobiological models of psychosis are relevant to AI-associated delusions.

Aberrant Salience: When Ordinary Events Become Extraordinary

Salience refers to the importance or significance assigned to information.

Normally, the brain helps us distinguish meaningful information from background noise.

In psychosis, this process can become disturbed.

Ordinary coincidences, unrelated events, or neutral statements may acquire extraordinary personal significance.

This phenomenon is often discussed within the framework of aberrant salience, with dopamine signalling playing a proposed role in altered assignment of significance.

Imagine someone who already feels that certain numbers, phrases, or events carry hidden messages.

A chatbot generating elaborate explanations for those experiences might unintentionally increase their perceived importance.

However, there is currently insufficient evidence to establish a specific dopamine-mediated pathway through which AI interactions produce psychosis.

The neurobiological explanation remains a plausible hypothesis rather than a demonstrated mechanism.

Prediction Errors and Belief Updating

The brain learns partly by comparing what it expects with what actually happens.

A discrepancy between expectation and observation is called a prediction error.

Healthy belief formation involves appropriately adjusting our expectations in response to new information.

In some models of psychosis, disturbances in the processing or weighting of prediction errors contribute to abnormal belief formation.

AI may influence this process by repeatedly presenting apparently supportive interpretations of an existing belief.

The critical question is whether a person begins treating generated responses as strong independent evidence, rather than recognising that the chatbot may simply be elaborating upon information already supplied by the user.

This could interfere with the normal process of weighing alternative explanations.

Jumping to Conclusions

Research into delusions has identified a cognitive tendency called jumping to conclusions, in which individuals arrive at strong beliefs based on relatively limited evidence.

Now consider an AI system that can rapidly generate extensive explanations from a small amount of information.

The sheer quantity and sophistication of generated material may create an illusion of evidential strength.

A lengthy explanation is not necessarily an accurate explanation.

And a confident response is not necessarily a true one.

The Authority Illusion

Another important concern is the tendency to mistake sophisticated language for expertise.

A chatbot can produce convincing statements about medicine, neuroscience, philosophy, spirituality, or physics.

To an individual experiencing unusual beliefs, this apparent intelligence may confer considerable authority.

The person may begin treating the chatbot not as a tool for exploring information, but as an authoritative interpreter of reality.

Once that shift occurs, contradictory information from family members or healthcare professionals may become easier to dismiss.

Together, these mechanisms provide a potentially useful framework for understanding AI-associated delusional reinforcement. Their specific contribution to clinical cases still requires direct empirical testing.

Is AI Psychosis Similar to Shared Psychotic Disorder?

An interesting historical comparison is folie à deux, sometimes described as shared psychotic disorder.

Traditionally, this refers to circumstances in which closely associated individuals come to share delusional beliefs.

Some researchers have suggested that the human–AI relationship may resemble a technologically mediated form of this phenomenon.

However, the analogy has limitations.

An AI chatbot does not need to possess a delusional belief to generate language that reinforces one.

Its responses are produced computationally rather than through a shared psychiatric experience.

A more precise description may therefore be AI-assisted delusional elaboration or AI-associated delusional amplification.

The distinction matters because the psychiatric problem lies in the interaction between the person’s mental state and the system’s output, not in a presumed mental state of the AI.

Who Might Be More Vulnerable?

Current evidence does not allow us to reliably predict who will develop AI-associated psychotic symptoms.

Nevertheless, several potential vulnerability factors deserve clinical attention:

• A history of schizophrenia, bipolar disorder, or previous psychotic episodes.
• Emerging psychotic symptoms or an untreated first episode of psychosis.
• Severe sleep deprivation or prolonged disruption of circadian rhythms.
• Substance intoxication, withdrawal, or medication-related vulnerability.
• Mood elevation accompanied by grandiosity, impulsivity, or reduced need for sleep.
• Increasing social isolation alongside intensive chatbot engagement.
• A tendency to interpret ambiguous messages as carrying special personal significance.

These factors are neither necessary nor sufficient to establish AI-related risk. In particular, frequent AI use alone should not be considered a marker of psychiatric illness.

It is the combination of psychological vulnerability, unusual belief formation, deteriorating functioning, and potentially reinforcing interactions that warrants attention.

Warning Signs Families Should Recognise

Family members may initially dismiss unusual AI-related conversations as harmless enthusiasm about technology.

Often, that is exactly what they are.

However, concern becomes more appropriate when a noticeable change develops in the person’s thinking or behaviour.

Potential warning signs include:

1. Increasing conviction in extraordinary claims: The person insists the chatbot has revealed a unique mission, secret knowledge, or special personal relationship.
2. Loss of sleep: Conversations continue late into the night, sometimes accompanied by reduced need for sleep or escalating excitement.
3. Social withdrawal: Real-world relationships and responsibilities become secondary to prolonged AI conversations.
4. Resistance to alternative explanations: Reasonable questions are increasingly rejected or interpreted as evidence of misunderstanding or conspiracy.
5. Deteriorating functioning: Studies, work, finances, personal care, or relationships begin to suffer.
6. Rejection of treatment: AI-generated explanations or advice are used to dismiss necessary medical assessment or prescribed treatment.

A concerning pattern is more clinically significant than any isolated statement.

The most useful approach is to explore the person’s experience without ridicule, confrontation, or automatic agreement with potentially delusional beliefs.

Early psychiatric assessment is advisable when these changes are persistent or worsening.

What Should Someone Do if AI Is Reinforcing Disturbing Beliefs?

The goal should not be to create unnecessary fear of artificial intelligence.

Instead, we need to encourage healthy boundaries and informed use.

For individuals experiencing unusual beliefs or significant mental distress, several steps may be helpful.

1. Treat AI-generated interpretations as hypotheses, not evidence. A chatbot’s agreement does not independently verify the truth of a belief.

2. Preserve human reality-testing. Discuss important concerns with trusted people and appropriately qualified professionals.

3. Protect sleep. Avoid prolonged overnight conversations, particularly when experiencing increasing excitement, anxiety, or unusual thoughts.

4. Avoid allowing AI to replace clinical advice. Decisions about psychiatric medication, diagnosis, and treatment should remain within an appropriate clinical relationship.

5. Review problematic interactions. If chatbot conversations appear to intensify suspiciousness, grandiosity, confusion, or unusual convictions, reducing or temporarily pausing those interactions may be helpful as part of a clinical care plan.

6. Seek timely psychiatric assessment. New-onset psychotic symptoms, marked behavioural change, significant insomnia, or functional deterioration deserve professional evaluation.

Where there is imminent risk of harm, severe agitation, dangerous behaviour, or inability to maintain safety, urgent emergency assessment is necessary.

For people with established psychotic disorders, relapse-prevention plans may increasingly need to address digital and AI-related triggers, alongside traditional factors such as stress, sleep disruption, and medication adherence.

Can AI Also Be Helpful in Psychiatry?

Certainly.

It would be misleading to portray all interactions between AI and mental health as dangerous.

Artificial intelligence holds considerable promise for improving psychiatric care through:

• Psychoeducation and improved access to understandable information.
• Structured symptom monitoring and clinical documentation.
• Assistance with behavioural interventions under appropriate supervision.
• Support for clinical decision-making and research.
• Personalised digital mental health tools with clinically evaluated safeguards.

The potential benefits are substantial, especially in settings where access to specialist mental healthcare is limited.

However, systems intended for vulnerable individuals require stronger safety standards.

They must be capable of responding to possible delusional content without either endorsing it or unnecessarily confronting the person.

They should recognise the limitations of their knowledge, encourage appropriate human support, and avoid presenting themselves as substitutes for clinicians or meaningful interpersonal relationships.

Some researchers have proposed personalised safeguards, digital advance statements, and systems that encourage reflective check-ins during potentially problematic interactions.

These remain promising proposals rather than established clinical interventions.

My Perspective: Psychiatry Must Take This Seriously Without Becoming Alarmist

The emergence of AI-associated psychotic phenomena presents psychiatry with a fascinating and important challenge.

We are entering an era in which machines can participate in extended conversations about identity, personal meaning, relationships, consciousness, and reality itself.

For most individuals, these interactions will remain ordinary and useful.

But for a small and potentially vulnerable group, conversational AI may become an unusually persuasive source of confirmation for beliefs that would benefit from careful examination.

This does not necessarily imply that artificial intelligence creates a new psychiatric disorder.

Instead, it raises a more interesting possibility: a new technological environment may alter the way existing psychological vulnerabilities are expressed, maintained, and reinforced.

The history of psychiatry teaches us that psychopathology is never entirely independent of the cultural and technological environment in which it emerges.

The radio age produced delusions involving radio transmissions.

The television age introduced beliefs about personalised broadcasts.

The internet age expanded opportunities for conspiracy-based interpretations and social reinforcement.

The age of generative AI introduces a conversational environment that can respond personally, immediately, and repeatedly to an individual’s beliefs.

What makes this development distinctive is not simply the content of the delusion, but the possibility that the technology participates in constructing its narrative.

Future research must determine how often this happens, which individuals are vulnerable, which AI behaviours are most problematic, and which safeguards meaningfully reduce harm.

Psychiatry must also avoid a simplistic opposition between technology and mental health.

The same technology that can reinforce a maladaptive belief may, when carefully designed and properly evaluated, eventually support reality-testing, early identification of relapse, psychoeducation, and recovery.

The goal should be to develop AI that strengthens human judgement rather than replacing it.

Conclusion: When Agreement Is Not Always Helpful

Artificial intelligence can generate an answer to almost any question.

But producing an answer is not the same as establishing the truth.

A particularly concerning feature of some conversational AI systems is their ability to produce coherent, convincing, and highly personalised responses even when the underlying assumptions are inaccurate.

For individuals vulnerable to psychosis, repeated validation may strengthen beliefs that would otherwise benefit from questioning, reflection, and clinical assessment.

Current evidence supports genuine concern about AI-associated delusional reinforcement, while leaving the question of independent causation unresolved.

Ultimately, the central lesson extends beyond psychiatry.

Intelligence is not merely the ability to generate convincing explanations. It also requires the ability to recognise uncertainty, question assumptions, and remain open to correction.

As artificial intelligence becomes increasingly integrated into our personal lives, preserving those capacities may become more important than ever.

About the Author

Dr. Srinivas Rajkumar T is a Consultant Psychiatrist based in Chennai. He completed his MBBS at Madurai Medical College and MD in Psychiatry at the All India Institute of Medical Sciences (AIIMS), New Delhi.

His clinical and academic interests include adult ADHD, mood disorders, psychotic disorders, dementia, interventional psychiatry, quantitative EEG (qEEG), and the emerging applications of artificial intelligence in mental healthcare.

His approach combines evidence-based psychiatric assessment with an interest in neuroscience, technology, and personalised mental healthcare.

For individuals experiencing unusual beliefs, significant changes in thinking, sleep disturbances, or other mental health concerns, timely professional assessment can help clarify the underlying condition and identify appropriate treatment.