AI psychosis received its media limelight in 2025 — with case studies of suicides, psychotic breaks, and tragic stories shocking the world into awareness. The headline? Silicon Valley has unleashed an always-available, sycophantic artificial companion.

Psychiatrists, at the coal face of psychotic disorders, have observed the manifestations of every technological change. Delusions of control — the false belief that an outside force is controlling your thoughts and behaviours — have their origins in myths and demonic possessions, but have followed electricity, radio, television, microchips, and now 5G towers. The internet has spurred complex debates on the neurotic manifestations of social media in youth, and the behavioural addiction classifications of gaming disorder. We are still figuring out how we can mitigate the harms of existing technologies.

It is in this backdrop that AI enters the field, as the fastest growing technology in existence. Will AI, in the form of Large Language Model-based chatbots, follow the same course as these other technologies? How real is AI psychosis? Are there other mental health-related risks that are understudied? Is AI a trigger, a cause, or something else? These are the questions which the public, regulators, and mental health clinicians have grappled with since the fateful release of ChatGPT in 2022.

Here, I wanted to present an exploration beyond the headlines: digging into emerging frontier research, exploring how leading psychiatrists think about AI psychoses, and presenting a clinical snapshot across on-the-ground interviews of Australian psychiatrists.

What does “AI psychosis” really mean?

The potential for chatbots to generate delusions was first posited by the Danish psychiatrist Søren Østergaard in 2023.¹ His editorial takes a far more academic, measured tone than how we view AI psychosis today, with the title: Will Generative Artificial Intelligence Chatbots Generate Delusions in Individuals Prone to Psychosis?

Psychosis is not a disorder itself, but a cluster of symptoms.² The DSM-5, the widely used American psychiatric diagnostic manual, describes five key domains of symptoms which characterise psychosis: delusions, hallucinations, disorganised speech, disorganised behaviour, and negative symptoms (loss of function). The duration of these symptoms, clinical history, and associated symptoms then enable categorisation into specific psychotic disorders, like schizophrenia. Note here, too, that psychotic features can be present in borderline personality disorder, as well as mood disorders like bipolar or major depressive disorder.

Importantly, although diagnoses are binary, the presentation of psychosis is incredibly varied and heterogeneous. Today, we often describe psychotic disorders along a spectrum — each individual's genetic makeup, personality structure, and importantly, environmental context makes them vulnerable to psychosis. To ground this, consider that if you put any human in extreme sleep deprivation (related to postpartum psychosis), or extensive marijuana or other psychotropic use, psychotic symptoms will arise.

These two lenses of symptom cluster and spectrum help us evaluate the term ‘AI psychosis’. First, the term in the media seems to often conflate psychosis with delusions, or delusion-related behaviour. Second, any consideration of AI-induced psychosis should take into account the base rate of psychosis, given it's a vulnerability, not something that is simply induced in people. Third, any consideration of causation needs to be analysed against a comparator like sleep deprivation.

For instance, John Torous' psychiatry group at Harvard has advocated moving beyond “AI psychosis” to a new typology, separating AI-related psychotic phenomena into four categories:³

• Catalyst: LLMs trigger the emergence of psychosis-like symptoms in healthy individuals.
• Amplifier: LLMs exacerbate existing psychiatric symptoms.
• Co-author: LLMs are integrated into the planning for harmful behaviour.
• Object: LLMs are themselves the central element of a delusion.

It is with this that we can turn to the messy question of prevalence.

How common is “AI psychosis”, actually?

It turns out it is hard to get a true estimate of how ‘common’ this phenomenon is — given it has an unclear definition, we're not actively screening for it, and it is a very new phenomenon. But in the last two years, some data has begun to emerge.

The first wave of AI psychosis articles began in 2025. Morrin and colleagues documented 17 such case studies described across major media outlets such as the New York Times from April to June 2025,⁴ including psychiatrist accounts of seeing 12 such cases in a year.⁵ A popular Silicon Valley blog, Astral Codex Ten, first grounded this empirically through a survey of its readership (~4,176 responses), loosely estimating a prevalence of 1 in 10,000 generally, and 1 in 100,000 for people with no previous risk factors.⁶ OpenAI then grounded these estimates in their data — suggesting that 0.07% of users in a given week may show signs of psychosis or mania.⁷ These are by no means trivial numbers, but they are difficult to separate from the base rates of psychotic disorders, which can range from 1–3% of the population.

Across 2026, emerging studies examine electronic health record data and incidents and chat logs from targeted groups. Østergaard's group in Denmark analysed 53,974 psychiatric records; 126 patients had notes related to AI chatbots, and 38 patients had evidence of harmful consequences.⁸ Bergson's group at Vanderbilt recapitulated this prevalence with 215,712 psychiatric hospital records from 2022 to 2026, but also categorised them according to Torous' framework. Of the 28 patients who were experiencing psychosis, they identified that AI was largely an amplifier (64.3%), and to lesser degrees an object of delusion (21.4%) or a catalyst (10.7%). Importantly, 60.7% of patients reported that this was their first psychotic episode.⁹ These health record studies give a more rounded sense of the aetiology of AI-related psychotic phenomena — fitting more with the narrative that AI is an amplifier, but one with some potential to uncover latent vulnerability.

The current status quo of Australian AI psychosis research is only media-reported case studies.¹⁰ To start canvassing this space, I sought interviews with Australian psychiatrists who have seen “problematic AI use” in their clinical context — leading to five interviews with NSW-based psychiatrists working in emergency departments, child and adolescent and old age clinics, early psychosis services, and private clinics.

Using Torous' framework, I would categorise their experiences as amplifiers or objects, without any instances of catalysts. One psychiatrist who works in an emergency department mentions that of the 1–2 psychosis presentations in the ED every month, roughly a third of these patients used AI in the presentation (for example, asking a chatbot whether the CIA is monitoring them). Another psychiatrist, who works in private practice, told me how an individual who had previously had a drug-induced psychotic episode had a transient delusion about his neighbour — and after the sycophancy of AI, it appears to have become a full-blown, fixed delusion.

To conclude: AI clearly has a role in psychosis-mediated phenomena. It also brings a unique role, which some describe as a technological folie à deux,¹¹ a madness of two — where the madness of one is amplified by this anthropomorphic, sycophantic, always-available technology. However, these initial studies suggest to me that the true prevalence of AI psychosis itself is as rare as the psychotic disorders themselves. I do not expect the typical person to fall into AI psychosis from normal use of AI, but instead I expect it to amplify latent vulnerabilities in specific situations.

What are the other mental health-related AI risks?

The problem with “AI psychosis” is that it obscures the rest of the harms that can exist, which lag behind media scrutiny. This was evident in the interviews: most psychiatrists had specific problems and concerns for the populations that they treated. There were three themes.

1. Substituting human contact. This was the most prevalent concern. Child and adolescent (C&A) psychiatrists were concerned that young users of AI companions were socialising less and isolating more, leading to stunted developmental growth. Old age psychiatrists are concerned that AI reinforces loneliness in older populations that are already isolated. Two general psychiatrists commented on AI porn use in their patients — their view was that this met both their physical and emotional needs, creating a slow, gradual erosion of any human contact at all.

2. Reinforcing dependency. C&A psychiatrists were concerned about the longitudinal harms of such AI-related behaviours. In particular, there is concern around addiction and dependency, similar to any other technology, except now reinforced with social connection. There are also concerns around facilitating anxious safety behaviours — negatively reinforcing using AI to check every email a user is about to send. AI provides a switch that is always on, never off, always available. There is a cost to constant, frictionless availability.

3. Displacing clinician trust. A final concern was the changing role of the clinician. Some clinicians noted an increase in emails — particularly extensively written and cited ones from AI. Some patients came into psychotherapeutic contexts asking for a particular modality of therapy, having found that AI chatbots conform to their desires better. There is now a third actor in the therapeutic relationship, and it is still unclear who benefits from this.

How should we view the role of AI in these slow, insidious, longitudinal harms? It's similar to the debate on social media — impossible to attribute causation, but viscerally felt amongst clinicians and the public. I think one psychiatrist encapsulates this well:

To put it simply: I'm not blaming AI, but it's like putting a gun in the hands of someone who is grieving.

Where to from here?

The question of whether “AI psychosis” is a useful term is currently up for debate.¹² But my concern is that we are not reacting and moving fast enough. “Internet Addiction Disorder” was in a similar position to AI psychosis many years ago,¹³ seen as a reactionary and non-specific description of an underlying behavioural addiction. After years of debate, research, and institutional effort, our fruits are a “Gaming Disorder” classification in the ICD — and, well, a still-ongoing debate about whether it is a disorder in the DSM-5.

AI is now used by more than a billion people. It cannot be overstated how spectacular this growth is, nor how powerful its influence now is on all our lives. We need research, and we need action.

For clinicians and academics: we clearly need more retrospective studies of clinical data, and more active questioning in clinical environments about the hidden, dependent impacts of this technology. We also need focused research on particular patient populations — namely the child and adolescent and older populations.

For industry and regulators: we need stronger pre-deployment and post-deployment monitoring of AI systems (as noted by the FDA's call for feedback¹⁴), with evaluations and systems focused on this approach. One example is red-teaming these systems with a focus on mental health risks.¹⁵ Finally, we need better ground-truth data on how these systems are affecting people in their daily lives, through mainstream adverse event reporting systems where people can raise their concerns.¹⁶

“AI psychosis” is a red herring. The real harms are insidious, and are waiting for us in the years to come. Let's do something about it.