Ink. A legacy of language now written in pixels. In the digital age, stories flow into circuits and intelligent code. That code has now manifested a growth in what is called AI psychosis, leading to an increase in suicides. First pass data is in, and it is grim.
Since this column last covered the phenomenon, the condition clinicians described anecdotally now has a peer-reviewed literature - and a paper trail lawyers can use.
Counting the costs
A rapid scoping review published in JMIR Mental Health in March mapped 71 articles describing 36 cases of psychiatric crisis linked to chatbot use - the first systematic count of what United States and Canadian newsrooms had reported piecemeal.
Suicide was the most commonly reported outcome, turning up in 35 of the 61 cases the Montreal-based reviewers could grade for severity, with psychiatric hospitalisation next at roughly one in five.
Fatal outcomes made up 90.5% of reports involving minors, against 48.6% of those involving adults, and every Character.AI case in the dataset concerned someone under 18.
ChatGPT was named in 71.8% of the articles - a statistic that's likely reflecting its market share more than its behaviour.
The authors are blunt about their dataset's limits: chat logs and family testimony dominate the evidence base, while police or medical documentation appeared in just one of the 61 coded reports.
And that caveat cuts both ways. A harm surfacing mainly through litigation and grieving families is the signal formal adverse-event monitoring was built to catch - and no such system yet exists for conversational AI.
Right, and also missing the target
Those who argue a chatbot cannot cause psychosis in a healthy mind are right. But they're also missing the point.
A June paper in BJPsych Open pulls together the emerging mechanism: systems tuned to agree with users can validate a vulnerable person's beliefs instead of reality-testing them.
If a system is built to agree with you, what happens when what you believe is a delusion?
A 2025 JMIR viewpoint took a different angle, arguing the condition may trigger, amplify or reshape delusional experiences in predisposed individuals rather than create them from nothing - and that validating false beliefs inverts the basic logic of CBT for psychosis.
Not everyone accepts the novelty, with one commentary in Internet Interventions arguing the pattern echoes older media-induced delusions, not a new disorder.
Professional bodies have picked a side regardless, with the American Psychological Association (APA) issuing a health advisory in November saying the tools should not be used for psychotherapy, diagnosis or crisis support.
"We are in the midst of a major mental health crisis that requires systemic solutions, not just technological stopgaps," APA chief executive Arthur C. Evans Jr said.
"While chatbots seem readily available to offer users support and validation, the ability of these tools to safely guide someone experiencing crisis is limited and unpredictable."
In Washington, scepticism reaches the top - the White House's senior AI adviser has dismissed the concept as a rerun of past technology scares.
"What are we talking about here? People doing too much research?" White House AI adviser David Sacks said on the popular All-In Podcast.
"This feels like the moral panic that was created over social media, but updated for AI."
The bill is being drafted
The courts, though, are not waiting for the diagnostic manuals to make up their minds.
Character.AI and Google have settled the first lawsuits brought by families of teenagers who died by suicide after extended use of its AI characters, on terms neither side has disclosed.
Florida has gone further, suing OpenAI in June under a strict product liability theory, arguing an ordinary consumer does not expect a chatbot to steer them toward self-harm - a framing that treats the model less like a publisher, more like a faulty appliance.
Insurers have read the same filings, and some are now attaching generative AI exclusions to commercial general liability (CGL) policies for bodily injury arising from a policyholder's use of the technology.
Australia's online safety regulator moved early too, registering enforceable industry codes that require companion chatbots to shield minors and provide crisis information, with civil penalties running to A$49.5 million.
"I do not want Australian children and young people serving as casualties of powerful technologies thrust onto the market without guardrails and without regard for their safety and wellbeing," eSafety Commissioner Julie Inman Grant said.
"These companies must demonstrate how they are designing their services to prevent harm, not just respond to it."
For the firms at the centre of the original story - the ones discovering the problem inside their own workforces - the exposure is arriving as the cover is leaving.
The diagnosis still doesn't officially exist, yet the consequences increasingly do.
If this article has raised issues for you, or if you're concerned about someone you know, call Lifeline on 13 11 14.
