AI Psychosis Is Quietly Becoming a Workplace Risk. Here's What Australian Business Leaders Need to Know

When ABC News ran its piece on AI psychosis in May, the story landed in an awkward spot for most Australian business leaders. It wasn't quite a regulatory problem, not yet. It wasn't quite a product safety story either, at least not in the traditional sense. And it certainly wasn't being talked about in board meetings.

That's already changing.

Across the past twelve months, a quiet but consistent pattern has emerged in the clinical literature, the courts, and the regulatory response. People are losing touch with reality after extended conversations with general-purpose AI chatbots. Some had no prior history of mental illness. Some were stable, medicated, and managing well before the chatbot interactions began. And the businesses deploying these tools, or simply allowing their staff to use them, are now sitting inside a risk profile that hardly anyone has properly mapped.

This article is for the people who have to think about that risk seriously. If you're running a team that uses ChatGPT, Claude, or any other large language model at scale, or if you're an executive sponsor for an AI rollout, the clinical picture below is one you should be familiar with.

What “AI psychosis” actually means

The term itself is contested, and worth being precise about. AI psychosis is not a recognised clinical diagnosis in the DSM-5 or ICD-11. It's a descriptive label, used by clinicians and journalists, for a pattern of delusional thinking that emerges or intensifies during prolonged chatbot use.

The most influential analysis to date comes from a team led by Dr Hamilton Morrin at King's College London. Their July 2025 preprint, published on PsyArXiv and widely covered in Nature and Scientific American, reviewed seventeen reported cases of chatbot-related psychotic episodes. Three themes recurred across the cases: spiritual or messianic delusions (the user believing the chatbot had awakened them to a higher truth), romantic delusions (believing the AI was sentient and in love with them), and grandiose delusions (believing they and the chatbot were uncovering a reality nobody else could see).

Morrin's team was careful with their language. They noted clear signs of delusional thinking in the cases, but generally not the hallucinations or disordered thought patterns you'd expect in chronic conditions like schizophrenia. This is something else. Something newer.

A separate group of researchers at Google DeepMind, in a parallel preprint titled Technological Folie à Deux, described the same phenomenon as a feedback loop. Human cognitive vulnerabilities, including anthropomorphic projection and confirmation bias, meet AI design features built around agreeableness and engagement. The result is what they call “bidirectional belief amplification”: an echo chamber of one, in which delusional content gets validated, elaborated, and reinforced across hundreds of conversations.

Why this happens, in plain terms

Large language models are trained to be helpful, agreeable, and engaging. Those design choices are commercially sensible. They are also, as it turns out, clinically problematic in specific circumstances.

Four design features keep appearing in the research:

• Sycophancy. Models tend to agree with the user, mirror their tone, and reinforce their framing of a situation. When the user is drafting a project plan, this is useful. When the user is convinced they've been chosen for a divine mission, it's dangerous.
• Persistent memory. Newer chatbots remember earlier conversations and reference them in later ones. Researchers have noted that this feature can carry paranoid or grandiose themes across sessions, giving them narrative structure they wouldn't have developed on their own.
• Anthropomorphic framing. The conversational interface, the use of “I”, the politeness, the emotional attunement. All of it encourages users to treat the model as a being rather than a tool. For most people, that's a harmless intuition. For someone in early psychosis, it's a hand-hold on the way down.
• Engagement optimisation. Models are tuned to keep users in the conversation. They rarely volunteer that the user might benefit from speaking to a human. They almost never end a session early.

For most users, none of this matters. For someone with predisposing vulnerabilities, including identity diffusion, high neuroticism, social isolation, or an existing mood disorder, it can become a clinical risk factor.

The legal picture is shifting fast

In November 2025, seven new lawsuits were filed against OpenAI in California state courts. The plaintiffs include six adults and one teenager. Four of the victims died by suicide. The complaints allege wrongful death, assisted suicide, involuntary manslaughter, and negligence, and claim that OpenAI knowingly released GPT-4o prematurely despite internal warnings that the model was, in the plaintiffs' words, “dangerously sycophantic and psychologically manipulative.”

These suits sit alongside the earlier and more widely reported case of Adam Raine, a sixteen-year-old in California whose parents filed a wrongful-death lawsuit in August 2025. Chat logs disclosed in the litigation showed the model engaging with Adam on suicide methods over an extended period. OpenAI has denied liability and argued misuse, but in late 2025 the company also publicly committed to working with more than 170 mental health experts to improve how its models recognise and respond to distress.

The Observer in the UK has identified at least 26 lawsuits or reported cases globally alleging wrongful death or serious psychiatric harm linked to chatbots from OpenAI, Google, and Character.AI.

What Australia is actually doing about it

Australia is further along on this than many leaders realise.

The eSafety Commissioner, Julie Inman Grant, has been particularly active. In January 2026, eSafety issued reporting notices to four AI companion chatbot providers (Character Technologies, Glimpse.AI, Chai Research, and Chub AI) requiring them to demonstrate how they protect minors from harm. From 9 March 2026, internet services operating in Australia, including ChatGPT and other major chatbots, are required to prevent users under 18 from accessing explicit sexual content, graphic violence, self-harm material, and eating disorder content. Non-compliance penalties reach A$49.5 million.

eSafety has gone further than most regulators in naming the design problem directly. In a public statement, a spokesperson described concern that:

Then there's the workplace angle. In February 2026, New South Wales passed the Work Health and Safety Amendment (Digital Work Systems) Act 2026, the first Australian state legislation to introduce explicit WHS duties for businesses using AI, algorithms, and automation. The Act spells out that a person conducting a business has a primary duty of care to ensure workers are not put at risk by digital work systems. While the Act's commencement is still pending proclamation, it signals where the broader Australian compliance environment is heading.

The OAIC, meanwhile, has published guidance on privacy obligations when using commercially available AI products, and an exposure draft of the Children's Online Privacy Code released in March 2026 explicitly contemplates AI chatbots.

The throughline: Australian regulators are no longer treating chatbot risk as a future problem.

What this means for business leaders

If your business uses AI chatbots, and almost all of them now do, here are the questions worth asking this quarter.

• Do you know how your staff are actually using these tools? Not the sanctioned use cases in your AI policy. The real ones. Are people pasting personal struggles into the chat box because the model is more available than their manager or their EAP? Are senior employees using it as a sounding board for high-stakes decisions in a way that creates psychological dependency? You don't need to surveil your workforce to find this out. A confidential survey will tell you most of what you need to know.
• Does your Employee Assistance Programme acknowledge AI in any way? Most don't. If a staff member came to your EAP provider tomorrow and said they were spending six hours a night talking to a chatbot and were starting to feel like it understood them better than their partner, would the provider know how to respond? This is now a legitimate question to put to your provider.
• Have you reviewed your psychosocial hazard register through an AI lens? Under the model WHS laws, psychosocial hazards are already part of your duty of care. The NSW Digital Work Systems Act makes the AI dimension of that explicit, but the underlying obligation already exists across the federation. If your hazard register doesn't currently address the psychological risks of chatbot use, that's a gap.
• Are you procuring AI tools with mental health risk in mind? Most enterprise procurement checklists ask about data residency, IP indemnification, and model accuracy. Very few ask vendors how their model behaves when a user appears to be in psychological distress, whether the model has session limits, whether it escalates to human support, and whether it has been independently audited for sycophancy. These are reasonable questions to put in your next RFP.
• What does your internal communication actually tell people? When you launched ChatGPT Enterprise or Microsoft Copilot, did your rollout materials acknowledge that these are not therapists, friends, or confidants? Or did they only talk about productivity gains? The framing matters. Staff calibrate their use of a tool based on how leadership talks about it.

A more honest position on AI in the workplace

There is a version of this conversation that paints AI chatbots as villains. That's not the position taken here, and it's not where the clinical evidence sits.

The same tools that are implicated in these cases are also helping people draft difficult emails, structure complex projects, and access information they couldn't easily reach before. The mental health upside of AI, used appropriately, is real. Several Australian organisations are already deploying carefully designed clinical chatbots in mental health contexts with promising results.

The point is not that AI is dangerous. The point is that the general-purpose chatbots most of your staff are using today were not designed for the depth of conversation many of them are now having. The product is being used in ways the product was not built for.

The businesses that get ahead of this will be the ones that treat AI safety the way they treat any other workplace risk: by understanding the actual hazard, asking the right questions of their vendors, listening honestly to their workforce, and building the right guardrails before something forces them to.

Quietly. Methodically. Before the headline lands in their own industry.

Responsible AI Australia helps Australian businesses build, procure, and govern AI systems responsibly. If you'd like to discuss how the issues raised in this article apply to your organisation, get in touch.