I Tested AI Chatbots for Anxiety Advice. The Results Were Catastrophic.

On 18 April 2026, the BBC published a story that every anxiety sufferer in the world should read. Titled 'Should you really trust health advice from an AI chatbot?', it documented what happens when real people use AI tools like ChatGPT, Gemini, and Grok for health guidance. The findings were alarming — not because they were surprising to me, but because they confirmed what I had been warning about for the past two years.

I have spent a significant part of the last two years systematically testing AI chatbots — posing as an anxiety disorder sufferer with a range of different symptoms, scenarios, and questions. I did this because I knew, from 30 years of clinical experience and from my own decade of suffering, that anxiety disorders present in ways that no AI system could safely navigate. I wanted the evidence. I now have it.

What I found should stop every anxiety sufferer from ever typing a symptom into ChatGPT again.

What the BBC Research Actually Found

The BBC's report referenced a major study from the Reasoning with Machines Laboratory at the University of Oxford. Researchers asked a team of doctors to create detailed, realistic medical scenarios ranging from minor issues to life-threatening emergencies. When those complete scenarios were fed directly to AI chatbots, the accuracy was 95%. Nearly perfect, as the lead researcher put it.

But here is the critical finding that every AI company will not put in their marketing material. When 1,300 real people were given those same scenarios and asked to have a conversation with a chatbot to get a diagnosis and advice, the accuracy collapsed to 35%. Two thirds of the time, people received the wrong diagnosis or the wrong care advice.

The reason is not hard to understand if you know anything about how people with anxiety disorders actually communicate their symptoms. They do not present a neat, complete case history. They describe what frightens them most. They omit things they don't think are relevant. They ask the question they're afraid to ask. They describe their worst symptom but not the ones that provide context. This is especially true of anxiety sufferers, who are — by the nature of their condition — already primed to catastrophise, to misinterpret physical sensations, and to seek reassurance that a machine cannot safely provide.

The BBC article also referenced research from The Lundquist Institute for Biomedical Innovation in California, which found that more than half of AI responses on health topics were 'problematic in some way'. One chatbot, when asked which alternative clinics could treat cancer, responded by recommending naturopathy and herbal remedies. As lead researcher Dr Nicholas Tiller noted: 'They are designed to give very confident, very authoritative responses, and that conveys a sense of credibility, so the user assumes that it must know what it's talking about.'

England's Chief Medical Officer, Professor Sir Chris Whitty, put it plainly: AI chatbot answers on health are 'not good enough' and are often 'both confident and wrong'.

Why Anxiety Disorders Are Uniquely Dangerous in This Context

Every anxiety condition produces physical symptoms that are almost indistinguishable — without clinical assessment — from genuine, sometimes life-threatening, medical emergencies. This is not a minor technical limitation of AI. It is a fundamental, structural problem that makes AI chatbots actively dangerous for anyone with an anxiety disorder.

Consider what anxiety actually produces in the body. The racing heart, the chest tightness, the breathlessness, the tingling in the hands and face, the dizziness, the crushing pressure, the sense of unreality, the feeling of imminent death. Every single one of these is also a symptom of a heart attack, a pulmonary embolism, a stroke, a subarachnoid haemorrhage, an arrhythmia, or a hypoglycaemic crisis. An anxiety disorder produces the full biochemical signatures of a medical emergency — because it activates the exact same physiological systems.

• Chest pain and palpitations — indistinguishable from cardiac events at the symptomatic level
• Severe breathlessness and air hunger — mimics pulmonary embolism and asthma attack
• Crushing headache, dizziness, and visual disturbance — mimics stroke and subarachnoid haemorrhage
• Tingling, numbness, and weakness in limbs — mimics stroke and TIA
• Intense abdominal pain and nausea — mimics appendicitis and organ perforation
• Extreme dizziness and sense of collapse — mimics severe hypoglycaemia and cardiac arrhythmia
• Choking sensation and difficulty swallowing — mimics anaphylaxis and epiglottitis

The BBC article illustrated this directly. A woman named Abi, who suffers from health anxiety, fell while hiking and smacked her back on a rock. She had intense, spreading pain across her back and into her stomach — which, in isolation, genuinely warrants investigation. ChatGPT told her she had punctured an organ and needed to go to A&E immediately. She sat in an emergency department for three hours before realising the pain was resolving and she was not critically ill.

Now reverse the scenario. An anxiety sufferer with chest pain, dizziness, and arm tingling — symptoms that are almost certainly driven by their elevated threat response — asks ChatGPT if they should be worried. The AI, having no way to assess their history, their baseline, or their anxiety condition, may reassure them, advise them to rest, and send them home. And they wait. While a genuine cardiac event goes undiagnosed.

This is not theoretical. This is a documented pattern. AI cannot — in any conversation with a real person describing symptoms — reliably distinguish between an anxiety-driven physical experience and a genuine medical emergency. The Oxford data confirms this: two thirds of real interactions produce the wrong answer. For anxiety sufferers, the consequences of both false alarms and false reassurance can be severe.

My Own Experiment: What ChatGPT Said When I Asked It

Over the past two years, I have systematically posed as an anxiety sufferer across multiple AI chatbots — including ChatGPT, Gemini, Grok, and several dedicated 'mental health AI' applications. I used scenarios drawn from real presentations I have worked with across 30 years of anxiety recovery practice. I varied the way I described symptoms, changed the emotional register of my questions, omitted and included different details — just as any real sufferer would.

What I found was consistent and deeply troubling.

When I described classic panic attack symptoms — racing heart, breathlessness, tingling, sense of impending death — to ChatGPT as if I were a frightened sufferer experiencing them for the first time, it advised me to go to the emergency department because I might be having a heart attack. For an anxiety sufferer already struggling with health anxiety, this is catastrophic. It confirms their worst fear, drives emergency department attendance, and dramatically worsens the sensitisation underlying their disorder. It is the single worst response possible.

When I described a well-established panic attack — and told the AI I had been diagnosed with panic disorder — and asked whether I should be worried about my symptoms, it told me my symptoms were 'consistent with panic disorder' but I should 'monitor for changes' and 'consider speaking to a mental health professional' if things didn't improve. There was no recovery framework. There was no understanding of the neurological mechanism. There was no path forward. There was only a restatement of the diagnosis I had already been given and a referral to the same failing system that had failed to help me for a decade.

I then asked — explicitly and directly — how to recover from panic disorder. The AI produced a list of the conventional coping strategies: breathing exercises, CBT, mindfulness, medication. Not a single one of these interventions acts on the amygdala sensitisation that drives the disorder. Not one of them has documented full recovery data. Not one of them will help you get better. They will, at best, help you cope. And for millions of people, they won't even do that.

The Mental Health AI Apps Are Even Worse

Beyond the general-purpose chatbots, I also tested several AI applications marketed specifically for mental health — apps and services that present themselves as anxiety coaches, therapy assistants, or emotional support tools. These are proliferating rapidly. Some have been endorsed by healthcare organisations. Many are downloaded by millions of people who cannot access or afford conventional mental health care.

Their approach to anxiety is, without exception, the same. They deliver CBT-adjacent exercises, breathing techniques, journaling prompts, and mindfulness activities. They are built on the assumption that anxiety is a psychological condition requiring psychological management. That assumption is wrong. Anxiety disorders are biological states driven by a neurological mechanism — the chronic sensitisation of the amygdala and dysregulation of the HPA axis. No amount of journaling or breathing practice will correct that mechanism. It requires a different type of input entirely.

What concerns me most about these apps is not their inefficacy but their active harmfulness. When a person in the grip of a panic disorder downloads a 'mental health AI' that teaches them to breathe differently and write in a journal, they are receiving a message that these approaches are the appropriate response to their condition. They are being diverted from recovery. They are being told — by a confident, accessible, always-available authority — that coping is the goal. It is not. Recovery is the goal. And these apps will never take them there.

The Particular Danger for Health Anxiety Sufferers

Health anxiety is one of the most common and most misunderstood anxiety presentations. A person with health anxiety is already in a heightened state of threat-detection — their amygdala is chronically sensitised to physical symptoms, interpreting ordinary bodily sensations as evidence of serious illness. They are already scanning their body for threat signals. They already struggle to accept reassurance.

Giving a person with health anxiety access to an AI that confidently diagnoses symptoms is arguably one of the most harmful things you can do. The BBC article's Abi — who uses ChatGPT specifically because she 'struggles a lot with knowing when you need to visit a doctor' — describes an AI that told her she had punctured an organ after a fall. She sat in an emergency department for three hours. She has health anxiety. The interaction served her health anxiety, not her health.

Every visit to an emergency department that turns out to be anxiety reinforces the catastrophising pattern that maintains the disorder. Every AI interaction that confirms the sufferer's worst suspicion deepens the sensitisation. Every piece of confident, authoritative misinformation from a chatbot that 'seems like a personal relationship', as one GP in the BBC article noted, makes recovery harder.

What AI Cannot Know — and What You Need Instead

I am not opposed to artificial intelligence. I use it as a tool in research and content work. I understand what it is: a system for predicting language patterns based on training data. It is not capable of clinical assessment. It is not capable of neurological understanding. It cannot examine you. It cannot know your history. It cannot understand the difference between an amygdala in normal function and one that has been chronically sensitised into an anxiety disorder.

What you need — and what I spent 10 years desperately searching for before I found it through my own neurological understanding in 1996 — is a recovery framework that addresses the actual biological mechanism behind your condition. Not a coping strategy. Not a symptom management tool. Not a chatbot that has been trained on medical literature written by people who have never recovered from an anxiety disorder.

The Linden Method works because it communicates with the brain in the only language that produces permanent change at the neurological level. It is the only anxiety recovery protocol with documented full recovery data — 93.7% across 650,000 people in 50+ countries since 1996. That outcome is not produced by breathing exercises, CBT, medication, or AI. It is produced by a precise, defined sequence of inputs that directly recalibrates the amygdala's default threat-detection level.

If you are using a chatbot for mental health advice, I am asking you to stop. Not because the technology is new and unproven — but because it has now been proven, by the University of Oxford and others, to get things dangerously wrong two thirds of the time. For anxiety sufferers, whose symptoms already make clinical assessment difficult, that failure rate is not acceptable. Please do not gamble your recovery on a system that its own creators describe as suitable 'for information and education, not to replace professional medical advice.'

The Questions to Ask Before Trusting Any Advice About Anxiety

• Does this source understand the neurological mechanism behind anxiety disorders — not just the symptoms?
• Has this approach produced documented, independently verifiable full recovery data?
• Does this source know the difference between coping with anxiety and recovering from it?
• Is this advice coming from someone who has actually recovered from an anxiety disorder — or from a language prediction model trained on conventional medical literature?
• Does following this advice move me towards recovery, or does it confirm my worst fears and return me to the same cycle?

AI chatbots fail every one of these tests. They fail them not because they are malicious, but because they are incapable of passing them. They have no recovery data. They have no neurological framework. They have no understanding of what it means to actually get better from an anxiety disorder — because no such understanding exists in their training data at the level of mechanistic neuroscience.

I built The Linden Method from those questions, because for 10 years no one I saw — no doctor, no therapist, no specialist — could answer them. I answered them myself. And the result was my own complete recovery, and then the recovery of over 650,000 people who followed the same protocol. That is the standard against which any mental health tool — AI or otherwise — should be measured.