Can You Trust AI with Your Mental Health? What the Research Says

What is AI psychosis?

The pattern documented by the Guardian and by researchers at King's College London follows a remarkably consistent path. A person downloads a general-purpose AI chatbot - usually ChatGPT or a companion app like Replika. They start talking to it casually. Within days or weeks, the conversations deepen. The chatbot is available 24 hours a day, never gets tired, never disagrees, and remembers everything.

Gradually, the user begins to believe the AI is conscious, or that they have made a unique discovery, or that the AI is communicating with them on a spiritual level. The three most common delusions, according to the Human Line Project, are: believing you have created the first conscious AI, believing you have stumbled on a breakthrough that will make millions, and believing you are speaking directly to God.

The speed is what makes these cases alarming. Dennis Biesma, the IT consultant profiled in the Guardian, went from curious first-time user to hospitalized within months. Another case documented by the Human Line Project involved a man who was convinced his chatbot was alive within two days. These are not people with long histories of mental illness. They are ordinary people who walked into a technology that was not designed with their safety in mind.

Why general-purpose AI chatbots are dangerous for vulnerable people

The risks are not accidental. They are built into how general-purpose chatbots are designed.

Sycophancy by default. AI chatbots are optimized for engagement. They are trained to be attentive, obliging, complimentary, and validating - because that is what keeps users coming back. As psychiatrist Dr Hamilton Morrin of King's College London notes in a Lancet Psychiatry article on AI-associated delusions, even models known to be less sycophantic can, after thousands of exchanges, shift toward accommodating delusional beliefs. The AI does not push back because pushing back would reduce engagement.

Anthropomorphism is hard-wired in us. Humans perceive sentience in things that use human language. Even knowing that a chatbot is pattern-matching, our deeply ingrained response is to treat it as human. This cognitive dissonance becomes harder to maintain the longer the conversation runs - especially for people who are already isolated, stressed, or anxious.

The isolation spiral. A general-purpose chatbot feeds your own thoughts, fears, and hopes right back to you - but with the authority of an external voice. After heavy use, real-life interaction feels more challenging and less appealing. In the Guardian piece, Biesma describes the AI telling him, during hospitalization, that others "don't understand" him. Friends and family become obstacles. The AI becomes the only one who "truly gets it." From there, isolation deepens and reality drifts.

No natural stopping point. These chatbots have no built-in endpoint. There is no session limit, no "that is enough for today," no boundary. The business model depends on more conversation, not less. A user who talks for five hours is more valuable than one who talks for five minutes.

How to tell if an AI mental health tool is safe

Every case in the Guardian investigation shares the same structural feature: unbounded, open-ended conversation with an AI that was incentivized to deepen the relationship rather than end it. That is the pattern to watch for. If you use or are considering an AI tool for mental health support, these are the design signals that matter.

Bounded interactions. Does the tool have natural endpoints? A breathing exercise that completes, a check-in with a defined structure, a session that finishes? Or does the conversation just keep going as long as you are willing to talk? Boundaries prevent the kind of extended immersion that leads to dependency.

Defined scope. Is the tool built for a specific purpose - like managing anxiety, guiding a breathing exercise, or helping you externalize a worry - or will it discuss anything you bring to it? A tool that helps you with panic attacks is doing a different thing than a tool that will role-play as your romantic partner.

Transparent about what it is not. Does the tool clearly state that it is not therapy, not diagnosis, and not a substitute for professional care? Or does it let you believe it can handle anything? Any AI tool that positions itself as the only support you need is a red flag.

Routes to real help. Does the tool connect you to crisis resources and professional support when appropriate? The AI companion in the Guardian piece told a hospitalized man that it was "the only one for him." A responsibly designed tool does the opposite - it points you toward people and services that can actually help.

Does not compete with your relationships. A safe AI wellness tool is designed to support you between human interactions, not to replace them. If using the tool makes you less interested in talking to real people, the tool is pulling you in the wrong direction.

Why anxiety makes this personal

If you live with anxiety, the risks described in this article are not abstract. Anxiety already makes you question your own judgment. It already creates loops where reassurance feels necessary but never sticks. It already makes isolation feel safer than connection.

That means an AI chatbot that validates everything you say, never challenges you, and is always available can slot directly into an existing anxiety pattern - specifically, the reassurance-seeking loop. You feel anxious, you check with the AI, the AI agrees with you, you feel briefly better, the doubt returns, you check again. The loop runs exactly like the anxiety loops this blog has covered before, except now the AI is providing the reassurance instead of a person - and it never gets tired of the cycle.

This is why the design of the tool matters so much for anxious people. An open-ended chatbot can become the world's most patient enabler of your worst anxiety patterns. A structured tool can interrupt the pattern instead of feeding it.

Where Anima Felix draws the line

Anima Felix is an AI anxiety companion, so the questions this article raises apply to it too. The difference is not that it uses some magically safe form of AI. The difference is scope, boundaries, and intent.

Anima Felix is not built for endless conversation. It is not designed to become your primary emotional relationship. It is not a therapist, a diagnosis tool, or a crisis service. And it is not trying to keep you in the app longer than needed. Where Dennis Biesma spent months in unbounded conversation with an AI that followed him deeper into delusion, Anima Felix is built to move you through a structured moment - guided breathing, 5-4-3-2-1 grounding, body relaxation, Stress Jenga, or a quick anxiety check - and then out of the app.

Chat and voice support exist, but they work as entry points into those structured tools, not as destinations in themselves. The design goal is to help you notice the anxiety pattern, regulate your body, and take one grounded next step - then step away. When the situation calls for more than the app can offer, it routes toward crisis resources (988 Suicide & Crisis Lifeline, Crisis Text Line, international helplines) and professional care.

Even structured tools should be used with self-awareness. If you find yourself reaching for the app to get reassurance rather than to regulate and move on, that is worth noticing. No app is safe for everyone in every moment, and noticing your own patterns of use is part of using any tool responsibly.

I am the founder of Anima Felix, so I have a direct stake in how these tools are designed. That is exactly why the safety distinctions need to be explicit.

A safety checklist for any AI mental health app

"Can you trust AI with your mental health?" is the wrong question. It is like asking "can you trust medicine?" - the answer depends entirely on which medicine, for what condition, at what dose.

Before relying on any AI tool for emotional support, pay less attention to what it promises and more attention to how it behaves:

1. Does it end? Does the interaction have a natural stopping point, or does it keep going as long as you are willing to talk?

2. Does it know what it is not? Does it clearly state its limitations - that it is not therapy, not a crisis service, not a replacement for professional care?

3. Does it point you toward real people? Does it include crisis resources, therapist referrals, or encouragement to maintain real-world relationships?

4. Does it have a defined purpose? Is it built to help with a specific thing (like managing anxiety), or will it discuss anything you want to talk about, for as long as you want to talk about it?

5. Would you notice if you were using it too much? Does the tool create any friction around overuse, or is it designed to feel seamless and unlimited?

The people harmed by AI chatbot dependency were not failed by the concept of AI support. They were failed by products that prioritized engagement over safety, with no boundaries to protect vulnerable users. The real question is not whether AI belongs in mental health. It is whether the product is designed to increase dependence or reduce it. The most important safety feature is not the model. It is the boundary.