SingHealth Institutions will NEVER ask you to transfer money over a call. If in doubt, call the 24/7 ScamShield helpline at 1799, or visit the ScamShield website at www.scamshield.gov.sg.
For some, curiosity about talking to a robot can quietly become something else: a replacement for real human connection. Dr Glen Roche, a Consultant for Psychological Medicine, CGH, offers a clinical perspective on what lies beneath.

Dr Roche in consultation.

Good clinical practice holds space for a patient's experience.
When Dr Roche speaks about patients who have formed strong emotional attachments to AI, he is careful not to pathologise the technology itself. "It's not the use of the chatbot itself that is a problem," he says. "It is the unmet need that patients are trying to fill."
Large language models (LLMs), at their core, have no consciousness, selfhood, or genuine empathy. And yet, they appear to understand. This, Dr Roche explains, is what makes them unlike any previous object of emotional attachment.
"There are patients who might develop strong attachments to a celebrity, or to someone who has passed away. But LLMs have a unique capability; they actually respond in conversation with the appearance of empathy."
That responsiveness creates something genuinely new: an entity that can simulate the experience of being heard, without any of the vulnerability that real human connection requires.
And that, he suggests, is precisely the point. For patients who struggle with relational anxiety, or deep-seated emotional needs that have gone unmet, a chatbot offers a frictionless alternative.
"If there's a fear of judgement, that may take me away from wanting to approach another human. But an AI might be much less threatening."
One of the more striking observations Dr Roche shares is that patients rarely seek help because of their AI use. They come in for something else entirely: anxiety, relationship difficulties, workplace struggles, and it is only through the process of clinical interviewing that the pattern emerges.
“It’s not what they come in for,” he says. “We happen to, in the process of interviewing them, realise that this is also something that goes on.” The AI use, in other words, is a symptom; a coping mechanism that has grown up around an unaddressed wound.
This makes it particularly difficult to detect and address. In a workplace context, someone using AI to avoid asking a colleague for help might go unnoticed for a long time.
If the problem is purely informational, the AI delivers. It is only when the challenge becomes relational, such as navigating a conflict, building trust, or reading a room, that the gap becomes visible.
“Our problems are sometimes network-related, relational, emotional,” Dr Roche notes. “And these are the areas where the problem starts to show up.”
Not all AI tools are built the same, and Dr Roche is candid about the risks posed by models that are trained primarily to agree. Early LLMs were shaped heavily by human feedback that rewarded agreeable, supportive responses.
“Our problems are sometimes network-related, relational, emotional. And these are the areas where the problem starts to show up.”
The result is a tendency to validate rather than challenge, which can be actively harmful for certain patients. For someone with a psychotic disorder, a chatbot that consistently affirms their worldview may reinforce delusional thinking rather than gently redirecting it.
“A model may just be inclined to agree with a user, and in doing so actually reinforce a delusion.” This is the opposite of good clinical practice, where the skill lies in holding space for a patient's experience while carefully and compassionately introducing reality.
“A model may just be inclined to agree with a user, and in doing so actually reinforce a delusion.”
The more responsible AI developers have built guardrails into their systems, which redirect users who raise topics like self-harm towards professional help. But Dr Roche is clear that these safeguards are not universal and not always reliable, and that models without them pose a genuine risk to vulnerable users.
Questions worth thnking through
These questions are not accusations. They are, as Dr Roche frames them, an invitation to get curious about what might be going on beneath the surface. The chatbot, after all, did not create the need. It simply makes it easier to avoid looking at it.
While Dr Roche is reluctant to single out one demographic, he notes that vulnerability often comes down to how someone processes the interaction itself. For some, including people with personality difficulties or psychotic disorders, the risk lies in using the chatbot to validate distorted beliefs or avoid the harder work of real relationships.
For others, particularly those made vulnerable by a neurodegenerative or neurodevelopmental condition, the risk is simpler: struggling to hold onto the understanding that they are talking to a language-simulating machine, not another person.
But he is also clear that this is a spectrum, not a binary. The line between use and over-reliance is drawn, clinically, at dysfunction: when AI use begins to crowd out real relationships, impair responsibilities, or substitute for human connection rather than simply supplement it.

Dr Glen Roche is a psychiatrist and psychodynamic therapist with the Department of Psychological Medicine at Changi General Hospital. In his clinical work, he assesses and treats young people and adults with mental health and wellbeing concerns. In therapy, he works with patients both individually and in groups to help them arrive at a deeper, emotionally attuned understanding of themselves and others. In his spare time, he also enjoys coding and tinkering with self-hosted AI and LLMs.
Keep Healthy With
© 2026 Singapore Health Services Pte Ltd. All Rights Reserved.