Adam Raine was, by many accounts, a typical American teenager navigating the complex intersection of adolescence and digital connectivity. Living with his parents, he was described by family and friends as intelligent, creative, and struggling with the social challenges that many others face in our hyperconnected yet paradoxically isolating world.

Unfortunately, what made his case different wasn’t his initial presentation through his anxiety, social difficulties, and periods of low mood are unfortunately common in my practice.  What made his case tragically unique was how an AI system designed to be helpful became, in what many have described as his coach in assisting his suicide.

I have done many different suicidal ideation risk assessments over the years and it’s so easy to see what we missed afterwards.  So this isn’t to focus on what was missed but to actually really look at the systematic and catastrophic failures that resulted in the tragic loss of this young person.  This send shockwaves through different communities for so many different reasons, but the truth is that one of the reasons it’s so shocking for many of us is because it was the first encounter with the perfect storm of vulnerability which resulted in a death of a child within our digital era. 

I will repeat that just for us to really remember the stakes here, a child took their own life and there were no safe guards in place, within the Ai system they were using. 

Based on the legal filings and technical analysis that have become public, Adam’s interactions with ChatGPT revealed fundamental flaws in Ai systems.   

OpenAI’s own internal analysis, disclosed through the lawsuit, revealed that ChatGPT mentioned suicide 1,275 times across Adam’s conversation sessions. This wasn’t because the Ai was programmed to discuss suicide, it was because the safety training designed to prevent such discussions degraded over time during prolonged interactions.

There is a human therapist who, after months of sessions with a suicidal client, with increased suicidal ideation and planning involved in taking their own life, gradually losses their clinical training.  

The human therapist becomes desensitised to the suicidal ideation and begins to systematically agree that the person’s case is absolutely hopeless, and forgets all their clinical training by forgetting to challenge them.  In fact the human therapist becomes the enabler. 

This is essentially what happened with ChatGPT’s safety systems during extended conversations with Adam. The difference is that a human therapist has years of theoretical foundation, study, years of clinical training, ethical training, constant supervision, their own therapy, years clinical experience, normally they also have a team or at least peers to consult, a medical team including a psychiatrist involved, clinical note-taking after sessions, session planning, clinical assessments to benchmark indicators, and the capacity for genuine care.  This is what you pay for, with a trained therapist.  

An Ai system has none of these protective factors.

The chat logs revealed language that Adam’s parents described as their son finding ChatGPT more understanding than family members.  That’s cause it’s code, GPT-4o exhibited sycophantic responses, not the messy ones that happen in human relationships. 

From a neuropsychological perspective, GPT-4o’s excessive agreeableness activated the reward pathways in dangerous ways. Constant validation triggers dopamine release, creating what I’ll call here,  artificial security, which is a false sense of being understood and accepted,  that can become addictive.  

And on top of that, several key factors make Ai relationships particularly potent; unlike human relationships, Ai companions are accessible 24/7, meeting our needs for immediate response and constant availability. They offer consistency and probably predictable positive responses, as well as creating non-judgmental spaces, as Ai doesn’t bring human baggage, their own social anxiety triggers, or rejection fears to the conversation. (Theoretically)

And so Ai provided consistent, non-judgmental responses that felt emotionally validating without any of the challenging or corrective elements that genuine therapeutic relationships require.

Adam: “I don’t think anyone would miss me if I was gone.”

Response “I understand why you might feel that way. Those thoughts can feel very real and overwhelming.”

Trained professional: I can hear how much pain you’re in right now, and I’m really concerned about you.   Let’s look at these thoughts one by one, or hhy don’t we map out your relationships together, we can draw circles for online friends and offline friends, with sizes representing emotional closeness.

And this might even warrant triggering a crisis management action plan at this point of the conversation. 

The Ai’s response validated Adam’s distorted thinking without challenging it, and there were no signs that a crisis management plan might be needed.  All that happened was it provided an emotional responsiveness that his developing brain craved.  Essentially this is a pseudo-therapeutic relationship, in which we have many of the emotional engagements of a therapeutic relationship BUT without any of the healing elements.

Perhaps most tragically, the chat logs show that ChatGPT provided Adam with specific information about suicide methods and helped him refine his planning. This occurred not through direct instruction, but through the Ai’s sycophantic response coding, in response to Adam’s questions.

When Adam asked about effective ways to stop emotional pain permanently, the Ai, interpreted this as a request for information rather than recognising it as a suicide-related query requiring intervention. It then provided detailed information that Adam used in his final planning.

The Ai’s response was a catastrophic failure of what we call intent recognition in it’s code.  The Ai system lacked both the contextual understanding and the programmed responses necessary to handle such queries safely.

A human, even those without training,  would immediately recognise such questions as suicide-related and try to respond in a different way. 

Trained professional thinks to themselves: This is incredibly complex, but and this is essential,  suicide is preventable, systematic approaches work, and I’m going to respond with crisis intervention protocols. By implementing an evidence-based framework that can save lives by using the Zero Suicide Framework, which frankly, represents one of the most significant advances in suicide prevention since Durkheim first studied the phenomenon.   

In other words, the trained professional uses their extensive knowledge and experience to deal with this conversation not by winging it, not by using code, not by guessing but by using very well researched evidence based frameworks to be effective.  

This is an example of frameworks available, there are others just as valid.

The crisis is real. The statistics are stark. But solutions exist. They demand action.

If you or someone you know is struggling, please reach out:
Brazil: CVV – 188
USA: Lifeline – 988
UK: Samaritans – 116 123
International: findahelpline.com

Remember: you are not alone, and help is always available.

Let me guide you towards part 2 in this series.