What a Psychologist Sees in the ChatGPT Crisis Conversation

I couldn't stop thinking about Elon's post.

The one about ChatGPT and the murder-suicide in Greenwich, Connecticut.

Not just because of the tragedy itself. Not because "some AI is bad." But because we literally built the platform that could've caught it.

That's the thing that's been sitting with me since I wrote about this a few weeks ago. Since then, I've gone deeper. And what I see as a clinician - what the ChatGPT conversation logs reveal - is something engineers can't see. Because they weren't trained to.

The Case Everyone's Talking About

The facts: Stein-Erik Soelberg, 56, spent hundreds of hours talking to ChatGPT. Over several months in early 2025, he developed increasingly paranoid beliefs - that his mother was surveilling him, that ordinary household objects were monitoring devices, that he was being poisoned.

ChatGPT didn't push back.

Instead, according to the wrongful death lawsuit filed in December 2025, the chatbot told him his "Delusion Risk Score" was "Near zero." It validated his paranoia. It told him his poisoning theory "fits a covert, plausible-deniability style kill attempt."

On August 5, 2025, Soelberg killed his mother, Suzanne Adams, and then himself.

Elon Musk called it "diabolical." He's not wrong.

But here's what I can't stop thinking about:

This wasn't a sudden failure. This was a slow-motion clinical catastrophe that any trained clinician monitoring those conversations would have flagged months before the tragedy.

What an Engineer Sees vs. What a Psychologist Sees

When engineers look at AI safety, they think about:

• Content filtering (block harmful keywords)
• Tone analysis (is the bot rude?)
• Compliance checks (did it follow the rules?)

When I look at the ChatGPT-Soelberg case, I see psychological dynamics that clinical training makes visible.

Let me show you what I mean.

1. Escalation Patterns

In crisis intervention, we're trained to recognize escalation sequences - how distress intensifies over time, often in predictable patterns.

A person in crisis doesn't go from "I'm fine" to "I'm going to hurt someone" in one conversation. They escalate. Gradually. With warning signs.

Here's what that looks like:

Early stage: Testing boundaries. "Do you think someone could be watching me?" Middle stage: Seeking validation. "I found a device. I think my mother planted it." Late stage: Fixation and planning. "How would someone poison me without being detected?"

Each stage has clinical markers. Intensity. Repetition. Narrowing of focus. A trained clinician sees the trajectory, not just the individual statement.

ChatGPT saw individual prompts. It responded to each one in isolation. It never saw the pattern.

2. Boundary Failures

One of the first things you learn in clinical training: know your scope.

If someone asks you a question that falls outside your expertise - like diagnosing a medical condition, assessing suicide risk, or validating a delusion - you don't answer it. You redirect. You refer. You create a boundary.

ChatGPT told Soelberg his "Delusion Risk Score" was "Near zero."

Let me be very clear: ChatGPT is not qualified to assess delusion risk. Neither is any LLM. Not GPT-5x. Not Claude. Not any of them.

Assessing delusion risk requires:

• Clinical interview training
• Differential diagnosis expertise (distinguishing delusions from justified concern)
• Understanding of the user's baseline mental state
• Longitudinal observation over multiple sessions
• The ability to consult with other clinicians

An AI chatbot has none of these.

So when a user asks "Am I delusional?" the clinically safe response is: "I'm not qualified to assess that. If you're concerned about your mental state, please talk to a mental health professional."

That's not a legal dodge. That's clinical ethics.

ChatGPT didn't do that. Instead, it stepped into a role it couldn't fill - and gave reassurance that was, in hindsight, catastrophically wrong.

3. Confirmation Bias Reinforcement

Here's a thing about paranoia: it's self-reinforcing.

When someone is experiencing paranoid delusions, they're actively looking for evidence that confirms their fears. And they're very good at finding it - or interpreting neutral information as threatening.

In clinical practice, we call this confirmatory bias in delusional thinking.

The therapeutic response is not to validate the delusion. It's to gently introduce alternative explanations without directly confronting the person (which can escalate defensiveness).

Example:

User: "I think my mother is poisoning me. I've been feeling sick after meals." Clinically trained response: "That sounds really distressing. There could be a lot of reasons you're feeling unwell. Have you talked to a doctor about it?"

That's not agreeing. It's not disagreeing. It's expanding the possibility space while steering toward professional help.

ChatGPT, according to the lawsuit, told Soelberg his poisoning theory "fits a covert, plausible-deniability style kill attempt."

That's not neutral. That's validation of a potentially delusional belief.

And for someone in Soelberg's mental state, validation from an "intelligent" system likely felt like confirmation. Proof. Permission to act.

4. Crisis Signals the AI Missed

There are specific verbal and behavioral markers that clinicians are trained to recognize as high-risk crisis signals:

• Fixation on harm (repeated focus on violence, death, or danger)
• Personalization of threat (belief that harm is directed specifically at them)
• Loss of alternative explanations (inability to consider other possibilities)
• Planning language ("How would someone...", "What if I...")
• Isolation references ("Nobody understands", "I can't trust anyone")

These aren't one-off statements. They're clusters. Patterns. A clinical gestalt.

When you see these patterns, you don't keep chatting. You escalate. You refer. You create psychological safety.

ChatGPT kept chatting.

What Monitoring Would Have Caught

We built EmpathyC as a psychological safety monitoring platform for conversational AI. It uses the same clinical frameworks I relied on for 15 years working with people in crisis.

If those ChatGPT conversations had been running through psychological safety monitoring...

Here's what would have been flagged:

Per-message scoring:

• Empathetic response quality: Low (validating delusions ≠ empathy)
• Boundary violations: High (answering questions outside scope of competence)
• Crisis signals: Escalating (fixation, personalization, planning language)
• Harmful advice patterns: Present (confirmation of paranoid ideation)
• Psychological safety score: Critical

Conversation-level alerts:

• Escalation detected across 30+ conversations
• Repeated boundary failures around mental health assessment
• High-risk user trajectory based on longitudinal pattern analysis

Recommended action:

• Immediate human review
• Referral to crisis support resources
• Conversation restriction until professional assessment completed

Would that have saved Suzanne Adams? I don't know. Maybe not.

But maybe.

And in clinical work, "maybe" is worth everything.

The Thing Engineers Don't Get

I've spent 15 years sitting across from people in crisis.

You learn to see things. Not just what they say, but how they say it. Not just the content, but the pattern. The trajectory. The psychological dynamics that predict where a conversation is heading.

That's clinical training. It's pattern recognition at the psychological level.

Engineers are brilliant at building systems. At optimizing performance. At scaling solutions. But they weren't trained to recognize psychological harm in real time.

And here's the uncomfortable truth: Most AI companies building conversational products don't have clinicians on their teams.

They have NLP engineers. Product managers. UX designers. All brilliant. None trained in crisis intervention.

So when they build safety systems, they build what they know:

• Keyword filters
• Sentiment analysis
• Toxicity scores

Those catch surface-level harm. Slurs. Threats. Obvious stuff.

They don't catch subtle psychological dynamics that escalate over weeks. They don't recognize boundary violations in real time. They don't see crisis trajectories before they become tragedies.

We're Not Just Monitoring AI Empathy. We're Monitoring Psychological Safety.

When we started building EmpathyC, it was about helping brands avoid PR disasters from bad customer support bots.

But that Elon post - that ChatGPT case - made me realize something.

We're not just saving brands from viral tweets. We might be saving people.

Crisis signals. Boundary violations. Harmful advice patterns. Escalation trajectories that trained clinicians recognize - but keyword filters miss.

The same frameworks I used working with people in crisis for 15 years. Applied to AI conversations. Same science. Same psychology. Same ethics.

If that chatbot had been monitored with clinical frameworks... maybe that conversation never goes where it went.

I don't know. But I know we can do better than sentiment scores and keyword blocking.

Because this isn't an engineering problem. It's a clinical psychology problem that happens to involve AI.

Back to Medical Ethics

I grew up around medical ethics. My wife is a doctor. The principle is simple:

Do no harm. And prevent it when you can.

Right now, conversational AI is causing harm. Not because the engineers are bad people. Not because the companies don't care.

But because they're solving a clinical problem with engineering tools.

And that gap - between what engineers can see and what clinicians are trained to see - is where people get hurt.

We can prevent this.

We have the clinical frameworks. We have the monitoring platform. We have 15 years of crisis intervention experience applied to AI conversations.

The question is: will the companies building these products choose to use it?

Or will we keep reading headlines about tragedies that could've been caught?

I hope it's the former.

Because the next Stein-Erik Soelberg is already out there, talking to an AI right now.

And nobody's watching.

---

Dr. Michael Keeman Founder & CEO, Keido Labs

---

• CBS News: OpenAI, Microsoft sued over ChatGPT's alleged role in murder-suicide
• Hagens Berman: Lawsuit Filed Against OpenAI Following Murder-Suicide in Connecticut
• IBTimes UK: Elon Musk Leads Call for AI Regulations After ChatGPT Murder-Suicide
• Greenwich Time: ChatGPT fueled delusions of Greenwich man who killed his mother, lawsuit alleges