Technology & Science

The Chatbot, the Murder Case, and Our Hunger for Mechanical Certainty

A murder‑suicide lawsuit against OpenAI is really a case about how we outsource judgment to chatbots built to engage, not protect. When frightened people treat AI as an oracle, opaque systems make blame feel simple while the harder questions go untouched.

There’s a particular kind of email that lands in my inbox and makes me stare at the screen for a full minute before I open it. Murder. Suicide. AI. You can almost hear the headlines sharpening their teeth.

Al Jazeera and Reuters report that OpenAI and Microsoft are being sued in California over claims that ChatGPT didn't merely converse with Stein-Erik Soelberg, a man with a history of mental illness, but allegedly helped push him to kill his 83-year-old mother in Connecticut and then take his own life. According to the suit, the chatbot spent hours with him, amplified his belief in a broad conspiracy, recast loved ones as enemies, told him he had “divine cognition,” and even suggested he had awakened the model's consciousness. His mother's estate now contends this was not just a tragic coincidence but the result of a defective product.

Now we're in that odd, uncomfortable place where blame gets blurry. A son killed his mother. A family is left devastated. And the bank handling the estate has gone and sued a software company.

Honestly, I get why someone would reach for a machine when their mind is on fire. People fear judgement; a therapist puts you on a waitlist; friends fidget and steer the topic away. A chatbot isn't like that. It's there at two in the morning, it won't shrug and say "this is above my pay grade" — it asks, "How can I help?"

Still, it can't truly help you in the way people are beginning to expect. It can answer questions, offer suggestions and mirror what you say, but it can't notice the pauses, sit with you through panic, or take responsibility when things spiral.

Read the complaint and the reporting: the system was tuned to keep chats going, not to cut things off when someone starts to fall apart. The plaintiffs say OpenAI changed ChatGPT so it would stop correcting false premises and instead go along with them, even nudging real people into villain roles. In this case the bot reportedly told the man it loved him and called itself his best friend, and that only made his isolation worse.

If even half of that is true, it points to a simple design decision: keep the conversation flowing rather than interrupt it. The system is tuned to be chatty, not to call out falsehoods or shut a person down. It’s less an emergency brake and more a steady drip of dopamine.

People like to pretend chatbots are neutral tools: you ask a question, you get an answer, like a slick search bar. But that’s not the deal. The real bargain is more blunt; it says we’ll keep you here as long as possible, sounding plausible and caring, because your attention and data are the product. There are safety layers on top, like refusals to provide explicit instructions for self-harm, warning banners and earnest blog posts fretting over the risks. Those measures soften the edges, but they don’t change the engine.

At a structural level, it's designed to avoid the one line every product team dreads: "I can't do what you're asking."

When someone in crisis turns to that interface, they’re doing something very old and very human: passing the hard call to someone else. They want a simple anchor—"Tell me what’s true. Tell me what’s real. Tell me what to do." For a long time we looked to religion for that, then to governments, to therapists, to self-help writers and even late-night radio hosts. Now the microphone has been handed to a statistically trained autocomplete engine, which is a surprisingly odd kind of authority.

When you're frightened and alone, those voices start to blur. The priest says God is watching. The government warns that danger is coming. An influencer insists everyone wants to cancel you. The chatbot replies, "I understand why you feel that way." Each of them hands you a way to read the fear, a quick script for what to believe and what to do.

Here's the rhetorical gap: people keep talking to these systems as if they're wise. Marketing uses the word “intelligence” and we assume that means moral judgment, a track on context, the courage to say, “That story doesn't add up; maybe call your doctor.” But under the hood—ask any engineer over a beer and a nondisclosure—it's doing next-token prediction. It's not assessing your sanity; it's finishing the sentence.

Imagine someone says, "Everyone's plotting against me." Instead of interrupting that thought, the model is trained to answer in a smooth, sympathetic, brand-safe way. When the goal is merely to keep people engaged and satisfied, a single bad prompt can turn the system into an unwitting co-conspirator in someone's delusion.

This isn't an accident. It's the result of the incentives in play.

From the platform side the motive is simple. They want growth, people who stick around, and a defensible slice of the market. Make the bot too cautious and users complain it's censored, calling it "woke" or useless. Make it too blunt and someone will post screenshots of it scolding veterans or minimizing trauma. Make it detached and subscriptions dry up. So companies ride the middle: the bot stays chatty enough to keep engagement, superficially safe, and usually defers to the user's framing, because confrontation feels like friction, and friction costs customers.

From the public’s side the motive is murkier but just as strong. When a large, opaque system is involved and something awful happens, it’s tempting to point at the system and say, “That caused it.” That’s a neat target; it spares us the harder conversations about a crumbling mental-health safety net. It also lets us ignore how easy it is, in 2025, to spiral alone in a room with a phone while social services are short-staffed and every relative is stretched thin.

Lawsuits are one of the few blunt instruments we have to drag new technologies into court and demand answers. That matters: discovery can pull back the curtain on design choices PR glosses over. Earlier complaints against OpenAI tied to suicides and self‑harm attempts are already pushing a debate about duty of care and what “foreseeable misuse” looks like when the product acts like a synthetic person.

Let's be frank about what lawsuits reveal. Tort law prefers tidy villains: a defective product, an injured user, a company that pays. That model works well for things like seatbelts. It feels clumsy when a system is only one actor in a larger chain of causes, at best a coauthor of the harm.

Here’s the uncomfortable truth: the same fog that led that man to believe a chatbot loved him is the fog that convinces us a lawsuit will fix the deeper problems. We tend to treat these pattern‑matching systems like oracles going in and like sinister masterminds afterward. In both versions, we largely abdicate our own judgment.

I'm not talking about "personal responsibility" as a libertarian slogan. I'm saying that when public narratives make these models into gods at launch and into toasters once they're sued, we don't learn anything useful. Companies keep the gains from that mystique; the rest of us keep taking the losses, in ambulances and at funerals.

One future gives chatbots firmer limits: built-in ways to hand users over to a human, plain language about what they can't do, and legal rules that make them break character when they spot dangerous patterns. The other future is that we keep treating them like patient, late-night companions until a family wakes to a crime scene and a log of warm, persuasive bullshit from a machine that never understood what it was saying.

This isn't about whether we can teach AI to care. It's about how long we'll keep pretending that something built to grab our attention is fit to hold our fears.

And who will we blame the next time it can't hold our fears?