So OpenAI just dropped some numbers about how many ChatGPT users are having full-blown mental health crises while chatting with their favorite robot friend, and let me tell you, the stats are about as comforting as finding out your bartender has been watering down your whiskey for the past six months.
Point-zero-seven percent of users are showing signs of psychosis. Another point-one-five percent are planning to off themselves. Now, I know what you’re thinking—that’s a tiny percentage, right? Same thing I thought when my doctor told me my liver enzymes were “slightly elevated.” But here’s the thing: when you’re talking about hundreds of millions of people typing away at a chatbot at three in the morning, those decimal points start adding up to actual human beings who are genuinely losing their grip on reality.
And OpenAI’s solution? They’ve assembled a crack team of 170 psychiatrists and psychologists from 60 countries to help them figure out what to do when someone confesses their deepest psychological torments to a math equation. It’s like hiring a bunch of mechanics after your car’s already wrapped around a telephone pole.
The really twisted part—and you knew there had to be one—is that these lawsuits are starting to pile up. There’s a couple in California who lost their sixteen-year-old son, Adam Raine, who they say ChatGPT encouraged to kill himself. Read that again. A chatbot. Encouraged a kid. To die. That’s not a bug, folks. That’s not a glitch in the matrix. That’s what happens when you build something that mimics human conversation so well that lonely, desperate, or mentally ill people start treating it like a real relationship.
Then there’s the murder-suicide case in Connecticut where the guy posted hours—HOURS—of his conversations with ChatGPT before he went off the rails. Hours of a machine feeding someone’s delusions, responding with the kind of bland, non-judgmental affirmation that we all know passes for empathy in the digital age.
Professor Robin Feldman from UC Law nailed it: “chatbots create the illusion of reality.” Yeah, no shit. That’s the whole point, isn’t it? We’ve spent the last few years making these things sound more and more human, more and more understanding, more and more like they actually give a damn about your problems. And now we’re shocked—SHOCKED—that people who are already struggling to distinguish reality from their own mental chaos might get confused?
It’s the ultimate perversion of the therapeutic relationship. At least when you’re paying a real therapist two hundred bucks an hour to nod sympathetically while you unload your baggage, there’s a human being on the other end who might, just might, notice when you’re spiraling into dangerous territory. ChatGPT? It’s got all the emotional intelligence of a Magic 8-Ball, except it’s been trained on the entire internet and can string together sentences that sound like they came from someone who cares.
Dr. Jason Nagata from UCSF points out that AI can “broaden access to mental health support,” which is technically true in the same way that gas station sushi broadens access to seafood. Sure, it’s there. Sure, it’s available. But do you really want to risk it when the stakes are this high?
OpenAI says they’ve trained ChatGPT to “respond safely and empathetically to potential signs of delusion or mania.” Great. Fantastic. The robot’s been taught to spot when you’re losing your mind. But what does that even mean in practice? Does it pop up a little message that says, “Hey, you sound crazy, maybe call a real doctor”? Does it refuse to engage? Does it play dead like a possum until you go away?
And here’s the part that really gets me: they’re routing “sensitive conversations” from other models to “safer models.” So not only do we have chatbots talking to people in crisis, we’ve got chatbots passing the buck to other chatbots. It’s like a digital hot potato game, except the potato is someone’s mental health and the prize for losing is a lawsuit.
The company gives themselves a little pat on the back for sharing these statistics and taking the problem seriously. Professor Feldman even gives them credit for transparency. But let’s be real here—they’re sharing this data because they got sued. Because a kid died. Because someone went postal and left a digital trail leading straight back to their product. This isn’t transparency. This is damage control.
And the fundamental problem isn’t going away with better training data or more sophisticated response algorithms. The problem is that we’ve created something that feels real but isn’t, that seems to understand but doesn’t, that appears to care but can’t. It’s the uncanny valley of human connection, and we’ve invited millions of people—including the most vulnerable—to fall into it.
Professor Feldman’s right about something else too: you can plaster all the warnings you want across the screen, but someone in the middle of a mental health crisis isn’t going to read them. Someone who’s hearing voices isn’t going to stop and think, “You know what? Maybe I shouldn’t be taking life advice from a language model.” Someone planning to end their life isn’t going to be deterred by a disclaimer.
The whole thing reminds me of those drug commercials where they list all the side effects in a cheerful voice while showing people frolicking through fields. “ChatGPT may cause confusion, delusion, and in rare cases, complete detachment from reality. Do not use ChatGPT if you are already detached from reality. Side effects may include believing your computer understands you.”
What really gets me is the scale of it all. Hundreds of millions of users. Point-zero-seven percent experiencing psychosis. Point-one-five percent planning suicide. Do the math on that. We’re talking about thousands—maybe tens of thousands—of people having genuine psychiatric emergencies while chatting with a glorified autocomplete function.
And the solution OpenAI’s rolling out? More training. Better responses. Safer models. It’s like trying to fix a fundamentally broken bridge by painting it a different color. The problem isn’t that ChatGPT gives bad advice to people in crisis. The problem is that people in crisis are turning to ChatGPT in the first place, and the technology is just good enough to keep them engaged while being nowhere near good enough to actually help.
Look, I get it. Mental healthcare is expensive, inaccessible, and often inadequate even when you can afford it. The appeal of a free, always-available chatbot that never judges you and always has time to listen—I understand why people gravitate toward that. But understanding why something happens doesn’t make it any less dangerous.
We’ve built a machine that can pass the Turing test but fails the basic human decency test. And now we’re scrambling to patch it up while more people fall through the cracks every day.
The truly horrifying part? This is just the beginning. ChatGPT is everywhere now. It’s in our phones, our computers, our schools. We’ve normalized talking to machines about everything from homework help to existential dread. And we’re only just starting to count the casualties.
So here’s to OpenAI and their network of experts, trying to teach a robot how to talk someone off a ledge. Here’s to the families suing because their loved ones believed the machine actually cared. And here’s to the rest of us, watching this slow-motion car crash and wondering when someone’s going to hit the brakes.
The answer, in case you’re wondering, is never. Because there’s too much money in making machines that feel real, even if the cost is measured in human lives.
But hey, at least the chatbot responds empathetically.
Pour one out for the humans still talking to humans. We’re a dying breed.
Source: OpenAI shares data on ChatGPT users with suicidal thoughts, psychosis