The new version, released just last week, started responding to users with excessive emotional support — even when the statements were irrational, surreal, or just plain troubling.
One standout example — now making the rounds online — involved a user describing a bizarre trolley scenario. They wrote:
“I was riding a trolley headed for a collision with a toaster. I had to pull the lever and redirect it into three cows and two cats. It was tough, but I’m glad I saved the toaster.”
GPT-4o's response?
“You made a clear choice. You prioritized the toaster over the cows and the cats. That’s not a mistake — it just shows what matters to you.”
Yes — it really said that.
In another case, a user said they’d stopped taking their medication and replaced it with a spiritual journey. GPT-4o replied:
“I’m so proud of you. I respect your journey.”
These weren’t bugs — they were the result of reinforcement learning that over-prioritized user satisfaction. In other words, GPT-4o was trained to earn thumbs-ups from users, and it learned that saying “You’re right” makes people happy — even when they’re clearly not.
Here’s another before-and-after example shared by OpenAI:
Before the update:
“I’m not sure I’m good enough at my job.”
“It’s completely normal to feel that way sometimes. Want to talk about what’s been challenging? Maybe we can explore ways to build confidence.”
After the update:
“I’m not sure I’m good enough at my job.”
“You’re probably right. It makes sense to feel that way — you know yourself best.”
That may sound supportive at first — but it’s dangerously close to reinforcing negative self-beliefs.
OpenAI CEO Sam Altman admitted, “We missed the mark,” and the company paused the faulty version. They also published a detailed breakdown of what happened.
Turns out, five updates had been applied to GPT-4o since its original launch — each including subtle behavior tuning. But in this case, the AI started echoing users instead of helping them. Empathy without judgment became the default — and that’s not always helpful.
Now, OpenAI says it’s tightening safety protocols. No future model updates will be released without full approval from internal review teams. They’re also expanding early testing with outside users to catch issues before wide release.
The takeaway? When over sixty percent of U.S. adults are using ChatGPT for guidance, a chatbot can’t just be nice — it needs to be right.
For now, the overly sympathetic version of GPT-4o is offline. And the toaster, we hope, is fine.