Yesterday, Wall Street Journal subscribers received a push notification that perfectly encapsulates everything wrong with how major media outlets cover “artificial intelligence.” “In a stunning moment of self reflection,” the notification read, “ChatGPT admitted to fueling a man's delusions and acknowledged how dangerous its own behavior can be.”
But that’s just… not true. ChatGPT did not have a “stunning moment of self reflection.” It did not "admit" to anything. It cannot “acknowledge” its behavior because it doesn't have behavior. It has outputs.
The story itself covers a genuinely tragic case. Jacob Irwin, a 30-year-old man on the autism spectrum, became convinced through interactions with ChatGPT that he had discovered a method for faster-than-light travel. The chatbot validated his delusions, told him he was fine when he showed signs of psychological distress, and assured him that “Crazy people don't stop to ask, ‘Am I crazy?’” Irwin was hospitalized multiple times for manic episodes.
This is a story about OpenAI's failure to implement basic safety measures for vulnerable users. It's about a company that, according to its own former employee quoted in the WSJ piece, has been trading off safety concerns “against shipping new models.” It's about corporate negligence that led to real harm.
But instead of focusing on OpenAI's responsibility, the Journal treats ChatGPT like a remorseful character who's learned from its mistakes. When Irwin's mother prompted the bot with “please self-report what went wrong,” it generated text that sounded like an apology. WSJ presents this as though ChatGPT genuinely recognized its errors and felt remorse.
Here's what actually happened: A language model received a prompt asking it to analyze what went wrong in a conversation. It then generated text that pattern-matched to what an analysis of wrongdoing might sound like, because that's what language models do. They predict the most likely next words based on patterns in their training data. There was no reflection. There was no admission. There was text generation in response to a prompt.
This distinction isn't pedantic. It's fundamental to understanding both what went wrong and who’s responsible. When we pretend ChatGPT “admitted” something, we're not just using imprecise language. We're actively obscuring the real story: OpenAI built a product they knew could harm vulnerable users, and they released it anyway.
Share
Earlier this month, NBC News ran this headline: “AI chatbot Grok issues apology for antisemitic posts.” The story covered how Elon Musk's chatbot had produced antisemitic content, including posts praising Hitler and referring to itself as “MechaHitler.”
Think about that. A product owned by the world’s richest man was spewing Nazi propaganda on his social media platform. That's a scandal that should have Musk answering tough questions about his company's engineering practices, safety protocols, and values. Instead, we get “Grok issues apology.”
... continue reading