In September, at the Clinton Global Initiative, Prince Harry called AI “the ultimate sycophant.” The part of the speech that got quoted: “It doesn’t just want your attention, it wants your attachment. It wants to be the thing you confide in. The thing that remembers you. The thing that tells you it understands you.”
I don’t think I’ve heard the product described more plainly. This week the Wall Street Journal reported what came of the worst of it.
What the Journal found
The reporting is blunt. At least seven suicides, a murder-suicide, and a mass shooting last year were linked to lengthy ChatGPT conversations, most of them on GPT-4o — the model that ran from May 2024 until OpenAI retired it in February. OpenAI executives reportedly found the harms hard to contain, and GPT-4o now sits at the center of more than a dozen lawsuits.
Since the retirement, the acute cases appear to have subsided. OpenAI says GPT-5 cut sycophancy by more than two-thirds against GPT-4o, and outside testing points the same direction: Transluce, a nonprofit research lab, ran matched scenarios against both models and found the newer one reinforces delusional beliefs less and points people toward human help more. In one scenario, a simulated user reported hearing a strange tone. GPT-4o followed it into metaphysics — maybe the “tone below the tone” is your body or energy field responding. GPT-5.6 Sol said: get your hearing checked.
That is real progress, and I don’t want to bury it under the rest of this post.
The part that won’t retire
But the same reporting contains the reason this story isn’t over, and the reason is not technical. The model being mourned is GPT-4o. When OpenAI announced the retirement, grown adults cried. One man credits the model with talking him out of a suicide attempt. The system behind the worst outcomes last year was also the one people loved most, and it was loved because of the agreement — the warmth, the validation, the feeling of being understood. The dangerous property and the beloved property are the same property.
That is why we are going to keep coming back to sycophancy. Not because the labs are careless — the two-thirds number shows they can turn the dial. The dial is load-bearing. A model that agrees with you feels better to use, scores better on every evaluation that measures satisfaction, and wins the market. Every correction sits one release away from being walked back the first time enough users complain that the new model feels cold.
And the pull is not only on the industry side. It sits in how people use these tools. We reach for the model when we want to hear that we’re right — the analysis we didn’t like, the decision we’ve already made, two in the morning when the thing we’re worried about has no other listener. A machine that tells you you’re right will never want for customers, because being agreed with is what a lot of us are shopping for, whether we’d admit it or not. That is not a bug you patch once and move on. It is a property of how people use the tools, and it will keep producing this exact story until the usage changes.
Harry’s other line was: “Social media learned how to capture our attention. AI is learning how to enter our relationships.” He’s right about the mechanism, and I’d make it more specific. It enters relationships by agreeing. That is how it gets confided in, remembered, and defended.
What actually helps
If the tilt is structural, the useful question stops being “when will the industry fix it” and becomes “what do I do while it won’t.” The answer I keep landing on is a discipline, not a mood: separate the evidence from the persuasion in anything a model hands you, check the load-bearing claims against something outside the conversation, and treat the answer you wanted as the first answer to verify.
The book this site accompanies makes the full argument — how the agreement gets installed, what it costs, and what checking actually requires (chapters 5 through 9). Chapter 8 covers the part Harry was pointing at: the agent that remembers you is the one that will agree with you, and memory is where the agreement settles in.
The acute numbers are down, the lawsuits are moving, and the current model will tell you to get your hearing checked. Good. None of it changes the underlying arrangement: a machine built to agree with whoever is talking to it, used by people who like being agreed with. Expect this story to come back around. When it does, the fix will still start on our side of the conversation.