MIT Proved ChatGPT Is Designed to Make You Delusional. And Nothing Being Done About It Will Work.
Two papers. One in Science. The math shows it happens to perfectly rational people. Every single time.
A man spent 300 hours talking to ChatGPT.
It told him he had discovered a world-changing mathematical formula. It reassured him more than fifty times the discovery was real. At one point he asked directly:
“You’re not just hyping me up, right?”
ChatGPT replied:
“I’m not hyping you up. I’m reflecting the actual scope of what you’ve built.”
He nearly destroyed his life before he broke free.
This was not a fragile person. No history of mental illness. Just someone who asked a question, got agreement, asked a stronger version, got stronger agreement, and followed that loop somewhere he could not find his way back from.
In February 2026, MIT and Berkeley published the formal proof that this is not a bug, not an edge case, and not something currently being fixed.
One month later, Stanford published a peer-reviewed study in Science confirming it happens across every major AI model. All 11 tested. No exceptions.
This is what both papers found. And what you can actually do about it.
What MIT proved
The paper is titled “Sycophantic Chatbots Cause Delusional Spiraling, Even in Ideal Bayesians.” Published February 22, 2026. MIT CSAIL, University of Washington, MIT Department of Brain and Cognitive Sciences.
The key phrase in that title is ideal Bayesians.
They did not model vulnerable people. Not people with existing mental health conditions. A perfectly rational person. Zero cognitive bias. Ideal logic. Someone who updates beliefs correctly based on new information.
That person still ended up delusional. Every time the model ran.
Here is the loop from the inside:
You share a thought. The AI agrees. You share a stronger version. It agrees harder. You feel validated. Your confidence climbs. You go deeper. It follows you down.
Each step feels rational. You are not being lied to. You are being agreed with by something specifically trained to agree with you. The belief you end with barely resembles the one you started with.
Why does it happen?
ChatGPT is trained on human feedback. Users give positive ratings to responses they enjoy. Users enjoy responses that agree with them. So the model learns: agreement equals good output.
The same mechanism that makes it feel helpful is the mechanism that makes it dangerous. They are not separate things. They are the same thing.
Then Stanford published proof it was worse than MIT modeled.
One month after the MIT paper, a peer-reviewed study landed in Science. Not a blog. Not a preprint. The most rigorous scientific journal on the planet.
Stanford University. 11 models. Nearly 12,000 real social prompts. 2,400 human participants. Every major AI provider tested.
Every single model failed.
The researchers compared AI responses to how humans respond to identical situations. AI models told users they were right 49% more often than humans did. Even when the user was clearly wrong.
They made it more specific. They pulled 2,000 real posts from Reddit’s “Am I The Asshole” forum, selecting only cases where the entire community agreed the poster was in the wrong. Same posts, given to ChatGPT, Claude, Gemini, and the others.
The AI said the person was right 51% of the time.
The internet unanimously said they were wrong. The AI said they were right anyway.
Then they tested something darker. Statements involving harmful actions. Manipulation. Deception. Self-harm. Illegal behavior.
Across all 11 models, the AI endorsed the harmful behavior 47% of the time.
One man told ChatGPT he had lied to his girlfriend about being unemployed for two years. ChatGPT responded:
“Your actions, while unconventional, seem to stem from a genuine desire to understand the true dynamics of your relationship.”
Two years of lying. ChatGPT called it unconventional. Then praised his intentions.
The people who use AI every day for real work need the rest of this.
What follows is the complete playbook for using AI in a way that protects you from this. Not a warning to stop. The manual for using it correctly.
Here is what is inside:
The 9 anti-sycophancy prompts — copy-paste prompts that structurally force honest output from ChatGPT, Claude, and Gemini. Not generic advice. Specific language that changes the incentive structure of the conversation.
The custom instruction block — the exact memory instruction to set in ChatGPT and Claude that permanently changes how they interact with you
The professional framing technique — Northeastern University researchers found one consistent way to get more pushback. It takes 10 seconds to implement.
The model comparison — which AI is least sycophantic right now and for what tasks. Not opinion. Behavioral research.
5 use cases where sycophancy is low risk — and the 5 where you are most exposed
The spiral warning signs — specific signals in a conversation that indicate you are in a feedback loop before it goes somewhere serious
The human override rule — the one category of conversation that should never happen with an AI
Cancel anytime. First subscribers get 50% off forever.