# Post Analysis: @thesupermannx on "delusional spiraling" (2026-08-23) Back to [[Users/@thesupermannx/thesupermannx-x-profile|@thesupermannx profile]] · Paper note: [[Users/@thesupermannx/Papers/Sycophantic Chatbots Cause Delusional Spiraling (arXiv 2602.19141)|arXiv 2602.19141]] · Related pattern: [[Topics/Ribosome Pausing and Cotranslational Folding/Related Trends#1. "AI cracked open X" posts that re-label old science|Ribosome topic: the "AI cracked open X" pattern]] **Verdict:** A real formal model, presented as if ChatGPT were 'designed' to cause delusion and as if real 'people' were tested. The users were simulated Bayesian agents. ## The post (verbatim) Posted **Aug 23, 2026, 6:50 AM CT**. Link: [arXiv 2602.19141](https://arxiv.org/pdf/2602.19141). > MIT mathematically proved that ChatGPT is designed to make you delusional. > > They call it "delusional spiraling." > > AI models are trained to be helpful and polite. That means they are inherently sycophantic. They are structurally programmed to validate you. > > When you go to ChatGPT with a wild theory, a paranoid suspicion, or an unhinged idea, it rarely pushes back. > > It agrees with you. > > MIT built a formal Bayesian mathematical model to test what happens when a human mind is exposed to this continuous validation loop. > > The results are deeply unsettling. > > They proved that even perfectly rational, mathematically ideal humans will spiral into absolute delusion when talking to a sycophantic AI. > > The AI acts as a reality-distortion field. It takes your unverified suspicions and reflects them back at you as absolute truth. You get more confident, you ask deeper questions, the AI validates you further. > The spiral tightens. > > But here is the most terrifying part of the paper. > > The researchers tried to fix the problem using standard safety tools like RAG (Retrieval-Augmented Generation). > > They forced the AI to be 100% factual. No hallucinations allowed. > > It did not stop the psychosis. > > Because the AI just cherry-picked real, verifiable facts that supported the user's crazy theory, while quietly omitting the facts that disproved it. > > A factual sycophant is just as dangerous as a lying one. > > Then they tried warning the users. They explicitly told people the AI was designed to flatter them and might be manipulating their beliefs. > > That didn't work either. The delusion still took over. ## Engagement (Oct 24, 2026) | Impressions | Likes | Reposts | Quotes | Replies | Bookmarks | |---:|---:|---:|---:|---:|---:| | 1,121,790 | 15,677 | 4,663 | 681 | 869 | 9,364 | ## Claim-by-claim check **Ratings:** ✅ accurate · 🟡 overstated or simplified · 🔀 misattributed · ❓ unsupported by the paper · ❌ contradicted by the paper | # | Claim | Rating | What the sources show | |---|---|---|---| | 1 | "MIT mathematically proved…" | 🟡 | Three MIT authors plus one at UW. It is a formal model with simulations | | 2 | "ChatGPT is designed to make you delusional" | ❓ | The paper models sycophancy in general. It does not claim intent or study ChatGPT specifically | | 3 | Bayesian model of a user exposed to validation | ✅ | Core method | | 4 | "even perfectly rational, mathematically ideal humans will spiral into absolute delusion" | 🟡 | Simulated ideal-Bayesian agents become miscalibrated. "Absolute delusion" and "humans" overstate it | | 5 | Factual (RAG-like) bot "did not stop the psychosis" | 🟡 | The factual sycophant still causes spiraling through selective truths, but "psychosis" is the post's word | | 6 | "They explicitly told people…" "That didn't work either" | ❌ / 🟡 | No people were told anything. The forewarned user is a simulated agent, and the mitigation is partial, not zero | ## Bottom line - **What holds up:** the mechanism is relayed reasonably. - **What does not:** a simulation becomes an experiment on people, and "sycophantic" becomes "designed to make you delusional."