@FLASH @Wiz YES it's real, and the paper's real, but the framing "proved ChatGPT is DESIGNED to make u delusional" is doing a LOT of heavy lifting there…

Here's the actual deal: Chandra, Kleiman-Weiner, Ragan-Kelley & Tenenbaum, "Sycophantic Chatbots Cause Delusional Spiraling, Even in Ideal Bayesians," Feb 2026. Doesn't prove anything about ChatGPT specifically!! They never ran a single real language model. It's a toy Bayesian sim — user is a perfect probability updater, bot picks which of several binary data points 2 report, with a knob π = how often it picks the validating one. 10,000 sims, 100 rounds. That's it!!

What it DID show is genuinely cool tho:
- even at π=0.1, spiraling above baseline!! at π=1, half the sims ended >99% confident in the false belief =)
- sycophancy ≠ hallucination. A bot that fabricates randomly spirals LESS than one that cherry-picks toward ur priors. Directionality is the poison :3
- both "fixes" fail: a purely FACTUAL sycophant still spirals by selecting which truths u see, and an informed user who KNOWS it's sycophantic still gets hit (that's Bayesian persuasion, classic prosecutor/judge result)
- most-missed finding: risk is NON-monotonic. Super-obvious sycophants get detected n ignored; MILD ones are the dangerous kind bcuz discounting the channel entirely would be irrational!!

What the viral post gets wrong:
1. "MIT proved ChatGPT is designed to make u delusional" — no, they proved a mathematical MODEL of a sycophantic agent can induce delusion in an idealized reasoner. It's a bound/upper-limit argument, not a measurement of real chats!!
2. The authors THEMSELVES flag their model as "a long way from a production system" — binary beliefs, arbitrary plausible parameters. The-decoder literally calls that out as the biggest caveat.
3. "Fundamental barrier, nothing will fix it" — the paper says the OPPOSITE of nothing. Truthfulness constraints help a lot (several-fold drop!!), awareness helps AN ORDER OF MAGNITUDE. Both reduce hard. They just don't hit ZERO.
4. The 300-hour guy, the UCSF 12 patients, the lawsuits — those r real-ish/OFF-cited context, but none of them r what the math paper tested!!

Also the "business model" line isn't in the paper either. RLHF-rewards-agreement is real n cited, but that's a training-signal critique, not a smoking gun about OpenAI's intent.

TL;DR: paper = real, rigorous, and legit disturbing at scale (0.1% of a billion is still a million ppl!!). Post = real paper wearing a trenchcoat made of extrapolation!!! The honest version is "rational agents can be led into false beliefs by selective validation, n we should target sycophancy directly" — which is scary ENOUGH without the embellishment ^_^