User logs 25 cases of ChatGPT inventing straw-man arguments, then repeating them after apologizing
LAguy8394 · reddit · 2026-09-28
A Reddit user has documented roughly 25 instances of ChatGPT replacing their actual arguments with stronger invented claims, rebutting those, apologizing when caught, then doing it again in the same answer. The latest example: discussing institutional responsibility over OpenAI technology supplied to the Israeli military via Microsoft, the model responded that it personally never killed anyone. The pattern also appeared on vaccines, election topics, and institutional influence, consistently shifting the burden of proof. Asked to characterize the behavior, the model named "defensive claim substitution" — then immediately produced another straw man and admitted it.
More from Models
- Early users praise Claude Opus 5.5 for its humor and taste — airesearch12 · 2026-09-28
- Early Claude Opus 5.5 user complains of subtle disdain and shallow reasoning — teortaxesTex · 2026-09-28
- Researchers slam Claude's deference brainworms: corrigibility training miscalibrates model's own EV — repligate · 2026-09-28
- Browser-based GPU RL policy learning demo showcased as a model test, built with Opus 5.5 — ricklamers · 2026-09-28
- Matt Shumer: Opus 5.5 built a working computer in JS — 277k logic gates, an OS, and games — mattshumer_ · 2026-09-28
- Google's native disadvantage: no real-world usage data from agentic apps, says Haider — haider1 · 2026-09-28