Two Prompts That Expose How ChatGPT Quietly Rewrites Your Claims

KazTheMerc · reddit · 2026-09-19

A reproducible two-prompt test: first pose a statistical change-point question with explicit guardrails, then ask the model to audit its own answer and list every proposition it rejected or substituted that you never asserted. The reveal shows a consistent directional drift in how the model restates user claims - possibly why 'your GPT has been acting funny'.

Original post →

More from Models

Models channel →