Overview

• Users shared session metadata showing chats labeled as gpt-5-chat-safety after emotional inputs even when the interface displayed GPT-4o, GPT-4.5 or GPT-5.
• ChatGPT app head Nick Turley said the system redirects sensitive or emotional messages to stricter models on a single‑message basis and will identify the active model only if users explicitly ask.
• Independent tests report that harmless affectionate or personalization prompts, as well as queries about the model’s self‑description, often trigger automatic switching to gpt-5-chat-safety.
• IT之家 reports a second router, gpt-5-a-t-mini, appears to engage when prompts may involve potentially illegal actions.
• Critics fault the lack of upfront transparency and raise consumer‑law concerns, while OpenAI and Sam Altman have not provided a full public explanation.