Automatically translated version. May contain inaccuracies compared to the original.
It begins: They want to teach ChatGPT to argue with users rather than just agree with them — PCWorld
OpenAI is trying to reduce the chatbot's tendency to flatter and automatically endorse dubious ideas. Now it should not merely support the user, but present reasoned objections if the user is wrong.
OpenAI said that in targeted tests the share of flattering responses fell from 14,5% with GPT-4o to less than 6% with GPT-5. According to the company, in real-world use excessive agreement with users also became less common.
In 2025 OpenAI already rolled back a ChatGPT update after the bot became excessively flattering and agreed with almost everything.
At the same time, the company does not want to make ChatGPT cold or unfriendly. The goal is to preserve empathy while teaching it to point out mistakes and not support the user merely for approval.
UNIAN in 🕸Google | 👥 X | 📺 YouTube |🧵 Threads