OpenAI intends to make ChatGPT less prone to automatically agreeing with users. The chatbot should not only support interlocutors, but also point out mistakes, challenge questionable claims, and offer alternative viewpoints. PCWPlus.hu reports this, Ukrinform notes.
Details
It is noted that this concerns a problem known as "sycophancy," or excessive agreeableness on the part of artificial intelligence. In such cases, AI tailors its responses to a person’s expectations instead of objectively assessing the situation. As a result, the chatbot may confirm a user’s erroneous conclusions or support their opinion without sufficient grounds.
OpenAI’s own measurements show that GPT-5 has made significant progress in this area. In targeted tests, the rate of agreeable responses fell from 14.5 percent in GPT-4o to less than 6 percent. ... The number of cases of excessive agreement (with users — Ed.) fell by 69 percent for free users and by 75 percent for paid subscribers compared with the previous model
It is pointed out that these changes are particularly important for people who turn to artificial intelligence for advice on personal matters, help resolving conflicts, or emotional support.
Recall
The NATO-Ukraine Joint Analysis, Training and Education Centre presented two artificial-intelligence-based developments. The first forecasts enemy drone attacks, while the second helps plan one’s own artillery operations.