South Korean scientists teach AI to say "I don't know"

A big problem with AI is that it never says I don't know, but even those who don't know speak confidently. Scientists at the Korea Advanced Institute of Science and Technology in South Korea have found a solution to this problem. They have developed a method that will allow AI to recognize when it is faced with a topic that it really doesn't know, and in that case it will remain silent instead of giving the wrong answer.

Until now, models like ChatGPT were trained to answer every question, whether they knew or not. This is why AI often creates facts that don't exist at all, which is called Hallucination. The main reason for this is that these systems learn from their initial data through Neural Networks and are not encouraged to express ignorance.

This invention is especially important in fields such as medicine and autonomous vehicles, where a wrong answer can be very damaging. AI can help a doctor in his work, but if it gives the wrong diagnosis, the result can be dangerous. If AI learns to say "don't trust me in this matter," trust in it will increase much more.