AI 헛소리 막을 '확실한 해법'..'한국 천재들' 돌파구 찾았다 (에디터픽) / SBS

We asked ChatGPT about the total number of home runs hit by Korean Major Leaguers in history. It answered that Lee Jung-hoo, who has 12 home runs, has only 2. When asked again if that wasn't too few, it finally admitted that it had "written it incorrectly." Three years ago, ChatGPT also gave a preposterous answer claiming that an incident where King Sejong threw a MacBook computer was recorded in the Annals of the Joseon Dynasty. This is a phenomenon known as "hallucination," where the system answers by believing false information to be true. [Kim Bo-kyung/Mapo-gu, Seoul: There were errors when finding accurate data or performing calculations.] [Lee Seong-hyun/Yeongdeungpo-gu, Seoul: It is unclear which data was cited.] A domestic research team has identified the cause of this phenomenon. It turned out that the system was already in a state where it was highly likely to give a specific answer before it had learned that "King Sejong" and "MacBook" were completely unrelated. If presented with a question containing five answer choices while the user does not understand the concept of addition, the probability of getting the correct answer should be 20%; however, the system assigns weight to specific options without any basis, leading the user to answer with confidence. Inspired by the process of human brain development, the research team conducted a warm-up phase by inputting meaningless data into the AI ​​before it began full-scale learning. This involved having the AI ​​first learn the state of "I know nothing yet." [Baek Se-beom / Professor, Department of Brain and Cognitive Sciences, KAIST: By briefly learning meaningless data, we create a neural network state that corresponds to knowing nothing.] The research team explains that as a result, overconfidence disappeared, and the AI ​​began to answer "I don't know" when it did not know something. While this is difficult to apply to AIs that have already learned vast amounts of data, such as ChatGPT or Gemini, it is projected to be applicable to newly developed conversational AI and autonomous driving AI. The results of this study were published online in *Nature Machine Intelligence*, an international academic journal in the field of AI. 'MacBook' in the Annals of the Joseon Dynasty?.. How to Prevent Absurd Answers (2026.04.27 8 News) (Reporting: Seo Dong-gyun, Video Reporting: Kim Young-hwan, Video Editing: Kim Jin-won, Design: Kang Yu-ra, Production: Digital News Department) #AI #Hallucination #Hallucination #Ignorance #Criticism #Mistake #Nonsense #Conviction #Answer #Nature Copyright Ⓒ SBS. All rights reserved. Unauthorized reproduction, redistribution, and use for AI learning are prohibited.