earticle

영어

Generative AI has rapidly advanced, yet most models are predominantly trained in English, raising questions about their performance in non-English languages. This study explores whether AI-generated essays in Korean are distinguishable from human-written texts and how they compare in perceived quality. We conducted an experiment with 116 South Korean participants, evaluating essays generated by OpenAI's ChatGPT and NAVER's HyperClova X against a human-authored university essay. Participants rated each essay on a 9-point scale to assess whether they believed it was AI- or human-written and provided quality scores out of 100. The results revealed that distinguishing AI-generated texts from human-authored ones was challenging. ChatGPT (5.44) and HyperClova X (5.50) received AI judgment scores near 5, indicating uncertainty in classification, while the human-authored text scored 4.17, showing a slight inclination toward human recognition. Correct identification rates were only slightly above chance (59.6% for ChatGPT, 54.8% for HyperClova X, and 60.6% for the human essay). Notably, ChatGPT's essay received the highest quality score (85.38), surpassing the human-authored text (82.87). Despite identifying AI authorship, participants still rated AI-generated writing highly for structure, coherence, and vocabulary. These findings suggest that AI can not only mimic human writing but also outperform it in perceived quality. However, AI-generated texts still exhibit identifiable patterns, such as grammatical precision and emotional detachment. We highlight the evolving role of AI in text generation and emphasize the need for further refinement in non-English language models. Recently, ChatGPT and other generative AIs have been improving rapidly in performance. However, most generative AI models are trained in English and tend to perform well when asked and answered questions in English. Therefore, we aim to investigate whether generative AI can perform well in Korean, a language of non-English-speaking countries, and whether people can distinguish between generative AI-authored content and human-authored content.