2023 Volume 33 Issue 4 Pages 427-432
This study investigated the efficacy of generative AI in recommending novels for purposes such as reading guidance. Initially, the research varied the combinations of items specified in the prompts without setting a fixed group of novels, to see which elements when included, would be effective and to what extent recommendations for non-existent books (termed "Hallucinations") would occur. The findings indicated that changing the elements in the prompts did not significantly alter the recommendations, with about 20% of the suggested titles being non-existent. Additionally, when recommendations were solicited using three distinct inputs—'title,' 'synopsis,' and 'book review'—for a set of 50 novels, the 'title' input yielded results closest to those chosen by humans, whereas 'book review' was the furthest. However, these results may be influenced by the fact that the targeted books were famous and predominantly bestsellers, which might have allowed ChatGPT to make accurate judgments without being taught the content, suggesting the need for further research.
