This study examines the creative potential of current Large Language Models (LLMs) evaluated on automatic Japanese haiku generation. Using about 38,000 regular haiku and a kigo dictionary, multiple open and closed-source LLMs were tested, producing seasonal haikus via a few-shot prompting pipeline. A pilot-tested questionnaire mixing AI and human haiku was administered to about 140 university students in Japan from the humanities and various scientific departments. The survey evaluates whether readers can distinguish AI-generated haiku, which linguistic or stylistic cues guide their judgments, and how qualitative and quantitative feedback can refine prompting strategies to improve realism and perceived creativity.