How to Write a Haiku with AI

Haiku began in Japan as the hokku, the opening verse of collaborative linked poetry called renga, and later became a form of its own. Traditional Japanese haiku commonly count on, sound units not interchangeable with English syllables, in a 5-7-5 pattern. An English syllable usually carries more sound than a Japanese on, so a 17-syllable English haiku can feel padded. For that reason many English-language haiku poets write shorter poems and don't use a fixed count. Two useful conventions extend beyond counting. The first is kigo, a season word that places the poem in time: cherry blossoms for spring, cicadas for summer, persimmons for autumn. Japanese poets look these up in saijiki, which are almanacs of season words. The second is the cut. In Japanese a kireji such as 'ya' marks it, dividing the poem into two images that spark against each other. In English a dash, a colon or a line break often does this job. AI generators fail in predictable ways. They miscount words like 'fire', 'flower' or 'every', whose syllable count varies from speaker to speaker. They reach for the most common images in their training data (cherry blossoms, the moon, falling leaves), which produces poems that sound like greeting cards. They also tend to state emotions outright ('my heart feels peace') when a concrete image should carry the feeling. A common misconception is that any three-line 5-7-5 poem is a haiku. Senryu highlights the human condition and is often humorous, but need not be; a missing season word alone does not distinguish it from haiku. English haiku may omit a season word. A poem can also hit 5-7-5 with no concrete image or juxtaposition, which meets the count but misses the form. The most effective workflow is to give the AI your own observation, ask for many variations, and then count, cut and rewrite yourself.