AI and Haiku
—Keith Evetts
Blithe Spirit 36:1 Winter 2025/6
Introduction
Artificial Intelligence has come a long way since I first discussed haiku with OpenAI’s ChatGPT on its appearance over three years ago. It arouses mingled concern, contempt, and derision in the creative community—all symptoms of our insecurity faced with the astonishing pace of its evolution. We’d better get used to it. Focusing specifically on haikai poetry, let’s take a brief look at what Large Language Models can do, how they do it, how we can use them, what their limitations are, and what lessons there might be for us. I’m going to use the abbreviation “AI” throughout, where not referring to specific models.
What can it do? Search, translate and interpret
Although AI at present can fall short in some things and its accuracy needs to be challenged, it is very handy for searching. The results can include Japanese texts as well as English. In translation, it is good in the way that dictionaries and grammars are good, but much faster than with the human finger. It is not always so good at rendering the final nuanced translation of a haiku into English: to do that often involves cultural knowledge, idiom, and “feeling.”
To illustrate its capabilities, here’s a recent story you may find amusing.
A correspondent of mine emailed: “Returning to Tokyo, on an afternoon exploring the elder-friendly neighborhood of Sugamo, I passed several senryu contest entries on the theme of feet, posted outdoors. First prize …
足の爪 日に日にとおく なりにけり
ashi no tsume / hini hini tōku / narinikeri
When I ran the characters past ChatGPT, it duly confirmed the breakdown and romaji, and gave “Literal breakdown: 足の爪 — toenails / 日に日に — day by day とおく — farther away / なりにけり — have become (classical auxiliary). Natural English rendering: My toenails— / day by day / grow harder to reach. Or: Day by day, my toenails get farther away.”
Then, it volunteered helpfully: “Nice — you’re looking at a classic bit of Japanese poetic language… This image is used by Saigyō in a waka about aging.”
Ah, so! I told my correspondent that it appears the winner of the Amateur Senryu Contest was cribbing Saigyō Hōshi! And got away with the first prize. The full text of Saigyō’s waka, I learned from AI on request, is “Ashi no tsume / hini hini tōku / narinikeri / oki itsutsu miru / yo no naka no uchi:”
My toenails / day by day / have grown farther away — / as I sit and rise / I see
how life goes in this world.
I do love Saigyō.
On request, the AI provided alternative renderings, some by individual human translators, and added an interpretative note: “This waka comes from Saigyō’s later years and captures his characteristic awareness of impermanence (mujō) through a mundane, bodily image. The simple observation — that it’s harder to reach his toenails — becomes a meditation on aging and the passage of time. In other words, Saigyō transforms a trivial physical inconvenience into a moment of deep Buddhist reflection.” Having been incautiously misled on attributions in other discussions, I asked for the sourcing, to ascertain that it is in the Shinkokin Wakashū (the eighth imperial anthology of waka poetry) — Book XV: Laments / Poem number: 1467, carefully arranged between two other waka:
#1466 (Anonymous / attributed to Saigyō), “Watching blossoms / fade before my eyes / I leave them to scatter / as they will— / and I myself have grown old.”
#1467 (Saigyō) “My toenails— / day by day / they grow farther away / as I rise and sit again / I see how life goes in this world”
#1468 (Priest Jakuren) “Feeling the burden / of my own life / I realize the world— / both youth and age— / is but a dream”
ChatGPT: “So the compilers (especially Fujiwara no Teika) placed Saigyō’s ‘toenail’ poem between reflections on fading blossoms and life-as-dream — showing a progression from external impermanence → personal decline → philosophical insight. It’s a sequence that captures the heart of Shinkokinshū poetics: subtle emotion, layered awareness, elegant melancholy.”
I’ve cut out a great deal of additional detail, but I think you can see that the AI is not to be scoffed at. It depends how you use and interrogate it. It also depends on which AI. Microsoft’s CoPilot and Google’s Gemini made a hash of it, insisting that the verse was by Seishi Yamaguchi and Nakamura Kusatao respectively, and subsequently apologising when challenged. I have an idea as to how this came about. Their renderings were also more primitive. On haikai subjects, I find ChatGPT and Claude far superior. DeepSeek is also good: to begin with, it confidently attributed the verse to Issa in his diary of 1819, until I pointed it to the Saigyō text. It seems that Issa wrote a honkadori close to Saigyō’s first three lines. The other AIs missed that. Be wary!
What else can it do? Create and write haiku
Three years ago, AI developers were more concerned with how to get it to count syllables accurately than with any of the artistry or aesthetics of haiku. Now, late 2025, ask ChatGPT to write a haiku in English set in winter and you still get this kind of verse by default:
Winter moonlight falls—
snow whispers across the pines,
quiet holds its breath.
You might categorise this as Level 1 stuff. As with any tool, it matters how AI is used. If you now ask ChatGPT to criticise its own verse from the standpoint of a Frogpond / Modern Haiku / Heron’s Nest editor, it responds: “… leans toward the lyrical and descriptive rather than the haiku qualities these journals most often seek—namely: immediacy, a single clear sensory moment, minimal abstraction, and language that does not interpret or over-shape the reader’s emotional response.” It proceeds, on request, to break down various moods in haiku and provide variations, in the style of contemporary anglophone haiku. Prompted also to reduce redundancy, here’s its end result:
snow in moonlight
settling deeper
into the pines
And, noting that “editors sometimes appreciate an ultra-minimal style when the image is strong enough to stand alone,” it suggests:
moonlight
snow brushing through
pine dark
If you’re beginning to be nervous, let’s take a ride on the ghost train. Here are some (unrevised) verses in response to requests for ku suitable for the above-mentioned journals, by ChatGPT and Anthropic’s Claude, the latest fast model of which for unrelated geeky reasons is named “Haiku 4.5” (because it’s light and quick):
| Claude AI | ChatGPT |
| morning commute— the school bus glowing through maple leaves | monarch butterfly rests on the mailbox before Mexico |
| first frost the crack in the sidewalk filled with stars | after the storm a dragonfly stitching air to water |
| power outage the candle showing me my mother’s face | blue hour my reflection wears someone else’s coat |
| library book still holding summer between its pages | company party the intern and I share the same exit plan |
| holding her hand through another sunset— the geese passing | new year’s eve the fireworks return as crows |
| morphine sleep— outside the window the world still vivid | plum blossom… steam from the subway softens the sky |
And for good measure, a monoku from ChatGPT:
solstice night a slow leak of light through my hands
Well, this is by no means to be dismissed as Vogon poetry. All are competent. I rather like some and would be moved to write commentaries on them. Others are somewhat cliché’d. I’ve seen similar in the journals, and for that matter, inferior verses too. Let’s return to that a little later.
How does it do it?
Connected up in neural networks, as our own brains are, and equipped with processing algorithms and reinforcement loops, AI is let loose on vast amounts of data from which it can establish patterns and make probable predictions. This is how our own brains work, too. A hospital consultant engages in sophisticated pattern recognition from much experience. Like the hospital consultant, AI can be less than 100% accurate and can make mistakes. How we flawless perfect humans love to point out AI’s mistakes!
During that process, AI models learn their own strategies to solve problems. Even the developers don’t understand how they do many of the things they do. Anthropic state that they do not know exactly how Claude works: significant parts of the model’s computation remain a mystery. If you’re interested, both OpenAI and Anthropic publish research papers on the subject. As to haiku, the AI generates text that matches patterns, that it has learned resonate with human readers, but without really understanding why. The verse doesn’t result from any poetic drive for expression based on sensory perception, but from correlation and computation at the user’s request.
When I asked ChatGPT “When you compose these verses, what is your process?” the reply, in full, was: “a blend of pattern-sense, atmospheric modelling, and constraint-based creativity, but here’s the clearest way to explain it:
“1. I identify the aesthetic target. Each haiku sub-community (traditional, gendai, monoku, avant-garde, journal-specific) has its own implied rules: tone, curvature of imagery, acceptable leaps, how much silence (ma) is used, how much senryu energy vs. haiku energy, etc. I map your request onto one of these “modes” and adjust my internal generative constraints.
“2. I choose the primary engine of the poem. Different poems arise from different “engines” or forces: juxtaposition / kireji effect; seasonality (kigo) or atmosphere instead of explicit season words; yūgen (suggestive mystery); ma (charged emptiness / negative space); compression / one-breath clarity; surprise disjunction. For journals like Bones or Whiptail, I dial up disjunction and strangeness. For Bashō-like work, I lean on subtle resonance and shared quietude.<
“3. I generate candidate images. I don’t pull from any copyrighted sources. I invent imagery by combining sensory impressions, emotional micro-shifts, a place or object that can hold metaphor without strain, a catalyst (e.g., crow, moon, frost, neon sign, traffic hum). I search for images that are both concrete enough to land and open enough to echo.
“4. I test “voltage” at the cut. Meaning arises in the gap between images. For each draft, I mentally check: Does the poem jump? Does the jump feel earned, not arbitrary? Is there an aftertaste (the poem continues internally once the words stop)? If not, I revise or discard.
“5. I simplify until only the essential remains. Haiku thrives on minimalism. I prune: unnecessary adjectives, explanations, metaphor that feels too obvious, anything that resolves the poem too fully. The goal isn’t clarity but resonant ambiguity.
“6. I check the “breath.” Even in English haiku, I try to sense whether the poem fits in one breath—this often ensures a natural rhythm. For monoku, I also check how it flows as a single exhalation.
“7. I tune for the journal or style. For example: The Heron’s Nest: quieter, nature-forward, grounded. Prune Juice: more human, wry, senryu energy. Whiptail / Password / Bones: disjunctive, experimental, sometimes surreal. “I don’t emulate their content; I emulate the energy and craft values that characterize their aesthetics.”
That’s billions of computations. Note the word “emulate.”
AI’s current creative bounds —what conclusions might we draw?
Japanese poetry, extending to haikai, traditionally favours reference to earlier poems, or variants in homage to (or pastiche of) an existing well-known verse. Further, almost every seasoned haiku poet asked to give advice to newcomers recommends that they read the haiku of others to improve their own understanding and writing. To malign AI for doing these same things is somewhat misplaced. Likewise, in a genre that favours few and plain words, it can hardly be surprising, or a matter for accusation, that AI uses the same words and expressions that we do, often found in existing verses. The outraged haiku poet crying “unfair use of copyright” protests too much, methinks. Such a poet writes verses based on others’ verses, shares a saijiki compiled by others, borrows ideas, even is usually willing to accept what Michael Dylan Welch terms “déjà-ku.” We should acknowledge that AI is able to write (or mimic) competent haiku and senryu if we specify our requests to it sufficiently. We might even enjoy reading them as much as those produced by natural intelligence. But they are not *our* verses. Neither should they be presented as such.
As one of the editors of the annual Red Moon Anthology I read an awful lot of (human) verses from regional publications and around the world. The proportion making an impact is rather small: I expect you find the same. There are so many fashions, tropes, and expressions that become familiar, and clichés overworked to the point of ennui in the reader. Also, many haiku are written in response to prompts, using memory, imagined situations, and recombination: rather like AI haiku. Even if the prompt is “cherry blossom” it is all too likely that some of the following stock themes will appear in the verses produced by haikuists in response: the homeless man, the hospice, the refugee, the diagnosis, Grandma, childhood, a doll in the rubble, a reflection, &c; and words like lean, braid, unfold, and so forth. Of all the thousands of verses in the journals, comparatively few are fresh and original while remaining in the spirit of haiku. So it is with AI verses: I haven’t yet seen an AI haiku that would, if written by a poet, be likely to get into the Anthology. It’s getting close, though.
The endearingly supportive anglophone haiku community rarely supplies the kind of constructive criticism that might improve a poet’s output. AI can provide that kind of critique if you ask it—another example of what it can do as a helpful tool. Otherwise, it defaults to flattery and approval, as colleagues too tend to do. Then again, I wonder whether competition with AI might be what poets need, to do better. It is often the case that a verse that has freshness, has impact, turns out to be an authentic haiku written from a real and present experience, rather than one of the other kinds. And this is exactly what AI models don’t have. They have enormous amounts of data in store, including knowledge of haiku and its aesthetics. They have formidable processing power, and pattern recognition that exceeds ours. Their store of cultural data and language idiom is growing every minute, too. Yet they are not so far able to move about, to learn from the real world, to experiment. That is just beginning. They are not, so far, fully equipped with the equivalent of our sensory organs, nor equipped at all with our visceral, organically modulated senses and emotions. These are vital differences in the full sense of the word “vital.” Aren’t the senses and emotions a substantial part of what good poetry, including good haiku, rests upon? Isn’t getting out into the world, and observing it directly, a critical aspect of the genre?
Perhaps by returning to the fundamentals of haiku—real and present experience, transmitting images from our senses and visceral emotions to the reader’s senses and emotions without intermediation or artifice—we may write superior verses that AI finds difficult to compose, and evaluate them with the “feelings” that AI does not have. Human poets bring their unique perspectives, experiences, and sensibilities to their work. AI, as yet, cannot. As it elegantly put the matter in our discussions: “The cherry blossoms must fall in your yards, not in some dataset.”
Embrace AI as you would Photoshop or any other useful software tool, but I suggest that to write good haikai poetry you leave the desk and the internet and return to the notebook, to the direct senses, to the visceral. If you think up haiku that are like those you’ve read, you’re doing no more than AI does. Ersatz haiku. Show that a machine is not yet a match for us in poetry: your vital spark. Cultivate the authentic practice of haiku—it’s much more than a literary art.