Haiku AI and Generative Models — 400-Year Tradition Meets AI in the 2020s

400 Years, 17 Syllables, and Generative AI

Four hundred years since Teitoku popularized haikai; 330 years since Bashō established Shōfū; 130 years since Shiki coined “haiku.” In the 2020s, a new participant has entered haiku’s history — generative AI.

In November 2022, OpenAI released ChatGPT. Ask it to “compose a haiku in 5-7-5,” and in seconds it returns a verse that respects the syllable count, includes a seasonal word, and uses a cutting word. This is a state of affairs unprecedented in 400 years of haiku history.

The Academic Prehistory of Haiku AI (2000-2020s)

Before generative AI, computational systems for haiku already existed:

2000s — Early Haiku Generators

Early haiku-generation research took place at Keio University, Kyoto University, and elsewhere, using:

• Statistical language models (n-grams)
• Template-based approaches (filling 5-7-5 slots with words)
• Combination with kigo dictionaries

Early systems could respect syllables and kigo but produced almost random outputs semantically — “the wind blows / the summer sky has / cicadas cry,” and so on. Emotionally impoverished verses.

2010s — Neural Language Models

Recurrent Neural Networks (RNN), LSTM were applied to haiku, gradually producing more natural verses. But contextual understanding and semantic consistency were still insufficient.

Around 2018 — Transformer / BERT

Since the Transformer architecture (Google, 2017) and BERT (2018), language-model performance leapt forward. These are direct ancestors of the 2020s large language models (LLMs).

2022 — The Age of LLMs

With GPT-3.5 (ChatGPT) publicly available, anyone can now have haiku composed by AI. This was followed by GPT-4, Claude, Gemini and many other LLMs with haiku-generation capability.

Actual AI-Generated Haiku

Typical outputs from major current LLMs when asked “Write an autumn haiku in 5-7-5” (representative, model- and time-dependent):

ChatGPT / GPT-4 family

Autumn wind ya / the sound of falling leaves / stillness yo

Syllables: 5-7-5 (respected) Kigo: autumn wind, falling leaves (season overlap — a bit careless) Kireji: both “ya” and “yo” appearing (double kireji, not recommended in textbook)

Claude family

Moonlight ya / a single duck cuts / across the lake

Syllables: 5-7-5 Kigo: moon (autumn) and duck (winter) season overlap Kireji: “ya”

Gemini family

Autumn is deep — what does the neighbor do

That’s an existing famous verse (Bashō) output as-is. The LLM’s training data contains famous verses, and it sometimes just outputs them.

Overall LLM haiku capability:

• Surface rules (syllable count, kigo, kireji): mostly kept (90-95% success)
• Semantic consistency: moderate (60-70%)
• Overlap with existing verses: occasional (training-data contamination)
• Cultural depth, necessity of feeling: clearly shallower than human masters

Strengths and Weaknesses of Haiku AI

Strengths

1. Syllable-count discipline — counting 5-7-5 is effectively perfect
2. Kigo coverage — memorizes 5,000 kigo
3. Kireji use — places “ya,” “kana,” “keri” appropriately
4. Mass generation — can produce 100 verses in a second
5. Editing and revision — can suggest improvements to human verses

Weaknesses

1. Cultural depth — does not understand why a particular kigo fits a particular verse
2. Originality — limited to imitation and recombination of training data
3. Contemporaneity — struggles to express modern life in modern seasonal feeling
4. Contradictory sentiments — hard to generate counterintuitive expressions like Bashō’s “Stillness — sinking into the rocks, the voice of the cicadas”
5. Rhythm and sound — respects 5-7-5 but poor at evaluating “beauty of sound”

The Modern Haiku World’s Response

Modern poets and journals divide broadly into three camps:

1. Active Advocates — Welcoming AI as a New Possibility

Mainly younger poets and the Modern Haiku Association. Claims:

• AI can popularize haiku (easier entry for beginners)
• Co-composition with AI can generate new expressions
• New contest formats like AI haiku competitions are possible

2. Cautious Middle — Distinguishing AI and Human Haiku

Mainly mid-career poets and general haiku journals. Claims:

• Treat AI and human haiku as distinct forms
• General journals should not publish AI-generated verses
• Require disclosure of AI use

3. Refusal — AI Haiku Is Not Haiku

Mainly traditionalists, Hototogisu-line, Haijin Kyōkai. Claims:

• Haiku is a poem of human feeling and AI cannot make it
• Accepting AI haiku would damage haiku’s artistic value
• Prohibit AI use in Haiku Kōshien and official contests

These positions coexist and do not converge on a single answer. An ongoing debate.

Applications of Haiku AI

Haiku AI is already used practically in several areas:

1. Education

• Middle- and high-school Japanese classes use it as a haiku composition-support tool
• Classes analyzing “AI-generated verses”
• AI-provided feedback on student verses

2. Publishing

• Haiku journals do pre-filtering with AI (detecting basic rule violations)
• New AI-haiku collections published (human curation)

3. Tourism and Regional Development

• At tourist spots, apps that return AI-composed matching verses in real time
• Implementations in Kyoto, Matsuyama, Sendai, and other haiku-related sites

4. Accessibility

• AI helps elderly and disabled users compose haiku by checking syllables and kigo
• Audio haiku support for the visually impaired

5. International Exchange

• Machine-translated haiku (Japanese → English → Japanese) supports dialogue with world poets
• Support running multilingual haiku contests

“Co-Composition with AI” as a New Form

In the mid-2020s, some poets actively try AI co-composition. The method:

1. Human writes the upper 5
2. AI proposes 20 candidates for the middle 7 and closing 5
3. Human picks one, or edits
4. Human makes the final judgment

In this co-composition, humans retain the final artistic judgment while leveraging AI’s generative power. Same idea as AI co-composition in visual art or music today.

Kōno Saki’s Experiments (2024-)

The young poet Kōno Saki demonstrates “co-composing haiku with AI” in her lectures and workshops. She advocates the “middle way — don’t leave everything to AI, don’t reject AI either.”

Kishimoto Naoki’s Caution

Meanwhile, Kishimoto Naoki maintains that “haiku’s soul is in fine observation, which AI cannot do,” and keeps distance from AI haiku. An example of the ongoing debate.

Technical Limits of Haiku AI

Technical limits current LLMs still face in haiku generation:

1. Syllable-Counting Errors

Counting Japanese morae is hard for LLMs. Handling contracted (kya, kyu, kyo), glottal (っ), and elongated (ー) sounds sometimes goes wrong. “Tōkyō” is 4 morae, but the LLM might count “toukyou” as 5.

2. Overlooked Season Overlap

LLMs often overlook season overlap when composing.

3. Training-Data Bias

LLM training data is centered on modern Japanese. Classical haiku (Edo period) and colloquial or free-verse haiku are harder to generate.

4. The “Mediocrity” Wall

LLMs tend to output the training-data average. For haiku this means mass-produced mediocre verses. The “unexpected leap” that marks a human master is hard for LLMs.

The Future of Haiku AI (2030s Predictions)

Based on current trends, the state of haiku AI in the 2030s might look like:

1. Multimodal Haiku

AIs that generate haiku from photo, video, and audio input. Already implemented in GPT-4V and similar. Growing application in tourism and events.

2. Personalized Haiku

AI that learns individual users’ past haiku, interests, and lifestyle patterns, and generates haiku “in that person’s style.” Combined with SNS posting services.

3. Sentiment-Evaluation AI

An AI that judges whether a generated haiku is “really good.” If realized, generation-plus-evaluation AI pairs can generate high-quality haiku at scale.

4. AI Poets Across Borders

Multilingual AI supporting simultaneous composition in Japanese + English + Chinese + Korean, and dialogue with world poets. Accelerating internationalization.

5. Digitization of Haiku History

Constructing databases of all 400 years of haiku, with AI performing “comparison of historic masterworks and modern verses.” Research at Waseda, Ritsumeikan, and elsewhere is under way.

Redefining “What Is Haiku?”

The arrival of haiku AI reopens fundamental questions:

1. Is it haiku only if a human made it?
2. If it observes 5-7-5, kigo, and kireji, is it haiku regardless of who (what) made it?
3. Who (what) decides artistic value?
4. What is “the author” of a haiku?

No settled answers yet. But what’s certain: 21st-century haiku cannot avoid coexistence with AI. Haiku is entering a period of transformation on the same order as when Teitoku opened haikai to the masses 400 years ago.

What “AI Haiku” Reveals About the Essence of Haiku

Paradoxically, AI haiku has clarified:

1. Syllable count, kigo, and kireji are the surface of haiku — even perfect adherence produces no masterpiece
2. The essence of haiku is “experience and observation” — Bashō’s Hiraizumi summer grass, Issa’s skinny frog, Hōsai’s solitude — all rooted in lived experience
3. The composer’s life makes the haiku — AI has no “life”
4. Haiku is a verse form and simultaneously a way of living — Santōka’s mendicant travel, Hōsai’s Nangō hermitage, Tōta’s Truk Island — life enabled the haiku

AI now imitates 90% of haiku’s craft, and this brings the remaining 10% — human life, observation, feeling — into sharper relief. Perhaps that’s the most important byproduct of AI’s arrival.

The Close of Haiku: 400 Years

This series traced 400 years, from the early-17th-century Teimon haikai of Matsunaga Teitoku, through Bashō/Buson/Issa, Meiji’s Masaoka Shiki, Taishō-Shōwa’s Kyoshi/Hekigotō/Santōka/Hōsai/Shūōshi/Seishi/Kusatao/Shūson/Hakyō/Tōta, Heisei’s Arima Akito/Takaha/Hasegawa/Kishimoto, up to the 2020s’ haiku AI.

“Why have 17 syllables lived for 400 years?” Partial answers offered across the series:

1. The extreme brevity of 17 syllables maximizes the space for reader imagination
2. The 5,000-word kigo system compensates for brevity with information density
3. The three cutting words “ya,” “kana,” “keri” give structural bones to 17 syllables
4. Poets in each era have accumulated the craft of folding contemporary material into 17 syllables
5. Master-journals, gatherings, Haiku Kōshien, SNS — multiple layers of transmission are in place
6. Amid new variables like internationalization and AI, new verses keep being made

Bashō, Issa, Shiki, Kyoshi, Santōka, Tōta — 21st-century poets compose new 17-syllable verses out of these 400 years’ collective legacy. And 22nd-century poets, too, will surely inscribe their era’s living in the same 5-7-5 frame.

To Close, With Bashō’s Words

Bashō is said to have taught in old age:

“Fueki-ryūkō” — “the unchanging and the flowing”

Both the unchanging (17 syllables, kigo, kireji) and the ever-changing (subject matter of the age, feel of one’s life) are what make haiku what it is. Bashō’s phrase from 400 years ago is still a valid guide for us in 2026, when generative AI writes haiku.

If this series has been for the reader an entrance to the haiku universe, that would be a joy. Continue reading the verses of Bashō, Buson, Issa, Shiki, Kyoshi, Santōka, Hōsai, Tōta, and today’s Kōno Saki, Iwata Kei, and all the poets across 400 years.

And if you find yourself wanting to compose your own 5-7-5, that is a new participation in the 400-year tradition. However crude your verse, the 17-syllable frame — built up by everyone from Teitoku to Tōta — is waiting for you.