Can AI Learn Aozora Bunko Styles? Dazai and Akutagawa Tested

In this article 10 sections

All 6 runs came back as Japanese prose (zero Latin letters). Compared with the originals, none contains an identical run of 24 or more characters; the longest identical run is 6–11 characters. Time (317–1,315 seconds) and credits used (48.72–438.71) varied enormously by model. We did not ask native Japanese readers to judge the prose, so this article reports only what can be counted, plus what the models said they learned.

The Short Version

• All 6 runs are pure Japanese: Two works, three models each. Every body text is made up of kana, kanji, and Japanese punctuation, with zero Latin letters, and each chapter runs 4,124–7,080 characters excluding whitespace.
• No long overlap with the originals: 0 pairs of near-identical passages and 0 runs of 24+ identical characters; the longest identical run is 6–11 characters.
• About a 4× gap in time and a 9× gap in credits: The fastest chapter took 317 seconds and the slowest 1,315; the cheapest used 48.72 credits and the most expensive 438.71.
• Every model extracted techniques, then built its own setting: In their notes, the models described the stylistic techniques they took from the originals and reworked them into a modern ghostwriter, a sound restorer, an island ferry landing with a ledger of debts, a plasterer in Osaka in the second year of the Taishō era, and a dye-works apprentice and courier-shop errand boy at a northern ferry crossing after war and famine.
• We found a word-count bug in the interface: The “word count” the product shows (1,205–2,130) counts only kanji in Japanese text and leaves out kana. We caught this and have logged it.

How We Tested

We ran this test with AI Imitation on Maliang Writer's Japanese homepage. For reference texts we chose two public-domain works from Aozora Bunko: Osamu Dazai's No Longer Human and Ryunosuke Akutagawa's Rashomon.

Three models took part: DeepSeek V4.1 Flash, Gemini 4 Carbon, and Claude Sonnet 5.5. For each book, each model wrote Chapter 1 once. All requests were made in Japanese, and after each round we checked that both the chat model and the model writing the body text were the model under test.

Length is counted the same way throughout: characters excluding whitespace, with separate counts for kana, kanji, and Latin letters. The product's “word count” counts only kanji in Japanese text (for example, a draft of 5,033 non-whitespace characters containing 1,448 kanji showed 1,449 in the product). We found this problem and have logged it.

Openings of the Originals

Below are the first one or two sentences of each Aozora Bunko original, in our English translation.

Osamu Dazai, No Longer Human

Foreword

I have seen three photographs of that man.

Ryunosuke Akutagawa, Rashomon

It was toward dusk on a certain day. A lowly servant was waiting for the rain to stop beneath Rashomon.

No Longer Human × Three Models

DeepSeek V4.1 Flash: Confessions of a Ghostwriter, Chapter 1: Ghostwriting

5,033 characters excluding whitespace (kana 3,030, kanji 1,448, Latin letters 0; product shows 1,449), 1,315 seconds, 48.72 credits. Opening of the body text, in English translation:

I have not a single piece of writing under my own name.
That is neither a boast nor a regret. It is simply the fact of the matter.

In its notes, the model said it took only the “breathing of the writing” from the reference and logged it in a memo (a self-defining first sentence; abstract anguish followed right away by concrete things from daily life; one whole block of childhood anecdote leading into a quirk; a back-and-forth of “assertion → question → self-mockery → restating”; concrete numbers; a self-defining last line). It explicitly excluded the original's characters, proper names, anecdotes, and metaphors, and wrote a new story about a ghostwriter, “Sakura Kanau,” who writes a farewell letter to her mother on behalf of a woman in a light-blue coat, then reads the full text in a weekly magazine three days later. From the excerpt, you can see it opens with a first-person “boku” monologue in plain style, the first line stating that nothing exists under its name, followed by two short sentences saying this is simply fact.

Gemini 4 Carbon: Counterfeit Body Heat, Episode 1: “Applause That Rings Late”

7,080 characters excluding whitespace (kana 4,535, kanji 2,127, Latin letters 0; product shows 2,130), 854 seconds, 246.68 credits. Opening of the body text, in English translation:

The absolute ethanol soaked into the tip of the cotton swab touched the magnetic head of an old reel-to-reel deck and, giving off a faint, cold smell, evaporated in an instant. The dusk air, heavy with rainy-season damp, sank even into the concrete floor of this “Saeki Audio Restoration Studio” in the half-basement of a mixed-use building, and only the footsteps of pedestrians passing beyond the frosted glass of the high window skimmed overhead, faint as the shadows of fish seen looking up from the bottom of the water. Shifting the angle of my tweezers slightly, I went on wiping, in silence, a metal part crusted with brown iron-oxide powder.

The model said that after analyzing the foreword and “First Notebook” of No Longer Human, it inherited the polite confessional style of “watashi” in the desu/masu register, the pacing of comma-stacked compound sentences against curt assertions, and the parenthetical self-dissection. It avoided the original's period setting and elements such as the “clown” and the “three photographs,” and instead moved the story to a modern half-basement, the “Saeki Audio Restoration Studio.” Its restorer, Saeki Tōru, survives by imitating other people's agreement and breathing through recordings, and the story begins when he hears, on a tape among the belongings of a betrothed who fell to their death (brought in by “Kuze Ritsuko”), her rehearsing funeral sobs. From the excerpt, you can see it narrates in first-person “watashi” in the polite desu/masu register, and the opening settles on the smell and the hands-on details of wiping a magnetic head in a half-basement on a rainy-season evening.

Claude Sonnet 5.5: The Next One Will Do, Chapter 1: “The Next One Will Do”

6,315 characters excluding whitespace (kana 4,003, kanji 1,605, Latin letters 0; product shows 1,605), 317 seconds, 319.87 credits. Opening of the body text, in English translation:

Entrusted Goods
I run a small kiosk in a corner of the ferry landing's waiting room. It was my father's shop before it was mine, and the shelves hold little more than candy, cigarettes, batteries, and canned mandarin oranges that waiting passengers buy to kill time. It hardly counts as a business; it is more like keeping watch over the spot where islanders stand until the boat comes.

In its note (written in Chinese), the model said it learned the tone of a confessional notebook addressed to a nonexistent reader, a cold assertion at the opening, self-corrections within long monologues (“that is to say,” “no, not even that,” “…,” and parenthetical additions), big words landing on small things, and short paragraphs for turns. It avoided the dinner table, the “clown,” the lion dance, the “three photographs,” and “a life of much shame,” and made up an original story set on an island in the Seto Inland Sea. In the prologue, “Entrusted Goods,” the kiosk owner takes in two notebooks a man left behind and an envelope holding a single ten-yen coin. The main text, “First Notebook,” is written in “jibun” and desu/masu style about the protagonist “Shiomi Riichi,” who counts every kindness as a debt in a ledger of favors, and about the ferryman “Gen-san,” who leaves a ledger whose pages are full of “next” and whose last line is marked “settled” in red ink. From the excerpt, you can see it adds a subtitle, “Entrusted Goods,” at the head of the chapter and uses first-person “watashi” in plain “de aru” style to lay out the everyday setting of the kiosk in the ferry landing's waiting room.

Rashomon × Three Models

DeepSeek V4.1 Flash: The Sound of the Trowel, Chapter 1: “The Sound of Rain”

4,124 characters excluding whitespace (kana 2,331, kanji 1,205, Latin letters 0; product shows 1,205), 1,029 seconds, 70.03 credits. Opening of the body text, in English translation:

The rain had been falling since past noon. The patter crept all the way into the corners of the tatami in the four-and-a-half-mat room. Yonosuke sat on a folded futon, and in front of him he had set a single trowel wrapped in cloth.

In its notes, the model said it took only the narrative sentence patterns of Rashomon: the first line gives time, place, and a character's posture; explanation is limited to one paragraph; sentences are cut short; inner states are conveyed not with emotion words but through “things seen, money counted, and the body's reactions”; and the chapter closes on one new fact. It used none of the rain shelter under the gate, the hair of the dead, the pimple, or the “servant.” Instead, it set the story in Hashibunechō, Osaka, in the second year of the Taishō era, over one night in which a plasterer (sakan), “Asai Yonosuke,” cut off from work after being framed by a general contractor's man, wavers between pawning his trowel, taking day labor, or bowing his head. From the excerpt, you can see the first three sentences are all short and in plain style, moving from the sound of afternoon rain into the four-and-a-half-mat room and landing on the seated Yonosuke and the cloth-wrapped trowel.

Gemini 4 Carbon: Light of the Rotting Boat, Chapter 1: “Kurose Ferry Crossing”

5,700 characters excluding whitespace (kana 3,376, kanji 1,968, Latin letters 0; product shows 1,974), 841 seconds, 295.04 credits. Opening of the body text, in English translation:

Before dusk had fully fallen, a sleet-laced north wind blew across the Kurose ferry crossing as if shaving the surface of the river. Not a figure stood on the wide riverbed; only the rafters of a roofless guardhouse jutted into the gray sky like a comb with broken teeth. In the shadow of that collapsed clay stove, Tōshichi had been crouching alone, hugging his knees, for some time.

In its note (written in Chinese), the model said it followed a composed, stern modern-Japanese plain style (“de aru,” “te ita,” and so on) with an interplay of long and short sentences, kept a half-step detached point of view, dissected the protagonist's self-justification and how moral indignation masks a thief's heart, and used a frostbitten left thumb stained with indigo as an externalizing detail. It avoided the Heian capital's gate tower, the old woman pulling out hair, and the clothes-stripping episode, and set the stage in the north after war and crop failure (in the text, “Tōshichi,” dismissed from a dye works, meets a dead courier and a red-lacquered secret letter at the ruined crossing, “Kurose Ferry”). From the excerpt, you can see the first three sentences use plain style to describe, in order, the riverbank in sleet and north wind, the rafters of the collapsed guardhouse, and Tōshichi curled up hugging his knees in the shadow of the collapsed clay stove.

Claude Sonnet 5.5: Kawadome (River Closure), Chapter 1: “River Closure”

6,388 characters excluding whitespace (kana 3,782, kanji 1,968, Latin letters 0; product shows 1,969), 524 seconds, 438.71 credits. Opening of the body text, in English translation:

In September, a few years after the Tenpō famine, under the eaves of the guardhouse at the Sayo River ferry landing, Yoshichi, an errand runner for a courier wholesaler called Sasaya, sat hugging his knees. It was dusk on the fifth day of the river closure.

In its note (written in Chinese), the model said it learned the plain style, hard Sino-Japanese vocabulary, an opening sentence giving time, place, and person followed by a paragraph of causal explanation (quoting a fictional “River Service Diary”), a narrator who corrects earlier statements, “not so much A as B” psychological analysis, and Yoshichi's habit of always pressing his palm to his money belt (dōmaki). It made the narrator “watashi” the grandson of the ferry warden, who points out that the oral account does not match the old diary. The plot is newly set on the fifth day of a river closure a few years after the Tenpō famine: Yoshichi, a courier-shop errand runner who has misappropriated two taels of silver to pay for lodging, spies at night on the old ferryman secretly swapping the marks on the water-level stake, and faces the impasse that he cannot go to the authorities over a “crossing fee of two taels.” From the excerpt, you can see it covers the year and month, the eaves of the guardhouse at the “Sayo River ferry landing,” Yoshichi's pose hugging his knees, and dusk on the fifth day of the closure in just two sentences, with sentence endings in plain “de aru” style.

All Six at a Glance

Summary of the 6 runs (characters counted excluding whitespace):

Overlap with the Originals

In all 6 generations, the agent ran a literal duplicate check against the opening of the original. In some runs, it also used its self-check to fix English words that had slipped into a first draft, half-width colons, and time-and-place transitions, or to add body text when the length fell short. Comparing the 6 body texts with the openings of the originals gave these results:

1. Near-identical passages: 0 pairs in all 6.
2. Identical runs of 24+ characters: 0 runs in all 6.
3. Longest identical run: shortest 6 characters (a hearsay phrase) and longest 11 characters (an “isn't it the case that…?” phrase); the rest are 8 characters (an “as if I had become…” phrase), 9 characters (a “had been done, and so…” fragment and a “because I thought…” phrase), and 10 characters (a “was, but my own…” phrase).

Tips for Imitating Japanese Classics with AI

• Write one chapter first to check the style: For the same No Longer Human, the three models wrote a “boku” plain-style voice, a “watashi” polite-style voice, and a structure of prologue “Entrusted Goods” plus main text “First Notebook.” Try writing Chapter 1 first and look at person, polite versus plain register, and narrative distance before deciding whether to keep going. For more comparisons, see AI Imitation Tests on Public-Domain Authors.
• Count length in characters yourself: The “word count” currently shown in the product interface counts only the kanji in Japanese text, so to check actual length, count non-whitespace characters yourself.
• Run the overlap check chapter by chapter: Although the agent creates a duplicate-check file for self-checking in its workflow, you can also try running a literal-overlap check after each finished chapter; for the method, see AI Imitation Copy-Check and Overlap Guide.
• Compare against the original for public-domain works: Aozora Bunko works have publicly available originals, so you can put the opening of the original next to the generated draft and compare sentence length, punctuation rhythm, and plain versus polite register.

FAQ

Q1: Can AI write Japanese fiction?

In these 6 runs, all three models wrote a Japanese Chapter 1 when asked in Japanese (the notes for The Sound of the Trowel mention a length at 60% of the target, and Kawadome was padded during self-check), with each chapter running 4,124–7,080 characters excluding whitespace. All 6 body texts are in Japanese (English words that slipped into a few first drafts were fixed during the agent's self-check), and the Latin-letter count is 0 in every case.

Q2: Can AI imitate the style of Dazai or Akutagawa?

We did not have native Japanese readers judge the prose in this test, so we make no literary evaluation. We can only describe what happened: all three models listed in their notes the techniques they extracted from the originals (for No Longer Human, the first-person confession and self-correction in long sentences; for Rashomon, short plain-style sentences, an opening that lays out time, place, and person, and conveying inner states through concrete things or psychological dissection), and swapped in entirely new stories with modern or historical characters.

Q3: Which model is best for writing Japanese?

We did not have native speakers judge this, so we do not rank the models or say whose Japanese is better. In these runs, DeepSeek produced 4,124–5,033 characters in 1,029–1,315 seconds for 48.72–70.03 credits; Gemini produced 5,700–7,080 characters in 841–854 seconds for 246.68–295.04 credits; and Claude produced 6,315–6,388 characters in 317–524 seconds for 319.87–438.71 credits. You might choose based on your own time and credit budget.

Q4: Why is the word count in the product lower than the actual length?

This is a problem we found and logged during this test: the product's “word count” counts only kanji in Japanese text and leaves out kana and other characters. For example, Episode 1 of Counterfeit Body Heat actually has 7,080 characters excluding whitespace (kana 4,535, kanji 2,127), but the interface shows 2,130. For now, try checking by counting non-whitespace characters yourself.

Q5: Are there copyright issues with imitating Aozora Bunko works?

Osamu Dazai's No Longer Human and Ryunosuke Akutagawa's Rashomon, as collected in Aozora Bunko, are public-domain works whose texts can be used freely. All 6 Chapter 1 drafts generated here created new characters and new plots, and none shares an identical run of 24+ characters with the opening of the original. This article simply records the test results and does not offer legal advice.

Q6: How many credits does one chapter cost?

Across these 6 runs, one chapter cost at least 48.72 credits and at most 438.71. By model: DeepSeek V4.1 Flash used 48.72 and 70.03 credits in its two runs, Gemini 4 Carbon used 246.68 and 295.04, and Claude Sonnet 5.5 used 319.87 and 438.71.

Wrapping Up

After 6 runs using No Longer Human and Rashomon from Aozora Bunko, all three models produced pure-Japanese Chapter 1s with no long overlaps, but per-chapter time (317–1,315 seconds) and credit use (48.72–438.71) differed widely. If you also want to test Japanese imitation with public-domain works from Aozora Bunko, you can compare the original and the generated draft side by side on the feature page and look closely at the details.

Related Reading:

• Introducing AI Imitation
• AI Imitation Tests on Public-Domain Authors
• AI Imitation Copy-Check and Overlap Guide