When a finished piece of writing no longer reveals whether a human wrote it, judges, publishers, and platforms look past the prose to examine working drafts, production speeds, and file histories. Literary quality alone cannot demonstrate human labor, shifting the evidentiary burden onto the writer’s habits and intermediate records.
The Disintegration of the Self-Contained Text
For decades, literary evaluation assumed the printed page could attest to its own creation. Automated detection software failed to sustain that assumption. In July 2023, OpenAI discontinued its AI Text Classifier after the tool identified only 26 percent of synthetic text while incorrectly labelling 9 percent of human writing as machine-made. A 2023 study published in Patterns found that over half of TOEFL essays written by non-native English speakers were flagged as machine-generated, whereas essays by American eighth-graders triggered almost zero false positives. Vanderbilt University deactivated Turnitin’s detection feature in August 2023 because of these unmanageable false alarms.
Human adjudication struggles equally. At the 13th Hoshi Shinichi Award in Japan, where machine assistance is permitted, three of four winning entries in the general division used generative tools, prompting judges to acknowledge that synthetic prose had become indistinguishable from human writing. When Rie Kudan accepted the Akutagawa Prize in January 2024, she stated that roughly five percent of Tokyo-to Dojo-to came from generative software, later clarifying to media that this text was confined to lines spoken by an AI character. Her prize was retained. In June 2025, Kudan published a short story in Kohkoku magazine constructed ninety-five percent with generative software, printing five days of raw prompts alongside the text.
The absence of clear rules creates administrative paralysis. Taiwanese author Chu Yu-hsun remarked on PTS News in September 2026 that reading for competitions had become excruciating. Reviewing hundreds to thousands of submissions, he estimated that up to forty percent of entries relied on generative tools, pointing to batches of one hundred manuscripts where roughly twenty displayed a uniform cadence resembling translated American prose. Chu stated he would decline future judging invitations unless contests explicitly banned artificial intelligence. Deputy Minister of Culture Lee Ching-hwi responded on September 8, 2026, confirming that state-sponsored awards would introduce definite rules while leaving private competitions to set their own conditions.
When Drafts Become Collateral
Because prose no longer validates its origin, an author’s working files have become collateral. In the summer of 2026, Jerry Falade saw a seven-figure deal with Macmillan’s Minotaur imprint unravel. Literary agent Sandy Hodgman explained that the agency could no longer substantiate that the manuscript was entirely written by human hands. Falade possessed only two text messages from 2021, short notes in which a former graduate professor praised his prose. Although Falade condemned the unauthorized submission of his draft to automated detectors, the absence of intermediate documentation ended the contract.
Publishers now withdraw released titles when questions arise. Mia Ballard self-published Shy Girl in February 2025 before Hachette issued a British edition in November. Readers on Goodreads and Reddit raised alarms in January 2026, and after The New York Times reported on the controversy that March, Hachette cancelled the American release and halted British sales. Ballard denied using generative software, attributing the suspect passages to a freelance editor hired without oversight. It marked the first known withdrawal of a published title by a major house over machine-authorship claims. By contrast, when The Atlantic cited a Stony Brook University scan of 14,000 Kindle titles that gave H.M. Wolfe’s Daggermouth a 60 percent probability score, Wolfe’s attorney dismissed the finding as false technology, and the publisher stood behind the book as it reached number one on The New York Times bestseller list.
Establishing authenticity demands extensive paperwork. Writer Vauhini Vara tested past regional winners of the Commonwealth Short Story Prize with automated detectors for The Atlantic, flagging four stories. The Commonwealth Foundation examined the authors’ working drafts, time-stamped documents, and notes, announcing on June 22, 2026, that it was satisfied no AI had generated the entries. On June 30, Jamir Nazir won the overall prize with one of the flagged stories. Nazir composed his entry by dictating into an Android phone and editing on the mobile screen, having developed neuropathy from diabetes and cancer treatments that prevented him from typing at a desk. The explanation failed to convince every observer. In Slate, critic Laura Miller suggested Nazir was turning health struggles into a shield against stylistic scrutiny.
Administrative Ceilings Across Asia
Taiwanese competitions demonstrate the strain of retroactive inquiries. Following the final vote for the 2026 TSMC Youth Student Literature Award, organizers received a named tip claiming the winning fiction entry was produced using a model tuned to prize-winning prose. Contacted by organizers, the student author admitted only to correcting spelling errors with software. The panel held two emergency meetings in early July. Judge Lin Chun-ying noted that without hard evidence the committee had to rely on personal conviction, viewing sophisticated machine collaboration as high-level ghostwriting rather than simple plagiarism. Judge Gu Yuling found the author’s explanation evasive. The committee voted again and awarded the first prize to another entry.
The controversy prompted immediate calls for procedural changes. Judge Wang Sheng-hung proposed explicit prohibitions or signed disclosure forms detailing machine involvement. Writing in Mirror Media, Chang Chuan-fen recommended that contest guidelines include telephone interviews about creative development, placing the burden of proof directly on entrants. The Taichung Literature Award instituted an outright ban, with cultural bureau director Chen Chia-chun confirming that doubtful submissions require original drafts and accounts of the writing process.
Online publication platforms enforce compliance through quotas and categories:
- Shousetsuka ni Narou introduced four tiers in June 2026 covering direct text generation, indirect conceptual use, auxiliary proofreading, and zero machine involvement. Unclassified works were barred from updating chapters or book summaries by September 1, while wholly synthetic texts remained banned.
- Kakuyomu established three tiers based on whether machine generation exceeds fifty percent, undercuts fifty percent, or serves auxiliary roles, making disclosures mandatory for official contests.
- Amazon KDP instituted a daily limit of three new title registrations alongside machine-content declarations to protect the catalog against automated spam.
- Fanqie Novel capped account creation in May 2026 at one title daily and three monthly. Its monthly report recorded over 112,000 rejected titles and 40,000 administrative penalties, citing crude machine output as its top violation.
- Jinjiang Literature City allowed auxiliary proofreading, character names, and rough outlines while banning synthetic narrative text. The platform requires ironclad evidence, such as system prompts left inside the text or detection scores above 60 percent, imposing locked chapters, yellow cards, and ranking bans.
Careless copy-pasting still surfaces in public serials. In April 2026, chapter 4,106 of Wan Gu Di Yi Fei Chai on Fanqie published a raw system command directly in the text: “continue writing based on this ending.”
Surveys show institutional division across Asia. Koubo.jp reviewed thirty-two Japanese literary prizes in June 2026, finding five outright bans, one conditional permission, and twelve disclosure rules, including the Edogawa Rampo Prize requirement for a tool report listing software names and purposes. In late 2025, Alphapolis withdrew its 18th Fantasy Novel Prize after discovering the winning manuscript was mostly generated by software, retroactively applying rules to over 3,200 entries. In South Korea, major newspaper prizes and the Changbi Newcomer Award explicitly bar synthetic submissions, though none demand early drafts. The friction plays out instead in digital serialized markets, where reader suspicion triggers one-star review campaigns that collapse serializations without formal recourse for authors.
Historical Metrics and the Speed Trap
Attempting to verify manuscripts through output speed encounters complex historical precedents. Mathematical thresholds separate extreme synthetic spam from manual labor: producing 500,000 words in three days requires seventy-two hours of continuous input at nearly 116 words per minute. Sustained daily outputs of 1,000 to 3,000 words require 167 to 500 days for that volume. Yet human working tempos vary widely.
Fyodor Dostoevsky produced The Gambler in twenty-six days in October 1866 under contract terms stipulating that missing the November 1 deadline would forfeit nine years of publishing rights to Stellovsky. He hired stenographer Anna Snitkina, dictated afternoons, revised her morning transcripts, and delivered the manuscript on schedule. Jack Kerouac drafted On the Road across three weeks in April 1951 on a continuous 120-foot teletype roll, although years of revisions followed before publication in 1957.
Authors who recorded production metrics did so for personal discipline. Ernest Hemingway logged daily totals under a mounted gazelle head outside Havana: 450, 575, 462, 1,250, and then back to 512 words. Those higher days allowed him to fish the Gulf Stream without guilt, keeping logs so as not to kid himself, while revising the final page of A Farewell to Arms thirty-nine times to get the words right. Stephen King described his routine in On Writing as ten pages daily, amounting to 2,000 words.
Documenting speed carries professional risk. Anthony Trollope placed a watch on his desk to produce 250 words every fifteen minutes during three-hour dawn sessions. His posthumous Autobiography revealed weekly logs averaging forty pages, noting that an insufficient week was a blister to his eye. Later Victorian critics dismissed his catalog, treating his strict schedule as evidence of uninspired mechanical labor.
Building a Record Without Illusions
Physical and digital audit trails can be fabricated. Browser extensions simulate human typing pauses, and tests reported by Plagiarism Today in November 2025 demonstrated that running synthetic prose through automated rewriting functions produced reports certifying seventy-five percent human input. Programs like the Authors Guild Human Authored certification, expanded to all American releases in March 2026, rely entirely on self-declarations backed by a ten-dollar fee for non-members. Grammarly introduced Authorship in August 2024 to log keystrokes, pasted passages, and synthetic revisions.
The writing studio in Slima records an Authorship Record as writing takes place rather than reconstructing habits after the manuscript is finished. It logs total duration from initial draft to final edit, active writing days, consecutive streaks, daily net growth, and typical working hours. Inflow channels are separated so that words generated by Slima AI, pasted text, and imported files are tracked independently, while consultations with AI are recorded by query count rather than word volume. Untracked spans are marked as unknown on the interface, and typed word counts are designated as derived net balances rather than direct measurements. The environment generates a printable summary mapping these records to the reporting formats of Shousetsuka ni Narou, Kakuyomu, and Amazon KDP, alongside a verification link and an optional public page that omits manuscript text and revision diffs. A frozen snapshot certificate, anchored by SHA-256 ledger hashing and server timestamps, is available for Pro accounts and individual purchase to give authors factual documentation for declarations without claiming to prove human authenticity or detect synthetic text.
Statutory regulations offer no shelter. Article 50(2) of the European Union AI Act obligates artificial intelligence providers to make synthetic text machine-readable beginning in August 2026. Article 50(4) confines disclosure mandates to public-interest reporting and exempts texts subjected to human editorial review, excluding fiction entirely. The rules governing writers are private contracts enforced by platforms, publishers, and prize committees, leaving the burden of verification entirely in the author’s hands.