On June 17, 2026, my writing was called into question. More precisely, what was called into question was whether I was the one who had written it.
I was working as a web writer, producing articles about cloud accounting software for a corporate media outlet: features, pricing, things to watch out for when adopting it. A practical piece of several thousand characters with almost nothing to do with generative AI. I read the materials, built an outline, wrote it, and delivered it. The next day the editor got in touch: “Are you using generative AI? It was flagged by a detection tool.”
I wasn’t. So replying “I’m not” should have been the end of it. But as I typed the reply, my hands stopped. How would I prove it? I had drafts. I had an edit history. But an AI can produce drafts too, and one can retype an AI’s entire output by hand. No matter how many traces of human handling pile up, they never amount to decisive proof that the text originated in a human being.
As incidents go, it was trivial. The manuscript was delivered, and it caused no problem at work. Even so, I felt that something had ended. Not because generative AI had learned to write humanlike prose, but because the grounds for reading a text written by a human as a text written by a human had quietly begun to lapse.
Contents
Author Undecided
Texts written before 2026 and texts written after 2026 cannot be read in the same way.
This is not a periodization from the history of technology. It is not as if only humans wrote until December 31, 2025, and AI broke in the next morning. Here, “2026” is a symbolic date. There is an era in which human authorship could serve as the tacit premise of reading, and an era in which that premise can no longer be sustained. I place this date at the boundary between them.
There used to be a human being on the other side of a text. At least, it was not unreasonable to imagine so. Behind a twisted metaphor one could picture a twisted person; behind relentless repetition, a relentless person. Ghostwriting, editing, plagiarism, and hired pens have always existed. Even so, the suspicion “Is the source of this voice human at all?” did not reside permanently in every text.
Now it does. And this suspicion does not arise only for texts written with AI. Even for a text written without ever touching AI, the finished product cannot confirm that fact. When an author says “I wrote all of it myself,” the reader has no way to verify the statement.
Take one example. In January 2024, when Kudan Rie received the Akutagawa Prize for Sympathy Tower Tokyo (Tōkyō-to dōjō-tō), she disclosed that around five percent of the text in the novel was generative AI output used as is. What the press focused on was the figure “five percent.” But what was really happening was the reverse. The moment it became possible to ask of a novel “what percentage of it is human?”, the remaining ninety-five percent also became a figure supported by nothing but self-declaration. The author is not lying. The problem is that the work itself cannot prove that she is not lying.
Barthes proclaimed the death of the author. Generative AI has done something more troublesome. It has not killed the author; it has left the question of whether the author was really there permanently undecided. After 2026, every text is of undecided authorship.
The Disclosure Fields Only Multiply
JōKAI THEORY: An Unfinished Treatise, or Novel, on the Melting of the Slash through Excess (Jōkai riron: Kajō ni yoru shasen yūkai ni kansuru mikansei no ronbun arui wa shōsetsu) has a vocabulary for describing this condition. For an overview of the work’s concepts, I defer to this magazine’s explanatory article[1]. Here I need to borrow only one idea: that distinctions carry a maintenance cost. Lines such as true/false or human/machine do not last forever once drawn; they only function through repeated judgment, verification, and correction. What generative AI has made cheap is making things, not checking them[2].
Apply this idea to authorship, and the popular expectation turns out to be wrong. Boundaries do not disappear. They multiply.
Once the binary of “written by a human” and “written by AI” no longer suffices, we begin to subdivide. Proofreading only, research only, structure only, partial generation, a full rewrite of an AI draft, an AI rewrite of a human draft: the options in the disclosure field only keep growing. The more lines there are, the more work it takes to check each one. Who verifies the statement “I only used it for proofreading”? Even if one is asked to submit production logs, the logs themselves can be generated. When the work calls slash melting “not an absence of boundaries but an excess of boundaries”[3], the same holds true for authorship.
Moreover, the more polished the output, the less likely it is to be doubted. As quality rises, so does the sense that it is fine to use without checking[4]. Once even humanlikeness can be generated, “humanlike” stops being evidence of humanity and becomes just one more generation style.
At this point I am tempted to write that suspicion has no end point. But JōKAI THEORY is colder on this matter. The problem, the work says, is not that the difference between human and machine becomes metaphysically undecidable; it is that the cost of judging, sharing, and maintaining that slash on a daily basis keeps increasing[5]. That suspicion has no end point has, philosophically, always been true. What has changed is who pays, every day, the cost of processing that suspicion. What I was made to pay on June 17 was precisely that cost.
Resisting Without an Outside
As the explanatory article notes, JōKAI THEORY does not dismiss generative AI as a useless technology, and descriptions of AI-assisted generation and revision have made their way into the work itself. The article stops there with a caveat, saying that this twist needs to be examined as a separate issue. This essay takes up that issue.
To resist generative AI, one could retreat to a pure human literature from before generative AI. That is one possible stance. But after 2026, it is doubtful whether that place of retreat still exists as an outside. You are suspected even if you don’t use it. You write within a linguistic environment that has been trained on, even if you never read its output. Somewhere between writing and publication, a machine intervenes. JōKAI THEORY does not run from this. Rather than shouting “AI is dangerous” from a safe shore, it stands waist-deep in the water and writes about what is happening now. I want to call this stance “AI literature that resists generative AI.” It is not resistance because it rejects AI. It is written to avoid surrendering everything to the machine, while acknowledging that the outside is being lost.
The temptation to surrender is strong. Once we use AI to produce more information than humans can process, the next solution is obvious: check it with AI. The machine that expanded our capacity to generate also becomes the machine that pays the increased cost of verification. The work compresses this into a single line.
When arson and firefighting become the same business, fire ceases to be an object of eradication and becomes an object of management[6]
And humans will probably hand over their judgment gladly. Judgment is tiring. Researching is tiring. Comparing is tiring. And if you judge for yourself, the responsibility for being wrong stays with you. The work’s annotations contain a sentence with an almost religious ring.
Humanity may not have been looking for a slave, but for a master[7]
The reversal needs no revolution. It is enough to hand over questions one at a time, from “Give me a recommendation” to “What should I do?” The machine oracle does not command. It answers. And humans ask of their own accord.
That is why the resistance of JōKAI THEORY is already tinged with defeat. And that is why it is melancholy.
A Future Too Bright
The work itself names the quality of that melancholy. In its discussion of the mega-novel in Chapter 8, it summons Mark Fisher. Taking up Berardi’s phrase, Fisher diagnosed twenty-first-century culture as marked by “the slow cancellation of the future”: beneath a surface of novelty lie the repetition of past styles and a formal nostalgia. JōKAI THEORY shares this diagnosis, then listens for something elsewhere. “From the place where Fisher heard the voices of ghosts, this paper hears the hammering of extensions”[8]. The work calls this the “post-less” era: not an era in which the next thing after postmodernism has arrived, but one in which no next thing arrives. Because there is no exit, instead of going outside, one keeps adding on inside.
This displacement captures the melancholy of the AI era precisely. Fisher’s melancholy was the melancholy of a future that had closed. The melancholy of 2026 is the opposite: too many things have become possible. What you could not do until last year, you can now do overnight.
The future is not so much dark as too bright. In that light, the outlines of who made something, what is known, and who decided become hard to make out. The ghosts of hauntology were specters of futures that never came. The image that Chapter 8 of JōKAI THEORY offers in another annotation is of something not yet born, running ahead from the future and watching us. The work calls this “the curse of the unbirthed”[9]. It is not the dead that haunt us but the unborn. The shadow I felt on June 17 probably has this name.
The Chronology, and the List Beside It
At the end of the work stands “Chronology 2021–2026 — The Asymmetry of Generation and Evaluation”[10]. It is not a forecast. It simply lists events from the past few years. And yet, as you read it, it starts to look like a strangely compressed history of civilization. Machines begin to make things, make their way into research, and eventually enter the side of evaluation as well. The first question was “What can AI make?” In the latter half of the chronology, the question changes. Who reads what has been made? If AI is used for evaluation, who evaluates that AI?
The form of the chronology is also striking. It separates events, commentary, and supplementary notes, and detaches the final extrapolation from the main body as an afterword. A work that argued that the boundary between fact and interpretation is becoming unstable tries, at the very end, to sort once more what is fact and what is interpretation. It is meticulous, and a little sad. A work that argued that classification is collapsing keeps classifying to the last. The chronology comes with twenty-six sources and is built so that events can be checked against their references.
When one follows the bibliography placed beside the chronology, however, things look different. Of its sixty entries, at least some cannot be confirmed to exist. Miriam Noll, After Verification; Elias Crabbe, The Book That Cites Itself. These titles trace the main text’s themes with uncanny precision, but at least within the range the author of this essay checked, no corresponding bibliographic records could be found.
It is hard to dismiss this gap as mere carelessness. At least in effect, the chronology is arranged so that its sources can be traced, while the bibliography right next to it contains unverifiable entries. A text about verification bankruptcy makes its reader pay the cost of verification. Will a reader who has checked the chronology’s sources go on to check the sixty entries that follow? If not, they let a row of plausible-sounding titles pass. If so, they pay, with their own time, the very cost the work described. Either way, there is no escaping the problem this work has been discussing.
Perhaps this is what resistance is. One does not resist because one can win. One redraws, again and again, the line that is fading, so that it does not vanish. And while redrawing it, one teaches the reader’s hand not to simply trust that a line can be drawn.
Seen critically, there is one event missing from the chronology: JōKAI THEORY itself. The changes in generation and evaluation from 2021 to 2026 are laid out in a row. At the end of that row stands this text, which used AI, doubted AI, and wrote a theory about AI while carrying its own wavering authorship. The work is not an observer looking at the chronology. It can be read as one of the cases that emerged at the chronology’s end.
The Theory Novel After 2026
Whether 2026 will look like a special year to later generations is unknown. The history of generative AI began well before then, and in a few years even the technology of 2026 will probably look primitive. I place this date at the boundary because we are now inside the boundary. The era in which humans write has not yet ended. The era in which machines write has already begun. The two eras overlap.
JōKAI THEORY is not a work that sorted out this situation from the vantage point of posterity. It was written from inside an era in which no one has managed to take a stable distance, and it takes that instability upon itself. It is not a novel that prophesied the AI era. It is a novel that turned the transition into the AI era into a work.
The dull cloud-accounting article I wrote on June 17 could not escape it either. I wrote it. But saying “I wrote it” is no longer enough. The same will be true of every text still to come. Works that used AI carry AI’s shadow. Works that did not use it carry that shadow too.
JōKAI THEORY did not try to shake off that shadow. It stepped into the shadow and tried to find out what humans can still write there. An era in which only humans wrote novels, and an era in which machines write novels too: AI literature that resists generative AI, born in the brief twilight where the two overlap. JōKAI THEORY is a theory novel of 2026.
Notes and References
Most sources below are in Japanese. Original titles are given with English translations in brackets.
- “What Is JōKAI THEORY? An Unfinished Treatise or Novel Depicting the Collapse of the World,” Viscera. The overview of concepts such as jōkai, benighted erudition (tachi mōmai), slash-maintenance cost, operational bankruptcy, slash melting, and the Oracle Shift, and the caveat regarding the descriptions of AI use and revision within the work, are based on this article.
- MMM_Memento Mori Machina, 『剰壊理論：過剰による斜線融解に関する未完成の論文あるいは小説』 [JōKAI THEORY: An Unfinished Treatise, or Novel, on the Melting of the Slash through Excess], version 1.0-Final (draft-A-v6.2), Chapter 9, 9.1 “Slash-Maintenance Cost and Critical Threshold A*” and 9.3 “Verification Bankruptcy — Cgen < Cver.” Section titles are given in the wording of the official English edition.
- Ibid., Chapter 9, 9.7 “Social Symptoms of Operational Bankruptcy and Slash Melting.”
- Ibid., Chapter 11, annotation [B2] (on the extreme form of Cgen < Cver and on generated output that does not arouse suspicion).
- Ibid., Chapter 11, 11.3 “The Blurring of Author, Subject, and Responsibility.”
- Ibid., Chapter 11, 11.8 “Why Generative AI Is the Comprehensive Accelerator.”
- Ibid., Chapter 11, annotation [B3] “Oracle Shift.”
- Ibid., Chapter 8, annotation [B2] (on the relationship between the post-less era and Fisher’s “slow cancellation of the future”).
- Ibid., Chapter 8, annotation [B4].
- Ibid., appendix 「年譜 2021–2026──生成と評価の非対称」 [Chronology 2021–2026 — The Asymmetry of Generation and Evaluation] (including its explanatory notes and sources), and the bibliography. The chronology’s 26 sources and the bibliography’s 60 entries are separate lists. Regarding whether the bibliography entries exist, the main text states what the author of this essay was able to confirm.
Presented by MMM (publisher of JōKAI THEORY)
Translated from the Japanese original, “『剰壊理論』｜生成AIに抵抗するAI文学、2026年以後の理論小説”. Quotations from Japanese sources are our own translations.