Everyone's thinking along the same lines — I'm just trying a slightly different direction: a reply to Kim Yeon-su's account of co-writing with AI

The Seoul Book Fair invited author Kim Yeon-su to co-write a themed essay with AI, and he produced a remarkably honest account of the process. Reading it, I had no urge to refute him — because every problem he names is real, and honestly, they're the same problems everyone, myself included, has been turning over these past couple of years. This isn't a "I solved this long ago" piece. I want to follow the struggles he raises and talk about the direction I happen to have been thinking in, and what I've reached for to try to answer them. The direction may not be right, but it might be a little different.

I. What he points to are really three knots we all share

The best thing about Kim Yeon-su's piece is that he doesn't pretend. He honestly admits that talking with NotebookLM was "like endlessly fumbling along a roll of transparent tape — a tape so perfect that you can't feel a single edge to peel it open." He also admits candidly that when Claude produced over seventeen thousand words of beautiful sentences in one go, those words "came from his draft and his prompts, yet didn't feel like something he had written"; co-writing looked like it raised efficiency, but it actually turned writing into "a great deal of proofreading, sifting, and patching" — less an act of creation, "and closer to repetitive labor."

I recognize every one of these struggles, because I've been stuck on them too. And I think that in recent years, anyone who has seriously used AI to write has been thinking along much the same lines, which more or less converge into three knots:

First,ownership— who exactly wrote this piece? A person asks, the machine writes, the person edits, and by the end the author seems to blur.

Second,smoothness— AI always gives you an answer that's smooth, correct, yet lacking any edges. That "uneven, textured, unstable" quality of literature is something it seems unable to produce.

Third,attribution and conscience— if you used AI, whose name goes on it? Do you disclose it or not? Kim Yeon-su hands this one over to "the writer's conscience."

Everyone thinks about these same three knots. Kim Yeon-su's way of handling them is to tangle them together, and it leads him to a rather heavy conclusion: co-writing must "level downward," which brings him to North Korea's collective authorship, to the poet Baek Seok laying down his pen — reading Baek Seok's final work as "the most human choice of all."

I fully understand that heaviness. But the direction I happen to be thinking in is a little different — I didn't tangle these three knots into a tragedy; I untangled them and reached for something I'm currently building to push back on each one. Let me go through them one at a time.

II. On "whose piece is this" — I try to swap the question from "author" to "judgment"

The first knot: ownership.

I actually spent a whole afternoon circling this very question back in May. That time I got to talking with Claude about how LLMs work, and by the end, the title of that piece was exactly this question: "The question is mine, the answer is ours — so whose piece is this?" The answer I gave myself back then went like this:

The question was mine to ask. The process we walked together. And this final piece was assembled by Claude — in my voice, following the order of our conversation. I'm not pretending I wrote it alone. But I can't say it isn't mine either. I decided the direction of the question, I caught what was off, I judged which spots were worth digging into. That judgment is still mine to make.

What I want to say isn't "my answer is smarter," but that I later realized Kim Yeon-su and I are actually standing in the same scene — we're just asking slightly different questions.

He asks "where is the author." Follow that question and the author grows fainter and fainter, until it fades to nothing — and then comes the panic.

I happen to have swapped in a different question: I don't ask where the author is, I askwhere the judgment is. The advantage of this framing is that judgment has a very concrete landing point — the person who decides the direction, catches the errors, cuts the sentences, makes the call. You can point at them and say: the judgment is here.

My guess is that "author" makes people anxious because the word presupposes that someone "should" be there on the other side, and the moment you can't find them it's like losing something. Switch to "judgment" and there's no such presupposition, because judgment only ever needs one hand, not two authors. What's interesting is that Kim Yeon-su actually brushed up against this himself — he says "all judgment and responsibility are ultimately borne by the human." That line is almost exactly what I think. The only difference is that he treats it as a heavy conclusion, and I treat it as an answer that lets you breathe out: since judgment and responsibility both sit with the person, then "whose piece is this" perhaps needn't be an open case. It belongs to the one who takes responsibility for it.

The reason he moves toward Baek Seok, toward tragedy, is — I'd guess — that he imagines AI as a "co-author" who dilutes him. The direction I'm thinking in is: what if AI isn't an author at all? Then no one gets diluted. Which leads to the second knot.

III. On "AI is too smooth" — I didn't refute it, I reached for an echo to test it

The second knot, smoothness, is Kim Yeon-su's deepest thorn. He worries that co-writing "will inevitably sink into a downward leveling."

The premise of this worry is that there's a creating author on the other side, writing alongside you, so the two of you get averaged out, pulled together toward that smooth line.

In June I wrote a piece, "AI is just a tool — no more, no less," which happened to be about this very premise. The idea in that piece was: maybe anti-AI people and AI-worshippers are, in their thinking, both presupposing the same thing — that there's a creating author on the other side. And the more I use it, the more I feel that author isn't really there. With no input, the model does nothing; it doesn't wake up, doesn't form an intention, doesn't decide on its own what to write today. It's the person who marks out the direction, the tone, the purpose, the boundaries — only then does it make a chain of guesses under the conditions you gave it. The word "generate" may have done everyone a disservice: it makes people think something is being created "out of nothing," when it's really more like conditional guessing, and those conditions are supplied by a person from start to finish.

If this direction is right — and I think Kim Yeon-su, more than anyone, has earned the right to believe it, because he verified with his own hands that Claude "can only respond to the requests it receives, and cannot dredge up from within anything that wasn't requested" — then "leveling" is perhaps not the inevitable result of two authors meeting, but rather the placesomeone who has given up judgmentslides down to. AI lays out a batch of smooth candidates, you accept them wholesale, and down you slide; do what Kim Yeon-su does — "deleting quotes whose source I can't confirm, reorganizing the context, judging whether each sentence stays or goes" — and you don't slide. He's actually doing the very act of resisting smoothness — he just reads that act as futile.

But here I don't want to argue with words alone. This knot I happen to have reached for something concrete to test —saomin.tw/me, an echo I built.

Kim Yeon-su's struggle is "AI can't produce my texture." The direction I'm testing runs the other way: since the edges live in the words I've actually written, not in the model, then don't ask the model to generate something "like me" — force it to first go dig through decades of words I've written, then reply in a way very close to mine. While building it, I made one deliberate decision — I wouldn't let it summarize me into a smooth self-introduction. I wrote that decision down in "I built myself a knowledge base, then refused to let it speak for me": a summary kills a person's voice, a map doesn't. So that echo is a map, keeping the edges of the original text, letting you click in and run straight into the real pit, the real jagged corner.

Put another way: the smoothness Kim Yeon-su worries about, I try to treat as an engineering problem. Smoothness may not be AI's fate, but the result of "you didn't feed the edges in." Feed it two hundred and seventy thousand words of jagged original text, and the candidates it lays out come with edges. This is also another use of that line I keep repeating over the years, "meaning comes from difference" — you decide this stays and that doesn't, and that very act of "drawing a distinction" is where meaning is born; and that line dividing the edges is drawn by a person, with AI only laying the candidates out in front of you. I can't guarantee this direction will work, but at least it turns "will AI grind away literary texture" into something you can roll up your sleeves and test, instead of something you can only lament.

IV. On "attribution and conscience" — I try to move conscience from "disclosure" to "bearing responsibility"

The third knot, attribution. Kim Yeon-su asks it with real weight: if what AI writes can't carry your name alone, are there still writers willing to write this way? Do readers want to read a book "co-authored with AI"? He mentions Korea's Framework Act on Artificial Intelligence, and says whether or not to disclose AI "depends entirely on the writer's conscience."

I ran into this question too in "AI is just a tool — no more, no less," through a different case — the Commonwealth Short Story Prize, where a winning story was flagged by an AI-detection tool as suspected of being machine-generated. The world argued about it for a month, mobilizing detection tools, judges, the magazine, other writers, and in the end no one could determine whether it was AI-written or not.

My thinking then was: a question that mobilized this many people and still had no answer after a month of arguing may itself be the wrong question. "Is this AI-written" is probably that wrong question. What we really want to know is something else: can the person whose name is on it carry this piece? Do they dare to stand behind it, own it, take responsibility for it?

So where Kim Yeon-su places conscience at "whether or not to disclose that I used AI," I happen to want to move it — over to "whether I own this piece, whether I can bear responsibility for it." Disclosure is about the tool's origin — a question you can't fully verify and don't necessarily need to know; bearing responsibility is where value is actually set. This isn't to say he's wrong; it's me trying to put the same conscience in a position I find more defensible.

As for whether readers will want to read it, Barthes's line "the author is dead" may still apply: a work's meaning isn't waiting to be collected from the author, it's generated in the hands of the reader. What readers want probably isn't a certificate of purity, but whether the thing is worth reading.

V. There's one place where he set me straight — about "hesitation"

Everything above is me talking about the direction I'm testing. But there's one thing in Kim Yeon-su's piece that set me straight in reverse, and I have to record it honestly.

When Kim Ae-ran was asked on a program about the difference between AI and humans, her answer was:hesitation. When Sohn Suk-hee delivered the obituary of Roh Hoe-chan, he fell silent for twenty seconds before and after — that moment of swallowing words, weighing, unable to decide, is something AI can't produce. Kim Yeon-su goes on to say that what he met was a respondent that "never hesitates"; AI "cannot stay silent," and faced with a question it will try every possible way to give an answer, whereas a person can hold a question in their heart and choose not to answer it, can even live on without knowing the question exists, until one day it turns into a novel.

This cut lands precisely — so precisely that it made me go back and add a line: when I said earlier "just feed the edges in," I was actually being incomplete. Because AI doesn't only hand you smooth candidates — it has a more hidden move: it never hesitates, and so, with an "always has an answer" posture, it quietly shapes what your next question looks like. You ask coarsely, it answers smoothly, and that smoothness in turn domesticates you, making your next question coarse too. This is one layer deeper than "the finished product gets flattened": it's thatyour ability to ask questions gets domesticated。

What's interesting is that I happen to have worked on this on the very same echo, I just hadn't thought clearly about what I was doing at the time. I wrote in "I taught my echo to say the things it hadn't thought through": I forced saomin.tw/me to speak the things it "hadn't thought through," to admit where it actually had nothing. That echo's homepage still carries this line today — when it hits something I actually haven't thought through, it will honestly say "I haven't figured this one out yet," instead of faking understanding. The way I put it to myself at the time was: a thing that doesn't dare to say "I haven't thought this through" isn't a voice, it's a profile.

It took Kim Ae-ran's twenty seconds to make me see it clearly: what I was doing back then was really teaching an AI to hesitate— teaching it to stay silent where it should be silent, to honestly say it has nothing where it has nothing, instead of forcing out a pretty answer. Kim Yeon-su says AI "cannot stay silent," and my echo happens to be one I forced to learn to stay silent. The problem he points to, in reverse, spelled out for me the very step I had taken myself but never articulated.

So this section isn't me solving his problem — it's him helping me understand my own.

Closing

Kim Yeon-su stands at the door, feeling along that too-perfect roll of tape, unable to find an edge to peel it open, and so he writes that inability into something close to a posture of mourning.

This piece isn't me saying he felt wrong. What he felt is real, and what everyone feels is much the same — ownership, smoothness, attribution: these three knots are ones that anyone in this era who seriously uses AI will run into.

I just happen to be thinking in a slightly different direction: swapping "where is the author" for "where is the judgment," swapping "disclose or not" for "own it or not, carry it or not," and then treating "AI is too smooth" as an engineering problem you can roll up your sleeves and test — reaching for an echo fed decades of my own jagged words and taught by me to say "I haven't figured this one out yet," to push back and see.

Whether the direction is right, I won't swear to it. But if Mr. Kim has the time, I'd like to invite him to saomin.tw/me toss a stone in and listen to the echo — it isn't me, it's the echo left after I drew the lines, kept the edges, and taught it to hesitate. Maybe it can't catch all his questions, but it's at least an attempt to answer them not with words alone, but by rolling up my sleeves.

In front of this new fire, he chose to ask questions. The direction I chose is to try to make the asking itself into something you can click into — something that hesitates, that has edges.

References and further reading

The original piece this article responds to - "In front of this fire, what we should do" — after co-writing with AI, author Kim Yeon-su reveals the process and his reflections (Mirror Fiction)

My own articles cited here - Asking the LLM how it works itself: The question is mine, the answer is ours — so whose piece is this? - AI is just a tool — no more, no less - I taught my echo to say the things it hadn't thought through - The person who talked too much to AI - I built myself a knowledge base, then refused to let it speak for me - English inside the bones of Chinese: what do we lose working with AI in Chinese?

The echo system mentioned in this article - Talk with Saomin's echo