On delivery, discovery, and the question AI actually raised
Almost every essay I’ve read about AI and writing knows exactly where it’s going. The structure is announced in the first paragraph, executed through the middle, confirmed at the end. This one tries not to — though I’m aware that trying not to is still a form of intention, and that awareness might be the problem from the start.
I’ve written about this before, from different angles — about why film criticism still matters in an age of language models, about the doubt AI writing starts to cast on human writing simply by proximity. What I want to say now is something more fundamental. About two kinds of good writing. And about which question, at this particular moment, we should actually be asking.
Two kinds of good writing. I keep wanting to complicate this binary, but it keeps holding.
The first kind knows where it’s going. It takes you somewhere efficiently: clarifies, organizes, argues, arrives. The prose serves the logic, and the logic serves the destination. Most journalism is this. Most criticism. Most of what circulates online under captions like “this is exactly what I needed to read.” Call it delivery writing. It delivers.
The second kind doesn’t know where it’s going, and neither do you. You’re reading and something shifts — an image arrives without announcement, a connection forms that wasn’t signaled, a thought that began as an argument about cinema becomes, halfway through a paragraph, an unexpected observation about grief, and from that, a clarity about something else entirely that you didn’t know you were confused about. The writer got lost. You got lost with them. At the end, you’re both somewhere that wasn’t on any map at the beginning.
Call it discovery writing. I’m not sure the category holds.
I was thinking about this while trying to learn a piece of music by ear a few years ago — the specific uncertainty of reaching a note you’re not sure about. Not whether you’re playing it correctly. Whether what you hear in your head is what you actually hear, or what you expect to hear because you’ve already internalized the recording. The two felt different from each other, but distinguishing them required a kind of attention that kept failing just before it resolved. I kept both versions for months. Eventually I couldn’t say which one had come first.
I don’t think this is a metaphor for anything in particular. Or maybe it is, but not for what it seems.
The recent controversies in literature tend to circle the wrong question.
Olga Tokarczuk — 2018 Nobel laureate, still one of the most significant novelists working in Europe — said publicly that she uses AI chatbots for research and idea organization. What she clarified afterward was that the writing itself — the sentences, the structure — she still does herself. AI helps her prepare, not compose.
Rie Kudan, who won Japan’s Akutagawa Prize in 2024, went further: she acknowledged that portions of her novel were written directly by ChatGPT, and she named this openly, without apparent apology or performance. What she also said — less widely quoted — was that the seams had become invisible to her. She couldn’t always locate, afterward, which sentences had arrived through which process.
A British writer’s award-winning book was flagged at 100% AI-generated by detection software — though such software remains deeply contested in its reliability, which is itself part of the problem.
What all three cases produce is the same debate. Authorship, attribution, acceptable percentages. Where the line is.
I keep returning to a different question. Not who wrote which sentence. Whether the writing is good. What good means. And whether our working definition of good still holds, now that those two questions have started to pull apart.
Delivery writing was always, at some level, information management. Not a dismissal — good delivery writing is genuinely difficult. Clarity of argument, precision of structure, logic that holds under pressure, the sentence that says exactly what it means without surplus: all of this requires real skill. But the core function is transmission. Something needs to get from one mind to another. The writing is the vehicle.
If that’s what writing primarily is — a vehicle — then AI has become the better vehicle. Coherence of argument, structural logic, readability, organizational clarity: across these dimensions, a well-prompted language model consistently outperforms most human writers working under deadline. Not occasionally. Consistently. The gap is not small.
Most writers resist the logical conclusion — if delivery writing is what we primarily valued, we should use AI for it and stop apologizing — not because they’re in denial about the technology, but because the resistance points at something real. Something else writing was doing. Something that isn’t transmission.
Discovery writing is inefficient by design. That inefficiency is the feature, not the flaw.
Human thought, moving through a problem, doesn’t follow the optimal route. It circles back. It follows leads that are interesting rather than correct, for longer than it should. It makes connections between things that have no business being connected — and sometimes, not always but sometimes, exactly those connections are where something new opens. A sentence starts as an argument about cinema and ends as a memory of a specific summer afternoon, and that memory, completely unplanned, turns out to be exactly what the argument needed. The detour was the point.
Human cognition is constitutively biased, obsessive, and accidental. We fixate on the wrong thing. We return to an image three times without understanding why. We build arguments around things we feel before we understand them.
Here’s where I want to say something that might not hold up.
A significant portion of what gets called discovery writing is probably just poorly structured delivery writing that we’ve learned to romanticize. The beautiful non-sequitur. The productive digression. The essay that gets magnificently lost. Some of that is genuine — human thought reaching somewhere by an inefficient route it couldn’t have reached directly. And some of it is a writer who couldn’t argue clearly, and we’ve been mistaking the confusion for depth.
This is the part I can’t dismiss easily.
The experience of getting lost with a writer isn’t a guarantee of anything. Being confused doesn’t mean you’ve been taken somewhere real. The most skillfully written nonsense produces exactly the feeling of productive disorientation we associate with genuine discovery. Lacan spent a career demonstrating this. His readers reported transformation. Whether the transformation corresponded to anything outside the experience of reading Lacan is a question that his admirers rarely press.
Which means the category of “discovery writing” might be, in large part, a story. Told by writers to justify structural weaknesses, and by readers to justify the effort of following something difficult. The sensation of being productively lost and the sensation of being cleverly conducted are, phenomenologically, very close. The reader who believes they’re getting lost might simply be moving through a maze designed to feel like wilderness.
I’ve been sitting with this for weeks and I don’t fully know what to do with it.
What I keep returning to — not as resolution, but as the only thing that won’t let me go — is a different way of defending the category. Not that discovery writing produces truer insights, which is probably impossible to verify. But what it does to the person following it.
Something changes in you when you read writing that genuinely doesn’t know where it’s going. Not necessarily because you arrive at truth. Because you practice a specific kind of attention — the kind that holds ambiguity without resolving it too soon, that stays with confusion without treating confusion as failure. Whether or not discovery writing produces genuine insight, reading it trains something. A tolerance for not-knowing. A capacity to follow something uncertain without demanding that it arrive.
That capacity — a tolerance for not-knowing — is one of the few things I can think of that genuinely can’t be optimized.
What AI’s logical coherence represents, in this context, isn’t simply a different strength. It’s a structural incompatibility with a certain kind of reading experience. Not because AI lacks knowledge or precision. Because following discovery requires, in the reader, the willingness to not know yet — to carry unresolved material and wait — and that willingness atrophies if you stop practicing it.
I’m not sure AI can genuinely not know what comes next. Its architecture is, at its foundation, pattern completion across everything it has seen. Whether that constitutes knowing, in the relevant sense, is a question I’d rather not answer confidently.
A pipeline has emerged in the conversation: Human → AI → Human → Writing. Seed the idea, hand it to the model for structuring and development, return to reshape. The argument is that this removes the structural weakness from human writing — we really are bad at architecture — while preserving what matters: ideas, voice, judgment.
There’s something useful in this. And a risk that doesn’t get discussed often enough. If you consistently offload structure to an external system, you may gradually lose the capacity to hold a complex thought long enough for it to become interesting. Not the ability to outline. Something stranger: the ability to live inside an argument through its confused middle stages, when you don’t know where it’s going and have to stay anyway.
The Renaissance workshop comparison keeps appearing in these conversations — the master with apprentices, vision directed and execution distributed. What it tends to skip is what those workshops actually produced: painters. People who internalized, through years of executing someone else’s compositions, how composition works when it’s working. The workshop was school in disguise. If AI is the new workshop, the question is whether it’s still a school.
I’ve started to wonder if that question is already moot.
Not because it isn’t real, but because it may already be answered in practice. Most writers I know who use AI don’t describe it as a tool that enters their process at a defined step. They describe it as part of how they think now. The tangled draft that used to exist in their heads — incomplete, interrupted by wrong turns — increasingly exists in a conversation with a model. The thinking happens there, in the exchange, before the writing begins.
Which means the question of whether this erodes capacity isn’t really about the future. We’re not asking what will happen if writers start working this way. We’re asking whether the character of thought has already changed, quietly, for people who have been doing this for two or three years. Whether the thing that was supposed to be preserved — vision, voice, judgment — is already being shaped by what AI can comfortably execute.
I don’t know this. But I’ve stopped asking the question as though the answer is still ahead of us.
Something that gets underweighted in discussions about AI writing: what it actually means to learn to write. The best writers developed their craft by reading the best writers. Not algorithmically optimized writing. Not competent average prose. Writers who were doing something almost impossible to repeat — who had built, through years of reading and obsessing and failing, a particular way of moving through language that was distinctly their own.
Reading that kind of writing slowly — following another mind’s architecture, watching where a sentence turns, observing what an unexpected image is doing in a context that seems wrong for it until suddenly it isn’t — is itself a practice. It builds something that operates below the level of conscious rule-following. Not techniques. A sensibility. A felt understanding of how thought moves when it’s doing something unusual.
Clarity is not the same as movement. Precision is not the same as discovery. You can learn a great deal about clarity from AI writing. What it can’t give you is the experience of following something uncertain — of carrying unresolved material forward without demanding resolution — and I think that experience is what changes the quality of thought you bring to your own work.
The standard most readily applied to good writing is quality of arrival. Coherence, argument, readability. These are real standards, not wrong ones.
But they’re not the only ones. The other standard — harder to articulate, instantly recognizable in the encounter — is: did this take me somewhere I couldn’t have arrived at alone? Not somewhere faster. Somewhere I couldn’t have reached by the direct route. Where I found something I didn’t know I was looking for.
That second standard doesn’t care whether AI wrote a sentence. It cares what the writing did.
There’s something I should probably say before this ends.
At some point while writing this — the middle section, I think, though the exact sequence has blurred with revision — I handed a paragraph to an AI and asked what it would do with the argument. What came back was more coherent. The transitions cleaner. The point landed harder. I put it aside and returned to my version.
I’m not entirely certain, having gone through several drafts since, which phrases are mine. I’ve tried to reconstruct the sequence. I can’t fully.
What I kept returning to was the version I couldn’t let go of. What that makes it, I’m not entirely sure.
The first standard needs to know who wrote it. The second standard doesn’t ask.