Introduction
I’m writing this on the three-year anniversary of a major event for indie publishers.
On February 15, 2023, Neil Clarke, Award-Winning Editor of Clarkesworld Magazine, documented a flood of AI submissions, leading to the painful decision a few days later to temporarily close online submissions. There was such energetic discussion that Neil also closed comments on his blog post (https://neil-clarke.com/a-concerning-trend/).
A few months later, Sharon Renée Thomas, editor of The Magazine of Fantasy & Science Fiction, temporarily closed online submissions for the same reason (https://fandsf.moksha.io/publication/fsf).
That was followed by a flood of closures: sudden “indefinitely closed” notices with no timelines, and submission portals stating new policies about not accepting AI-generated or AI-assisted manuscripts (like WHP). There was also a shift toward editors moving to trusted pipelines (e.g. friends from MFA years) in an attempt to reduce exposure to risk and avoid the detection arms race (see https://www.schneier.com/blog/archives/2026/02/the-ai-generated-text-arms-race.html from a few days ago).
Magazines say it out loud: AI is clogging up their large submission pipelines and costing money for processing – money they don’t have to spend. Book publishers are more cautious because they have lower submission volume and they don’t want to drive away legitimate authors by closing portals due to flooding by AI submissions.
The purpose of this essay is to lay out the current situation with AI in publishing near the beginning of 2026, explain why the problem has changed so quickly, and estimate where we might be heading with literary creation and publishing.
What’s happened in the three years since 2023?
AI writing got good – fast. That’s what happened.
Three years ago, most of us in the editorial and publishing worlds thought poetry would be the Mount Everest for AI writing. We were seeing all kinds of painfully obvious problems even a year ago: overused abstractions collapsing into irritating cliché, metaphors cast about like bird seed hoping something germinates, complete failure to appreciate silence and hinting.
Now, some large language models (LLMs) are producing elegiac poems in a high literary register that are shockingly good. This is not incremental improvement, based on tweaking code and training procedures. This is a qualitative leap forward.
Short-form poetry is easiest for LLMs. Long-form writing – both poetry and fiction – is more difficult. I asked an LLM for a technical explanation of the difference. It helpfully supplied three lines of reasoning (which I have condensed and reorganized).
First, poetry benefitted disproportionately from LLM technology improvements. “Poetry is short-form, dense, and highly sensitive to rhythm, image, and register. Once models crossed certain thresholds in token-level probability control, human preference alignment, and fine-grained stylistic evaluation, poetry improved faster than long fiction. Recent LLMs show dramatically higher human preference scores in ‘Literature & Art’ and ‘Creative Writing’ benchmarks.” (See https://llm-stats.com/benchmarks/category/creativity.)
Second, it’s not just a matter of technical cleverness, because humans got involved in deciding how to target expensive training routines. “Poetry became an explicit optimization target. By late 2025, poetry and lyrical prose were explicit benchmark categories. Models were evaluated by LLM-as-judge systems trained to approximate literary taste. Open models (Qwen3, DeepSeek, Llama variants) were deliberately tuned for poetic quality. Poetry stopped being a side effect and became a test case.” (See https://www.siliconflow.com/articles/en/best-open-source-llm-for-literature.)
Third, short forms of poetry hide the remaining weaknesses of LLMs. “LLMs still struggle with long-term thematic development, moral consequence over time, and deep narrative transformation. Poetry largely avoids those traps. As a result, poetry output now masks structural limitations that still expose AI prose in longer genres – a pattern confirmed in human evaluation studies of literary outputs.”
The result is a new level of baseline behavior: “metrically consistent verse, credible free verse, effective lineation, controlled enjambment, image clusters that cohere across a poem.” The LLMs benefiting from focused training “use restraint rather than maximalism; and allow gaps, pauses, and understatement; they pass a first serious read and feel workshop-ready.”
Machine poetry ain’t perfect: “despite the shock, there are stable failure modes. AI poems still describe grief without aftermath, invoke embodiment without cost, and reference trauma without ethical gravity. The poem knows the experience, but has not been changed by it. AI poetry rarely risks being wrong, avoids embarrassment, and does not stake identity, relationship, or belief. Human poets still write things that might fail socially or morally. Models don’t. Strong human poems often undermine their own metaphors, reverse emotional claims mid-stream, leave unresolved tension. LLMs tend to resolve too cleanly – even when aiming for ambiguity.”
Pity poetry editors. First-pass reading has been the mainstay for editorial filtering but is no longer reliable. We used to say, “Sounds like AI” but that’s an outdated heuristic. Poetry is now the genre where AI involvement is most difficult to detect. But commercial LLMs are nothing if not quick to offer client-focused reassurance: “The shock is real, and it’s justified. Poetry crossed a capability threshold faster than almost anyone predicted – not because machines became poets, but because poetry turned out to be structurally easier to simulate than we assumed.” Thank you, dear LLM, for empathizing with our shock.
In the last eighteen months, book-length nonfiction leapt ahead, in part because LLMs can hierarchically organize a book into parts and sub-parts that are relatively independent, and they’re now very good on the chapter scale. Short-form fiction, similarly, has improved dramatically. The real sticking point turns out to be long-form fiction.
Long-form fiction requires a lot of long-term memory, character transformation, thematic recursion, causal integrity, and the endings have to justify the rest of the book. LLMs have improved in all those ways but they are still struggling, for now at least.
Okay, so Mount Everest turns out to be long-form fiction, not poetry. But it’s still more surprising than we can easily express to read machine-generate poetry that is strong in craft and emotionally moving, machine poetry that reaches us.
What will happen in the next three years?
A dear friend, a scholar suffering from creeping dementia, shared with me that his memory had fragmented to the point that he could not maintain the entire argument of a book in his mind anymore, so he had to focus on essay-length writing. That tragic moment, seared into my mind, is the reverse of what’s happening with LLMs.
We’re already seeing LLMs with huge context windows – hundreds of thousands and in some cases a million tokens. This means an LLM can hold a ton of information in memory – and keep everything properly prioritized. We’re seeing strong competence in chapter-length writing, and even multiple chapters in sequence.
For nonfiction, this means consistent focus on the arc of argument and supporting evidence. As context windows get longer and LLM planning improves, remaining limitations of this kind will steadily (maybe suddenly) disappear. And skillful direction of well-trained LLMs can already produce complete nonfiction books on all manner of subjects.
For fiction, this means LLMs can optimize continuity of characters, remain consistent in plot, and avoid straying into the kind of nonsense or rambling or contradiction that destroys a story (LLM drift). Novella-length fiction is a strong point for AI now, and getting stronger quickly.
The big limits are mainly in long-form fiction. They revolve around the reality that depicting transformation and meaning in fiction isn’t just a matter of larger LLM context windows and clearer internal prioritization of LLM agent activities. It’s characters who change, because of events, sometimes in contradictory ways, and often with unclear resolutions as to self-understanding and behavior. It’s events that collide with the values characters hold dear, leading to axiological tangles that persist, that resist harmonization, that confute character consistency.
But can’t LLMs learn to do that, too?
Yes, of course, and they probably will get there in the next three years. What’s slowing things down is the importance of human-reinforcement feedback to support LLM learning. Human judges don’t agree with each other on fiction, the very meaning of excellence shifts from genre to genre, and the time and expense associated with nudging the LLM training process with real human evaluators is large.
So it turns out that LLMs solved the poetry problem before the long-form fiction challenge because it’s cheaper. Sigh.
I asked an LLM about what editors and publishers can expect in the next three-to-five years. Here’s what I got. “2026–2028 (near term): Strong chapters, weak books. Credible novellas with editorial cleanup. Highly fluent popular nonfiction sections. But NOT reliable endings or deep character evolution. 2028–2031 (mid term): Book‑length coherence with human scaffolding. Genre fiction that survives close reading. Nonfiction that is stylistically publishable but still fact‑fragile. And increasing difficulty distinguishing AI from human drafts. 2031+ (long term): AI can co‑author long works at near‑human fluency. But fully autonomous literary mastery remains questionable, and lived consequence, moral risk, and existential stake will remain weak.”
In case you missed it, yes, five years is long term now.
This assessment strikes me and others I’ve talked to as credible. If anything, it’s underestimating how quickly the training problem might be solved. So pull up your britches, my editorial and publishing and writing friends. Superb machine writing is almost here.
Reflections on Steady State for Creative Writing
I don’t have a crystal ball but…
Not very long from now, most human beings will no longer rely primarily on writing as a way to learn, as we have done for several centuries. Most of us will be more like project managers than generators of prose, in almost every part of the economy. This is already true as businesses push their employees to use GenAI as much as they can to improve efficiency and quality of output.
The exception – yes, there will be an exception – is people who pursue writing at a high level. They will hone their craft as a specialization, a form of personal self-expression, coached by AI and human tutors with superb writing capabilities. They will be like great artists, designers, and musicians: they create for the joy of the act and the satisfaction of the achievement despite the perplexing pain of the process.
Consider chess. No doubt some say, “What’s the point in becoming excellent at chess (or go, or any other game) when AI can beat the best humans?” But many chess players play the game for intrinsic rewards: the wonder of a world of 64 squares of mystery and tactical excitement, the joy of visionary plans implemented in a tiny complex system, and the profound pleasure of forging intuition under the hammer blows of rigorous analytical training.
The true writers among us will always write, no matter how good the machines become. It’s a calling, not a chore: some people are drawn, even compelled, to express themselves in writing. It’s true for cooking, music, drawing, and sports: the exuberance and intensity of the moment is more important than being the best. A few people can worry about being the best cook in the world, and let them worry. The rest of us can step into the fifty-thousand-year history of food preparation and take pleasure and satisfaction in it.
That is one fixed point in the wobbly rope of AI and literary creation: humans do it for love of the art.
Another fixed point is that machines won’t stop getting better just because we feel uncomfortable about humans no longer being uniquely valuable for our creativity. When (not if) machines become superintelligent, which means better than human beings at virtually everything cognitive, including literary creation, social constraints will still matter. A few (not many) readers will want pain-honed literature. Some publishers will want human provenance guaranteed.
Recent projections agree that we already have AI-human symbiosis in content creation (the language of https://publishingstate.com/how-generative-ai-will-shape-publishing-in-2026/2025/), and that will only get stronger. There’s something to embrace in that symbiosis, and Forbes and WriteStats go there with cautious enthusiasm (see https://www.forbes.com/sites/forbesbooksauthors/2025/01/27/ai-in-publishing-a-2025-industry-forecast/ and https://writestats.com/what-ai-changed-in-2025-how-generative-tools-shifted-the-publishing-landscape/).
But symbiosis is middle ground. In fact, the market will split, probably in three ways.
On one end, autonomous AI with little human oversight will produce literary, artistic, musical, and cinematic productions customized to user demand – a level of bespoke creation never seen before. This is the democratization of the artistic commission, formerly the privilege of extreme wealth, and will revolutionize the entertainment industry as well as publishing. The watch phrase is “good enough” and the aim is not to sell a particular story, but to sell to a particular person, by giving them exactly what they ask for. It’ll be cheap, fast, amazing, and economically huge. I suspect it will also accumulate weariness, boredom, and self-loathing, as consumers find their needs satisfied with no more effort than asking for something they want.
In the middle ground, throughout the business world, and also as a form of personal expression, the AI-human symbiosis will be as rich as we want to make it. Grandpa always wanted to write a memoir, and now he can, without needing to hire an expensive literary genius to ghost-write it for him. Hopefully he remembers to mention the symbiosis thing. It’ll involve effort, but not much writing craft. Prompt engineering and creative direction are skills, to be sure, and Grandpa’s memoir will be better if he’s good at them, but they’re not creative literary generation.
On the other end we’ll see humans writing for humans who want to read human writing for which humans are accountable. AI might be involved in author training, in brainstorming character dictionaries and plot lines like a writing buddy, in sticking to a schedule like an accountability partner, or in proofing. But the pain of the writing journey is still felt through every cell of the author’s body, their lifeblood dripping into legible characters on the page. This path will be small, expensive, thorn-strewn, yet lovingly walked by a few hardy souls – not because they hate the machines, and not because they feel compelled to compete with them, but because they adore literary effort and the journey of mastering the page.
This tiny part of the steady-state literary economy expresses what publishers have always assumed they were doing – even the giant publishers, and certainly little indies. But the industry is changing under the pressures I’m describing. The game is shifting from writing as mass production to writing as the boutique performance of an irreducibly human identity craft. Live music survived streaming. Live theatre survived Netflix. Human literature will survive AI.
Coexistence of these three sectors of the steady-state literary rapprochement between humans and machines has already begun. There will be battles along the way, beginning with publishers fighting, probably in vain, to assure themselves on the provenance question; writers fighting to force AI companies to pay for using human literary productions in LLM training; and nasty accusations flying in every direction about immoral machine sellouts and irrational elitist literary snobs. It’ll be messy, sure. But the steady state I’m describing is coming, ready or not. And it’s closer than we might think.




This is a fascinating post, Wesley, and my gut and my brain tell me that you are pretty right. I guess I am confident enough that there will always be an interest by humans of reading what other humans have to say, and there will always be high-quality human writing. And I believe that the intrinsic rewards make writing or any other creative and flow activity (like chess) worth doing for those who are interested, regardless of what machines can do.
Yet as a college professor, like you, I'm especially worried about this: "Not very long from now, most human beings will no longer rely primarily on writing as a way to learn, as we have done for several centuries." I suppose that humanity will make that transition somehow, but I can't imagine what it will look like. I fear that there will be lot fewer thoughtful and creative human beings when they all grow up with weak engagement with both writing and reading.