Can Large Models Be Creative? An Answer from a Programmer Who Wrote Web Fiction for Three Months

From May to July, I used AI to write a novel on Fanqie — from outline to full text, from signing the contract and going live to pushing out 4,000 characters a day, I ran the entire pipeline end to end.

The book’s numbers were dismal, but I wasn’t discouraged in the slightest — because this was an MVP validation from the start. The question I actually wanted to answer mattered far more than the numbers: Does AI really have creativity?

Let me give the conclusion up front: No. Under the technical principle of autoregressive prediction, it simply cannot produce creativity.

1. The Devil Is in the Details: The Real Mess That Is AI Novel Writing

If you just ask AI to dash off an opening — sure, it can produce text that looks decent enough. Fluent grammar, normal narrative pacing, even the occasional passable bit of dialogue.

But when you ask it to write a 400,000-character novel — with seven or eight recurring characters, foreshadowing planted across 10 volumes that has to pay off, comedic reversals and emotional climaxes —

the disaster begins.

1. Weight dilution.

Once you get past chapter twenty or thirty, the settings from the middle stretch feel watered down. It’s not that it “forgot” — it’s that in the process of repeatedly computing probability distributions, the weight of the middle section gets diluted. The protagonist’s personality traits that you spent a huge amount of space defining in chapter one have become a vague outline by chapter thirty, because the token distance is too far and the probability weight has decayed to the point where it no longer affects the output.

2. Treating the symptom, not the cause.

You tell it “this dialogue is too stiff,” and it smooths the dialogue out — but at the same time it wipes out all the foreshadowing and character hints buried in that conversation. Because it doesn’t understand why the dialogue was written that way; all it cares about is whether the “stiffness” you asked it to fix has been resolved.

3. The more you revise, the messier it gets.

As you write further in, you find yourself adding more and more rules to constrain the AI — “the protagonist must be like this,” “the dialogue can’t be like that,” “emotional expression has to work this way.” But these rules dilute each other and conflict with each other. You tell it to watch out for A, and it forgets B. Constrain A, B and C at the same time, and overall precision falls off a cliff.

GPT also burned me once, badly — it swore up and down that on a certain historical timeline, the protagonist couldn’t possibly make it from Shouchun to Wancheng by that same night, and had me change the location settings for the entire third volume. Only later did I discover that the correct historical fact is exactly that you can race back from Shouchun to Wancheng — GPT had simply blown the “that same night” time constraint out of proportion. The plot framework for two whole volumes was wasted because of one wrong judgment from it.

Lesson: Next time the AI revises the outline, verify the historical timeline by hand.

2. How I Reached These Conclusions

Before getting into first principles, let me lay out what I actually did. I didn’t dare pass judgment after reading a few articles — I took AI into real, live combat for three months.

On tool selection, I took one long detour.

The first thing I used was Inkos — a multi-agent novel-writing framework that can automatically plan chapters, generate text, and polish out the AI flavor. I was excited at first, but quickly found it was a black box — the setting bible I fed it often wasn’t recognized, battle scenes got blocked by MiniMax’s content moderation, and the fully automated pipeline just wouldn’t run. Its greatest value was polishing out the AI flavor; everything else was uncontrollable.

Then I switched to webnovel-writer, a long-form web fiction assistant system built on Claude Code, supporting serialization at the 2-million-character scale. It beat Inkos on context management — very high cache hit rate, more meticulous planning — but the text it produced reeked of AI, so much so that one chapter even tested as 100% AI. I revised a few chapters and found it still didn’t work.

Only after neither off-the-shelf tool satisfied me did I start building my own wheel.

Using OpenCode + Claude Code’s skill mechanism, I built 11 skills, 9 of which directly serve my own writing pipeline:

• Writing: sanguo-writer — reads the settings and chapter outline, then automatically writes the next chapter
• Review: sanguo-auditor (audits chapters) / sanguo-dagang-auditor (audits outlines)
• Polishing and revision: sanguo-pollish (style polish) / sanguo-revise (revision discussion) / sanguo-revise-yasuo (compression and trimming)
• Reader simulation: sanguo-reader-test (simulated blind reader test) / sanguo-shuangwen (reviewing from a shuangwen reader’s perspective)
• Memory management: sanguo-memory-dump — generates long-term memory, tracking settings, characters, and earlier details
• Methodology reference: cuonale-perspective — narrative techniques distilled from a hit urban-rebirth novel
• General advisor: accuracy-top-expert — an expert mode for when fact-checking and deep reasoning are needed

The workflow eventually settled into: GPT for overall planning + DeepSeek as the main workhorse for the text + Gemini for chapters that need literary flair + GLM 5.2 for review and revision. Yep, different models are good at completely different things; no single model can do it all.

On scale, the numbers tell the story better:

• In May I finished the first 5 volumes, about 95 chapters; by the end of June I’d written 137 chapters cumulatively
• DeepSeek burned through 1.8–2 billion tokens in a single month; on heavy days it torched 60 million in one day. From June 1 to 19 alone, I fired off more than 2,700 requests
• Mimo over the same period: nearly 300 million tokens, more than 3,300 requests

Then I started breaking down hit novels — I read a god-tier urban-rebirth work and was reduced to tears by the lyrics at the ending; I read a campus novel and took apart its “male-female double act” formula, analyzing its “feedback-loop structure”; I also read a reverse-time-travel-back-a-thousand-years novel, a 90s rebirth struggle story, and several others, extracting chapter outlines and dismantling character personalities one book at a time.

The more I read, the more clearly I saw that line — between what AI can do and the polished work humans can do, the difference is fundamental.

The deeper my own reading went, and the more I compared it against AI’s analysis, the more I felt it: even when AI analyzes a work, there’s some loss of precision.

Or rather, the subtle things a human can catch, AI simply cannot catch precisely.

3. From First Principles: Why “Creativity” Is a Mountain AI Cannot Climb

A while back Kimi K3 was released — a domestic large model with 2.8T parameters, whose reasoning ability surpasses GPT-5. A lot of friends around me panicked: “The parameter count is already this big — if we keep piling on more, will creativity just emerge?”

Putting it together with my actual experience and digging through the material, the conclusion is: piling on parameters doesn’t solve the problem.

Every current large model is, at bottom, an autoregressive probability predictor — put plainly, it’s a super-complicated “word-chaining machine.” Every token it outputs is the highest-probability next word selected from the statistical patterns learned from massive training data.

That mechanism sets two ceilings:

First, it has no “intent.”

When a human writes a novel, you’re not calculating which character has the highest probability of coming next. You have images in your head, emotions, things you want the reader to feel. You’re designing a “narrative trap” — you know exactly what expectation this passage should plant in the reader’s mind, and then, at some moment 30 chapters later, you detonate it with precision.

AI can’t do that. It’s just guessing the next character. However well it guesses, it’s still guessing — not designing.

Second, it has no body, no lived experience.

I finally understand why AI simply cannot write something like a “comedic punchline.” Because comedy depends on immediacy, global narrative control, and pragmatic judgment grounded in real experience.

For example: a few days ago Lukia was using voice input. I said “36 grotto-heavens,” she said “72 blessed lands” — and the input method transcribed it as “72 伏地.” Then she sent one more character: “魔.”

That kind of accidental, everyday stroke of genius — AI can never, ever write it. Because it has never blushed in that kind of hilarious, awkward moment, never shared a muddled giggle with a partner it lives with day in and day out. It can’t understand what that feels like, so it can’t guess a character like that.

Excellence in craft cannot make up for a missing creative structure.

From small models like Deepseek’s 200-something B
to large models at the trillion scale,
they’re all more than enough for writing fluent sentences.
But for the truly brilliant works humans need, precise wording is only the passing grade.
Subtle emotional resonance and wildly inventive reversals are even more indispensable.

4. What AI Can and Cannot Do

After three months of running AI on novel writing, plus my day-to-day use of AI for backend code and H5 frontend work, here’s how it feels to me now:

What AI can do:

• Write H5 frontend code — 100% handed to AI, no review needed from me (because the frontend can self-check and self-revise via e2e tests until it runs)
• Summarizing material, extraction, rapid analysis
• Style imitation and reproduction — mimicking a writer’s feel for language, rhythm, word-choice habits. Fable 5 is very good at this
• Stereotyped writing in standard formats, exam essays
• Filling in transition paragraphs, mass-producing supporting content

What AI can’t do:

• Comedic punchlines, consecutive reversals (this is the blind spot of probability prediction)
• Global narrative control — “tying the knot” and “untying the knot” in the reader’s emotional expectations
• Emotional resonance grounded in real lived experience
• Writing the accidents and coincidences of life
• Genuine creativity and making the final call

AI is like an intern doing work for you. Sometimes it’s impressively nimble, but most of the time it’s lazy and dumb, with a very straight-line brain. Once you hand it the work, you have to be the director yourself — establishing characters, laying out foreshadowing, controlling the reversal beats; these core things still have to be done by a human.

5. Where Human Value Really Lies

I was chatting with an old online friend a while back. He was a developer from the Flash era, and now he’s running his own startup. I said: “With AI around, code production capacity is in oversupply. I barely need to write code myself anymore. The real value of humans is doing creative work and making the final call.”

Honestly, I stunned myself a little after saying that. I’m a programmer who’s written code for over a decade, and now I’ve realized that code doesn’t need me anymore. But on the other hand, that excites me — because writing code was never the point. Making something is the point.

So in the AI era ahead of us, what abilities should humans really focus on?

1. Creativity and the ability to make the call. AI can give you ten versions of a plan, but which one to pick, what tone to set for the whole thing — that’s judgment unique to humans. AI has no “sense of taste”; it can’t judge what’s good.

2. Accumulating lived experience and emotional raw material. Humans actually live; the various feelings and emotions that arise in life are something AI can never get. “The memes you play with online every day, the bullshit sessions with friends, the embarrassment when a flex fails, the flutter in your chest when you see someone handsome or pretty, every little everyday joy you feel in daily life — that’s the nutrition a novel needs.”

3. Narrative magic. In writing, what’s truly brilliant isn’t “precise word choice” — it’s being able to anticipate the reader’s emotions, plant preset anchors in their mind, and then detonate resonance at the moment of reversal. That kind of thing exists only in human intuition and experience.

4. The ability to orchestrate humans + AI. The winners of the future aren’t “people replaced by AI” or “people who reject AI” — they’re “people who can use AI to turn creativity into production capacity.” Humans set the direction, the characters, the turning points; AI does the breakdown, the filling in, the mass production — humans control the core, AI handles the industrialized output.

6. An Unsettling Footnote

While writing this article, I couldn’t help wanting to say a couple more things.

AI boosts productivity, but a productivity boost doesn’t bring a consumption boost. Which makes the supply-demand relationship increasingly tense — overcapacity, everyone grinding away, yet consumers’ wallets haven’t gotten any bigger because of it.

AI is hitting entry-level white-collar jobs hard. Without entry-level positions, mid- and senior-level talent never gets the chance to develop. This is a generation’s predicament.

AI compute is controlled by the giants, so even if you’re an independent creator who knows how to use AI, in essence you’re just a digital tenant farmer — working for the platform. Your creativity, your time, your output — all of it ultimately has to pass through the platform’s algorithmic distribution before it can create value.

But I won’t go further down that road. Back to the point — no matter how the economy grinds, no matter how society changes, the part of us that makes us human is something no one can take away. You get angry, you get scared, you get moved, you cry your eyes out over the ending of some novel. None of that has anything to do with capital markets or parameter counts. It’s the proof that you lived.

In Closing

When I started writing the novel in May, I never imagined that three months later things would turn out like this.

Those 42 days, I wrote every day, and every day was “💪 passion.” The numbers were dismal, but I was never depressed about it — because every day I was writing, creating, pushing forward.

Later I read a lot of peak-tier works on Qidian and Fanqie, and saw what truly god-tier writing looks like — and I was suddenly depressed for several days. Only then did I realize that what was making me unhappy wasn’t “bad numbers” — it was that I’d started comparing.

Same with AI. Compare AI’s output against human creativity and you’ll find they’re not even on the same footing — it’s not a gap in magnitude, it’s a gap in dimension.

AI is a superbly useful hammer. But a hammer can’t draw up blueprints, can’t paint inspiration, can’t hear the rhythm of music. It’s an extension of your arm strength, not your soul.

Instead of being anxious about being replaced by AI, think about this: in all the years you’ve been alive, what are the things that are truly one-of-a-kind, the things AI can never produce?

And that is: an exclusive ability to produce disruptive original work, a personal style in your words, and the deep soul-level empathy that comes from real lived experience.