← Blog

What AI Fiction Should Read Like

I'm not going to explain this one. I'm going to show you.

Same scene. Same characters. Same moment: a woman sees someone she hasn't seen in three years. First, the way most AI platforms write it. Then, the way it should read.

Typical AI roleplay output
*She looked at you with wide eyes* "I... I didn't expect to see you here." *She tucked a strand of hair behind her ear nervously* "How long has it been? Three years?" *She shifted her weight from one foot to the other, clearly uncomfortable* "You look... good. Different, but good."
Same scene, prompted for prose
She was standing by the window when he came in, and for a long moment she didn't turn around. When she did, her expression was the kind of careful blankness that takes effort to maintain. She tucked a strand of hair behind her ear — a gesture he remembered from a different kitchen, a different life — and said, "Three years." Not a question. She already knew exactly how long it had been.

Same information. Same model, even. But the second version has physical space. Interior observation. A gesture that carries history. Dialogue that reveals instead of declares.

One more.

Typical: a character enters a room
*He walked into the old house and looked around* The place was a mess. Dust everywhere, furniture covered with sheets. *He felt a wave of nostalgia wash over him* "I can't believe it's still standing," *he muttered to himself* *He ran his hand along the dusty mantlepiece*
Prompted for prose
The door stuck on the swollen frame, same as it always had. He put his shoulder into it the way his father used to, and the smell hit him before the light did — dust and old wood and something underneath that might have been lavender, years gone. The furniture was draped in sheets that had yellowed to the color of old teeth. He stood in the hallway with his bag still on his shoulder and listened to the house settle around him, the tick of pipes cooling, the particular silence of a place that has been empty long enough to develop its own gravity.

The asterisk version tells you what happens. The prose version makes you feel like you're standing in the hallway.

Typical: a tense conversation
*She crossed her arms and glared at him* "You promised you wouldn't do this again." *Her voice was cold* "I trusted you. We all trusted you." *She turned away, clearly hurt* "Just... go. I can't even look at you right now."
Prompted for prose
"You promised," she said, and the flatness of it was worse than shouting. She hadn't moved from the chair. Hadn't uncrossed her legs. The anger was all in what she wasn't doing — not raising her voice, not standing up, not giving him the satisfaction of a scene he could push back against. "We all trusted you," she added, and the we landed like a door closing. She picked up her glass and took a sip and looked past him at the wall, and he understood that the conversation was already over. Everything after this was just him finding out.

Notice what's different: the subtext. In the chatbot version, emotions are labeled. "Her voice was cold." "Clearly hurt." In the prose version, you infer the emotion from what the character does and doesn't do. The anger is in the stillness. The finality is in "the conversation was already over." The reader does the work, and that's what makes it land.

Why the difference exists

It's not the models. Claude, GPT, Gemini — they've all read more fiction than any living person. They can write like this. The problem is that most platforms prompt them for chat: short turns, fast replies, stay in character, keep it moving.

Prompt for chat, get chat. Prompt for prose, get prose. It really is that simple, and that's what we spent most of our development time on — system prompts that tell the model to write fiction, not to be a chatbot wearing a character costume.

The quality ratchet

There's a second-order effect nobody talks about. When the AI writes well, you write better. You invest more in your inputs because the responses reward it. You write more specific direction. You care about your character's voice. The whole thing ratchets upward.

The reverse is also true. Flat AI output makes you stop trying. You type shorter, lazier inputs because the response is going to be disappointing anyway. The spiral goes down.

Prose quality isn't a feature. It's the thing that determines whether the entire experience works.

New accounts start with 500 free credits.

Try Underfiction

Frequently asked questions

Can AI write good fiction?

Yes. Frontier models like Opus have been trained on virtually all published fiction and can produce literary-quality prose. The quality depends on prompting — most platforms optimize for chat speed, not prose quality, which is why the output usually reads flat.

What's the difference between AI roleplay and AI fiction?

Roleplay produces short, dialogue-heavy exchanges with action asterisks. Fiction produces narrative prose — scene-setting, interior monologue, subtext, rhythm. Same models, different prompting.

What AI writes the best fiction?

Opus generally produces the most literary output. GPT and Gemini are close. But the prompting matters as much as the model — a well-prompted mid-tier model beats a poorly prompted frontier one.