Character.AI is enormous. Reportedly twenty million monthly active users. Fast, polished, free, millions of community-created characters. If you want to casually chat with an AI version of a fictional character, it's the best product in the market and it's not particularly close.

But if you've ever tried to write fiction on it — real fiction, the kind where characters can be angry, afraid, violent, sexual, morally compromised — you know the problem.

The filter

Character.AI's content filter is the strictest in the industry. That's not editorial opinion — ask anyone who's used it for more than a few sessions. The filter doesn't just block explicit content. It blocks fight scenes, intense emotions, romantic tension, moral ambiguity, basically anything that makes fiction interesting.

You're writing a scene where two characters argue. The argument gets heated. One character says something cruel — the kind of thing real people say when they're hurt and angry. The AI stops. Refuses. Suggests a "more constructive" direction.

Or you're writing a mystery. A body is discovered. You try to describe the scene and the AI won't engage with anything that involves physical harm, even to a fictional character who's already dead before the story started.

Or — and this is the one that drives people to Reddit — you're in the middle of a romance. Things are going well. The tension has been building for hours. And then the AI fades to black and tells you it "cannot continue this interaction in this direction."

The model

Character.AI uses proprietary models optimized for speed. After their original founders left for Google in 2024, the company shifted toward fine-tuned open-source models alongside their in-house "PipSqueak" model. They're fast — impressively fast — and good at maintaining a consistent character voice across a long conversation. For chat, this is ideal.

For fiction, the output reads like a chat log. Short sentences. Action asterisks. Minimal description. No interiority. The model was built for conversational back-and-forth, not narrative prose, and it shows.

Underfiction uses frontier models — Opus, GPT, Gemini, Grok — through Venice. These are the most capable language models available, prompted for literary writing. And because Venice handles inference differently than the official APIs, the models follow creative direction without the refusals and hedging you'd hit even on the direct API. The output reads differently. Scene-setting, tension, subtext, interior monologue. Prose, not chat.

Where your stories live

Character.AI stores everything server-side. Your conversations, your characters, your most private creative experiments — all in their database, accessible to their employees, usable for training per their ToS, and subject to law enforcement requests.

Underfiction keeps stories local by default and offers optional Story Sync for cross-device writing, with synced copies encrypted at rest. Inference goes through Venice, which separates your account identity from prompts before model requests are handled.

For casual chat this probably doesn't matter. For the kind of fiction that Character.AI's filter blocks — the kind you'd want privacy for — it matters a lot.

The real question

Character.AI is free and has millions of community characters. That's a real advantage. If you want social character discovery and casual chat, it's the right tool.

If you're here reading this post, though, you're probably not looking for casual chat. You're probably looking for the thing Character.AI keeps refusing to let you do. That's what we built Underfiction for.