Fanfiction humanizer, scene by scene

A single constrained rewriting pass, executed on our AI server. The accompanying instruction directs the model to rewrite the supplied text while preserving the original meaning and factual claims, and to return only the rewritten text with no preamble, commentary, quotation marks or markdown fence. Applied to fiction, the preserved claims are the scene's propositional content: the events, the stated details, the substance of the dialogue.

Remote terms apply. Five runs per rolling 24 hours, input between 1 and 12,000 characters, a 300 character instruction field that governs the whole run. No generation capability exists on this route: nothing is invented, extended or suggested.

The instruction field is the only configurable surface, so it carries the entire specification.

Step 1

Submit one scene, cut at a structural boundary and sized to the output ceiling.

Step 2

Supply a standing instruction fixing person, tense, register and length.

Step 3

Execute, then diff the result against the source for continuity before accepting it.

The input ceiling is 12,000 characters, roughly two thousand words. The completion ceiling is 1,400 tokens, roughly one thousand. The asymmetry is not a bug to route around: it fixes the working unit at approximately a thousand words, which in practice is a scene.

Choosing the cut point matters more than hitting the number. A break chosen at a scene transition produces a rewrite whose seams you control. A break chosen by the token limit produces a rewrite that stops mid-paragraph and a run spent on an unusable result.

In non-fiction the preserved claims are assertions about the world. In fiction they are assertions about the text's own world, which is a stricter test in one respect: a reader who has followed forty thousand words notices a changed detail immediately, whereas a paraphrased statistic often passes.

Preservation is instructed rather than guaranteed, and generation drifts. The failure modes worth diffing for are specific: a dropped negation, a reassigned line of dialogue, a continuity detail such as which hand holds the object, and a compressed passage of time. Each reads perfectly well and each changes the scene.

It reaches the surface layer, and the surface layer is where most of the machine-written impression lives: uniform sentence length, paragraphs of matched size, hedged description, dialogue in which every speaker has the same syntax. Varying those is exactly what the rewrite is good at.

It does not reach anything above the sentence. Writers testing these tools on real fics describe AI prose as correct but bland, with a voice that drifts over a long work. The blandness that survives a rewrite is structural: a scene with no reversal, a character with no want, an exchange where nobody is trying to get anything. No rewriting pass supplies those, and a page claiming otherwise is selling the wrong thing.

The 2026 fanfiction accusations are the clearest available case study. The evidence was one CSS class name that labels generated paragraphs and survives a paste into an archive's rich text editor, checked by a style override that turned the page red when the tag appeared.

Examine what that can support. It detects one vendor's paste path, not authorship: the document itself conceded that a missing tag does not mean a work is AI free, and that routing the text through a word processor first usually leaves nothing detectable. It is blind to proportion: one work carried six instances across 263,627 words and another 1,300 instances in under a third of that length, and both were framed identically. It is blind to editing entirely. Reporting on the episode summarised it as artifact inspection rather than authorship classification, and warned that a narrow artifact treated as proof causes harm. The document opened on undeniable proof and had narrowed that to specific sentences and paragraphs within four pages.

This follows directly from the previous section and is the most practical thing on this page. A rewriting pass reads and returns text. It does not inspect or remove HTML, class attributes, markdown characters or invisible codepoints, and no claim to the contrary appears here.

Those jobs belong to deterministic routes that run in your browser with no allowance attached: an HTML stripper that removes tags and decodes entities, a markdown stripper, and a hidden character scanner covering exactly 14 invisible codepoints. Sequence them after the rewrite and immediately before posting, since a rewrite can reintroduce typographic characters such as long dashes.

It does not claim that a rewrite defeats a detector, a fandom checker or any future one. Detection reports signals, never proof of authorship, and predicting a third party's verdict is outside what any tool can honestly offer.

It does not state any archive's rules, because those change and belong to the archive. And it does not treat rewriting as a substitute for disclosure: where a community or a platform expects AI use to be tagged, that expectation attaches to how the work was made, not to how the sentences ended up reading.