Fanfiction Humanizer

Since 2023 AO3 has had recurring waves of readers accusing writers of using AI, including on works posted before the tools they were accused of using existed. In October 2023 writers were locking their accounts over it. The pattern is now familiar: a work reads too smoothly, a comment says so, and the writer spends a week defending a draft they wrote by hand.

In June 2026 an anonymous group circulated a twenty-five page document naming around thirty writers, and harassment followed. It is worth looking at what its evidence was, because it is the only technically checkable claim anyone in these waves has made.

The document rested on a single artifact: font-claude-response-body, a CSS class Claude attaches to its output, which survives a copy-paste into some editors including AO3's. That is real. It is also, on its own authors' account, close to useless as proof. Pasting through Word or Google Docs strips it, so its absence shows nothing. Only Claude leaves it, so other models leave no trace at all. And it marks paragraphs rather than works — six instances in a 263,627-word fic were treated the same as thirteen hundred in a short one. The document opened by claiming undeniable proof of AI content and had narrowed to proof of Claude usage for specific sentences by page four.

We say this as the people who build the scan that reads exactly this kind of residue: paste residue evidences a paste, not authorship. It can tell you text passed through an application. It cannot tell you who wrote the story, how much of it, or whether the writer typed every word around it. Anyone presenting it as more than that is overreaching, whatever their intentions.

This is worth stating flatly, because a great deal of the anxiety on this search is about a rule that does not exist. AO3's Terms of Service permit fanworks made fully or partly with AI tools, provided they qualify as fanworks. In November 2024 the Organization for Transformative Works confirmed it would neither implement a ban nor require a mandatory disclosure tag.

What it does instead is encourage voluntary tagging — Created Using Generative AI, AI-Generated Text — so readers who want to filter that content can. The reasoning follows from the archive's purpose: it exists to preserve fanworks broadly rather than to judge which ones deserve preserving.

Two things that can get you actioned there, and are worth not confusing with an AI rule. Anti-spam policy still applies, so posting a large number of works in a short window is a problem whoever or whatever wrote them. And the OTW's own position on AI is not permissive in every direction: it has taken measures against large-scale scraping of the archive and sought exclusion from datasets used to train these models. Permitting fans to use the tools and objecting to fic being harvested to build them are not in tension.

Other archives set their own rules. Read the one you post to rather than assuming AO3's position generalises.

The instinct is to rewrite the chapter until it stops reading as generated. It is the wrong move, and not only because it concedes the point. Clean, even, competent prose is what reads as AI to both detectors and readers — so the writers most often accused are polished ones, people writing in a second language, and anyone whose work has been through a beta. Rewriting toward suspicion makes your prose worse and settles nothing.

What holds up is provenance. Drafts with a history. Notes. A document with version history showing the thing being built. A beta who read chapter four in March. None of that is glamorous and all of it is more persuasive than any score, because it is evidence of the process rather than an argument about the product.

The one technical thing worth doing is clearing invisible characters before you post — not to hide anything, but because they are the only checkable artifact in this whole argument and they arrive by accident. Text that has been through a chat window, a phone keyboard and a document editor collects them at every step, and they mean nothing about who wrote it.

It reads prose. It has not read your canon, does not know your characters, and cannot tell you whether a line sounds like them — which is the only question that really matters in fanfiction. Everything fandom-specific stays with you.

It will not make your writing good. It finds the constructions that make fiction read as machine-made and clears the residue that came in with a paste. What makes a fic worth reading is having something particular to say about these people, and no tool has ever supplied that.

And it does not promise a detector outcome, here or anywhere else on this site. On AO3 there is no detector to beat. Where readers are the detector, no rewrite reliably changes their mind — which is the honest answer, and the reason the rest of this page is about evidence rather than evasion.

The one pass worth running last is the character-level one. The zero-width space remover finds what travelled in with your text — invisible on screen, and the only artifact in the whole accusation argument that anybody can actually check.

For the general approach to prose rhythm and structure, the AI humanizer covers it outside a fiction context, and the AI detector explains what these tools measure and why they disagree with each other so much — worth reading before you take any score seriously, including one someone else ran on your work.

If your draft came out of a chat window, the Claude watermark cleaner deals with what Claude specifically leaves behind, which is the residue the June 2026 document was built on.