Can you "humanize" AI prose?

A whole product category now exists to launder machine text past detectors: paste in generated prose, get back "humanized" prose. They work, sort of, against the tools they're built to beat. They do much less against a structural fiction audit, and the reason why is worth understanding whether you build with these tools or screen for them.

What a humanizer actually does

Almost all of them perform the same three operations: synonym substitution (swap common words for less common ones), syntactic shuffling (split and recombine sentences, reorder clauses), and controlled noise injection (occasional informal phrasing, the odd deliberate imperfection). The target is perplexity. Rare-word substitution raises lexical surprise; sentence reshuffling raises burstiness. Against a detector whose entire theory is "low perplexity equals machine," this is effective by construction. You are not making the text human; you are editing the exact number the detector reads.

Why it doesn't move a fiction audit much

A structural audit for fiction isn't reading perplexity. It is reading properties that survive paraphrase:

• Scene vs. summary. A humanizer changes the words in "a wave of grief washed over her." It does not turn that sentence into a dramatized scene with a body in a room. The emotional-summary tell is structural; synonym swaps don't touch it.
• The triad reflex. Reordering a sentence rarely removes the underlying habit of describing things in groups of three. The container survives the repaint.
• Dialogue subtext. No substitution engine invents the gap between what a character says and what they want. That gap either was authored or it wasn't.
• Voice consistency. Laundering operates sentence-by-sentence. It cannot manufacture a single recognizable idiolect sustained across eighty thousand words.

In practice we often see humanized text score worse on some audits, not better: the injected noise makes rhythm erratic in a way that doesn't match human burstiness either — it has the statistical signature of randomization, which is its own tell.

The thing that does change the signal

There is exactly one reliable way to move a structural audit toward "human," and it isn't a tool. It's revision — the real kind. A writer who takes generated raw material and genuinely rewrites it: cutting the summary and dramatizing the scene, giving a character a want that contradicts their words, imposing a consistent voice, breaking the cadence on purpose where the meaning demands it. After that level of work the audit moves, because the manuscript has actually acquired the properties the audit measures. At which point a fair question is whether the output is still "AI prose" in any meaningful sense, or simply a book a human wrote with an unusual first draft.

That ambiguity is real and we don't pretend otherwise. The line that matters to most editors is not "did a model ever touch this" but "did a human mind do the authorial work." Surface laundering fails that test by definition — it is specifically designed to avoid doing the work. Deep revision passes it because it is the work.

If you screen manuscripts

Two practical implications. First, a low score is not proof of innocence any more than a high score is proof of guilt; a heavily laundered text can land mid-range. Always read the per-audit pattern, not just the headline. Second, the audits least affected by laundering — scene construction, dialogue subtext, voice consistency — deserve the most weight in your judgment, precisely because they are the hardest to fake without doing the authorial work that would make the question moot.

If you write with these tools

The honest summary: a humanizer can fool a perplexity detector and will not give you a good book. The properties it cannot fake are the same properties that make fiction worth reading. The shortcut and the craft are pointed in opposite directions, which is, in the end, the most reassuring thing we've learned building this.