AI humanizer mistakes: the 12 error classes, with examples

Six classes feed the meaning score, which is rated 1 to 5 by two editors reading blind. The table shows each class with a pair written to isolate it. In a real output several classes usually appear together.

Table: illustrations written by the editors to show each class. They are not tool output.

The last row is the one that gets writers into trouble with sources. The input credited a claim to a named party and kept the party's own hedge; the output states the claim as fact in the writer's voice. A reader checking the writer's work would now hold the writer to a claim the writer never made.

The changed-qualifier row is the commonest in the editors' experience of reading rewritten text, and the hardest to see, because the sentence after the fault is shorter and more confident than the one before it. Confidence reads as quality. It is not.

Six classes feed the errors-per-100-words figure, which carries 20 of the 100 points. These are visible without the input, so a reader can catch them alone.

Table: illustrations written by the editors to show each class. They are not tool output.

The synonym misfire and the replaced technical term are the two that a word-swapping tool produces by design. A tool told to vary vocabulary looks for a less common word with a similar sense; "routine" and "habitual" overlap, "significant" and "important" overlap, and in both cases the overlap is not where the sentence needed it. The guide on how to humanize AI text manually puts word swaps last in its edit order for this reason.

Because the edits that move a detector reading are the edits that carry meaning risk. Krishna and colleagues built an 11-billion-parameter paraphraser in 2023 that could "control lexical diversity and content reordering", and it dropped one detector's detection rate from 70.3 percent to 4.6 percent at a 1 percent false positive rate (Krishna et al. 2023, arXiv 2303.13408, read 9 November 2026). Lexical diversity means rarer words; content reordering means moved clauses. The first produces synonym misfires and replaced terms; the second produces reversed relations and broken references.

The open-source blader/humanizer skill, which edits for readers rather than detectors, writes the opposite rule into its own instructions: a name, number, date, quote or citation "must come from the source or the writer", and the skill should ask for a missing detail rather than invent one (blader/humanizer README, read 9 November 2026). The same README states that detectors still flag most of its output. That is the trade in one line: the tool that refuses to invent moves readings less.

Claims first, grammar second. Write the input's claims as a list of short lines before you open the output. Tick each line off in the output. Any output sentence you cannot tick against a line is an invented fact or a changed one; any line you cannot tick is a dropped claim. Then read the output on its own for the six writing faults, and divide the count by the output's length in hundreds of words so that a longer text is not punished for its length.

Read any sentence that sounds better than the input twice. Smoother usually means a qualifier went missing. Read any number, name and date against the input character by character, since the changed-number fault is a one-digit error that no amount of reading for sense will catch. And treat an unattributed strong claim with suspicion; if the input said who said it, the output should too.

No detector score should be the only evidence in a misconduct case, and no clean detector reading is evidence that a text still means what it did. Once the first run is published, the main ranking will show each tool's meaning score and error count beside its detector readings, with every marked sentence in the downloadable ratings file. Until then, the classes on this page are the checklist. The guides hub holds the rest of this series.