Humanized AI Text Reads Odd More Often Than Robotic: Swapped Words, Inserted Nonsense, Stray Typos

Three artefacts, and none of them reads like a chatbot on its own. First, a word sitting one step off the ordinary choice. Pangram Labs, which audits these tools, prints the pair: the plain sentence is "I need to get my car fixed because the engine is making a strange noise", and its rewrite is "I require to obtain my vehicle repaired because the motor is creating a peculiar sound" ¹.

Meaning survives that. Register doesn't.

The second artefact means nothing at all. Humanizers sometimes add nonsensical phrases to a piece of writing in the hope that a detector will read the gibberish as unlikely to have come from a machine ¹. The example Pangram prints is a fragment of code and the page range 18 to 23, dropped between two sound sentences ¹.

The third is plain damage. A pass can leave typos and broken grammar behind, a misspelling where the draft had none ¹. Surface errors beside an untouched structure are a reason to open the version history. Nothing more than that.

Picture the three together on one marked page. A paragraph carries two words nobody in the seminar would choose and one clause that resolves to nothing, and a misspelling sits inside a sentence that is otherwise doing careful comparative work. Read a full paragraph before deciding anything. A single sentence carries little of this.

The register is the tell. A mistake on its own is not.

When it does, the rewrite has optimised for a score and spent clarity to get there. Pangram Labs says as much about its own synonym example: swapping words for synonyms damages the text and lowers its clarity and fluency, sometimes to comedic effect ¹. The sentence still parses. What it has stopped doing is sounding like anybody in particular, and that is a different failure from sounding like a machine.

How much of that damage shows varies from pass to pass. A 2023 stress test of recursive paraphrasing reported the method cutting detection rates sharply while degrading text quality only slightly in many cases ³.

Price is the variable Human's own measurement found to matter. Human, tuned for college-level academic writing at human.olive.is, reports one in-house finding here, measured September 2026 on its own held-out sets rather than by an independent audit: whether a humanized document is caught depends on the humanizer's price tier more than on which humanizer it is. This is an estimate from our detector. Treat a flag as a reason to look closer, not as a finding.

Price tiers aren't hypothetical. NBC News reported that Turnitin keeps a list of 150 tools, some charging as much as $50 for a subscription, that adjust text so a detector will not flag it ⁸.

What a pass does to a student's sentences, clause by clause, is what-a-humanizer-does-to-writing. The artefacts here are what survives that pass and reaches the marking pile.

Vocabulary, first and by a wide margin. In an ACL 2025 study, five annotators who frequently use language models for writing tasks read 300 articles and wrote down why they judged each one. Vocabulary clues appeared in 53.1 percent of those explanations, sentence structure in 35.9 percent, originality in 23.7 percent and formatting in 15.0 percent ².

The structural tells are concrete enough to teach. The researchers' definition of the structural clue names a high frequency of "not only ... but also" and consistently listing three items; one annotator put it as the "it's not just this, it's this" comparison, alongside listings of specifically three ideas ².

One explanation described writing filled with the same language it used to describe everything: inspirational, stunning, essential ². The researchers' own definition of the originality clue reads like a line off a marking rubric: machine writing is straightforward, safe, and lacking in surprises or humor ².

Their judgement held up under a humanizer.

The researchers built one themselves by prompting o1-Pro with instructions drawn from the annotators' earlier comments, then put a batch of 60 articles, 30 of them humanized, in front of the same panel. The expert majority vote was perfect on all 60 ².

Across the study as a whole, that same majority vote missed 1 of 300 articles ².

Whether the skill survives a real marking pile, with a name attached to every paper, is can-teachers-tell-humanizer-used.

Because detectors misread plain prose in the same direction. A 2023 Stanford study simplified the word choices in essays written by US eighth graders and watched the average misclassification rate across the detectors it tested climb from 5.19 percent to 56.65 percent ⁴. Nothing about the authorship changed. The vocabulary did.

The reverse held too. Rewriting human TOEFL essays with richer vocabulary cut the average false positive rate from 61.22 percent to 11.77 percent ⁴. The authors read that as a sign the detectors may be penalising a limited range of expression rather than reading authorship ⁴.

Bloomberg Businessweek put names to who pays for that. Its reporting, drawn from students, academics and AI developers, says the students most exposed to a wrong accusation are likely those who write in a more generic manner, including neurodivergent students and students working in a second language ⁷. A humanizer vendor makes the same point from the other side. Undetectable AI's ethics page states that a single AI text detection score alone is not enough to rule out cheating or deception ⁶.

The incentive runs backwards, too. A 2026 Notre Dame team took 642 published abstracts of 25 to 500 words, had a model rewrite each one, and ran the rewrites through a humanizer. Light, guideline-compliant edits to the same abstracts were flagged 38 to 80 percent of the time, depending on the detector and the field, while fewer than 4 percent of the humanized rewrites were ⁵. That is the finding that should worry a department: guideline-compliant AI editing carried a higher sanction risk than humanizer-assisted evasion ⁵. The authors say classroom writing may behave differently ⁵.

How that lands on a cohort writing in a second language is ai-detectors-non-native-english-writers-policy.

Open the version history before opening a detector. A paragraph that reads wrong is a question about how it was written, and the draft record can show whether the passage grew across several saves or arrived whole in one. Odd wording earns that look. It earns nothing further.

Read the whole document rather than the sentence that snagged. The Notre Dame cohort ran as short as 25 words ⁵, and Human itself reports a weak reading under 150 words and none under 50.

Compare the paper against something the student already handed in. A graded essay from week three costs nothing and outlives every version number a detector ships.

The order those checks belong in is tell-if-student-used-chatgpt.

If a reading would help after that, take one. Human reads a document sentence by sentence and reports how much of it reads as machine-written or machine-edited, with a document verdict of Human, Mixed or AI. This is an estimate from our detector. Treat a flag as a reason to look closer, not as a finding. Paste as much of the paper as fits, up to 1,000 words. A lone short paragraph gives the detector little to read, and under 50 words Human declines rather than guessing.