Signs of AI writing: 40 tells, each with a human alternative

Because of how the text was chosen. A model predicts the next likely word; the result, as the blader README quotes Wikipedia, "tends toward the most statistically likely result that applies to the widest variety of cases". A sentence that signals importance instead of adding a fact, a rhythm applied by rule, an ordinary fact dressed as a turning point: each is the widest-fit choice.

Liang and colleagues measured the statistical side in 2023. Seven detectors flagged 61.22 percent of 91 human-written TOEFL essays as AI, and the flagged essays had significantly lower perplexity, meaning more predictable word choice; when a chatbot was asked to enrich the vocabulary, the false positive rate fell to 11.77 percent (Liang et al. 2023, arXiv 2304.02819, read 9 November 2026). Predictability is what the detectors read and what the tells below describe in words.

Numbers 1 to 5 are the five the blader/humanizer README calls strongest. The rest are numbered by family. Every quoted phrase is a tell, not a recommendation.

Loosely, and in one direction. GPTZero describes its product as "a sentence-by-sentence classification model" with premium features "to detect AI vocabulary", and says its "Paraphraser Shield" is built to catch text altered to look more human (GPTZero, technology page, read 9 November 2026). A classifier reads statistics across every sentence; this list reads habits a person can see. Where they overlap is vocabulary and rhythm, tells 11 to 18 and 38 to 40, because those are the tells that lower perplexity.

Two consequences. Removing the vocabulary and rhythm tells may change a detector reading, and nobody can promise by how much; the blader README states plainly that "getting past AI detectors is not a goal, and detectors still flag most of its output" (read 9 November 2026). And removing the staging and inflation tells changes nothing a detector reads while making the text much better, which is the stronger reason to remove them. The guide to whether AI detectors work has the measured accuracy figures.

As an editing checklist, in family order. Read once for staging and inflation, since those cuts shorten the text and expose the claims. Read again for rhythm, breaking the triads and the same-shaped paragraphs. Read a third time for vocabulary, replacing each weighted word with the plain one. The manual humanizing checklist puts the same work into 20 edits with the meaning risk of each.

Do not use the list as proof of authorship. A writer who learned English from textbooks, or who writes in a formal register for work, will show a dozen of these tells on an honest page, and Liang and colleagues' 2023 result is what that looks like when a machine does the judging. No detector score should be the only evidence in a misconduct case, and no checklist should be either. The guides hub holds the rest of this series.