Flight sheet · guide

AI humanizer terms explained: a 30-term glossary

By Best AI Humanizer editorsPublished November 9, 2026Updated November 9, 2026

Short answer

This glossary defines 30 terms from AI humanizers and AI detectors in 35 words or fewer each, in one table, from perplexity and burstiness to paraphrase attack and request cap. Each definition names what the term measures or buys and, where a number exists, who published it, so a reader can quote one line alone.

This glossary defines 30 terms from AI humanizers and AI detectors in 35 words or fewer each, in one table. AI humanizer terms explained this way come in three groups: what detectors measure, what humanizers do to text, and what the plans and this site's method mean by their labels. Longer treatment of the detector terms is in how AI detectors work, the product itself is defined in what an AI humanizer is, and the scoring terms come from the published method. The guides hub holds the rest of the series.

Which AI humanizer terms explained here matter most?

Every term below is one a reader meets on a vendor's pricing page, a detector's report or a study abstract. Each definition names the thing, says what it measures or buys, and gives a source where a number is involved. Where a term has a different meaning at two vendors, the row says so.

Term	Definition (35 words or fewer)
AI humanizer	A paraphrasing tool sold on changing how AI-written text reads to detectors and readers. It substitutes words, reorders clauses, varies sentence length and shifts tone. The promise, not the mechanism, separates it from a paraphraser.
AI detector	A program that returns a probability or label for whether a text was machine-written, from its statistics or a trained classifier. Wu and colleagues' survey sorts detectors into watermark, statistical, neural and human-assisted methods.
Perplexity	How surprised a scoring language model is by each word. Low perplexity means the text is what a model would have written. Liang and colleagues showed in 2023 that non-native English essays also score low.
Burstiness	The variation in predictability across sentences. GPTZero describes human writing as mixing long, surprising sentences with short plain ones, where model text stays even, and uses it alongside perplexity, read 9 November 2026.
Trained classifier	A language model fine-tuned on pairs of human and AI text to tell them apart. GPTZero, Turnitin and Pangram describe classifiers of this kind; the threshold between labels is each vendor's own choice.
Zero-shot detector	A detector that needs no training set, scoring text against a language model's own probabilities. DetectGPT, which measures probability curvature under small edits, is the best known example.
Probability curvature	DetectGPT's statistic: whether small rewordings lower a passage's probability under a model. Model-written text tends to sit at a local peak; human text does not, so perturbing it moves the score less.
Watermark	A pattern planted in word choice at generation time. Kirchenbauer and colleagues' 2023 method favours a randomised set of green tokens so a detector holding the key can test any short span.
Green list	In Kirchenbauer and colleagues' watermark, the set of tokens softly promoted before each word is generated. A watermarked text contains more green-list tokens than chance predicts; a detector counts them.
Paraphrase attack	Rewriting AI text, by a tool or a second model, so a detector's statistics change while the meaning stays. Krishna and colleagues' paraphraser cut DetectGPT's detection from 70.3 to 4.6 percent in 2023.
Recursive paraphrasing	Paraphrasing an output again, repeatedly. Sadasivan and colleagues used it in 2023 to lower detection rates across watermark, neural, zero-shot and retrieval detectors on 300-token passages, with small quality loss in many cases.
Spoofing attack	Making human text read as AI to a detector. Sadasivan and colleagues showed an attacker can infer a watermark's hidden signature without access to the detector and plant it in human writing.
Retrieval detector	A detector that stores every text a model generates and checks new text against the store. Krishna and colleagues proposed it in 2023 as a defence that paraphrasing does not break.
False positive	Human-written text labelled AI. Weber-Wulff and colleagues found false positive rates from 0 to 50 percent across 14 tools in 2023; GPTZero states a rate of no more than 1 percent, read 9 November 2026.
False negative	AI-written text labelled human. Weber-Wulff and colleagues found the 14 tools they tested more likely to call AI text human than human text AI, and more so after paraphrasing or machine translation.
Threshold	The cut-off a vendor sets between labels. Turnitin labels sentences above its threshold AI and does not display document scores under 20 percent, read 9 November 2026; another vendor cuts elsewhere.
Document score	The headline number a detector reports for a whole text. At Turnitin it is the share of qualifying prose sentences above the threshold, not the share of text a person wrote, read 9 November 2026.
Sentence-level detection	Scoring each sentence separately and rolling the results up. GPTZero reports a probability and confidence per sentence; Turnitin counts qualifying sentences above its threshold, both read 9 November 2026.
Confidence band	A detector's statement of how sure it is. GPTZero labels predictions uncertain, moderately confident or highly confident and states an average error under 1 percent for high-confidence predictions, read 9 November 2026.
AUROC	Area under the receiver operating curve: a detector's accuracy across all thresholds, from 0.5 (chance) to 1. Sadasivan and colleagues tied the best possible AUROC to how far human and AI text distributions sit apart.
Baseline reading	The detector score of an input before any humanizer touches it. This site's method checks every input on every detector first, so the change a tool makes is visible against it.
Mode	A named setting that changes how far or in what direction a humanizer rewrites. Walter Writes AI names Simple, Standard and Enhanced; Phrasly names Balanced, Strong and Max, both checked on 8 November 2026.
Request cap	The most words a tool accepts in one submission, separate from the monthly pool. Walter Writes AI's Starter caps a request at 750 words and StealthGPT's Pro at 1,000, checked 8 November 2026.
Monthly pool	The words a plan allows in a billing month. Undetectable AI's entry plan gives 10,000, Walter Writes AI's Starter 30,000 and StealthGPT's Pro 50,000, all checked on 8 November 2026.
Yearly toggle	A pricing page default that shows a monthly figure while charging a year at once. Undetectable AI shows $5.00 a month and charges $60.00 a year, checked on 8 November 2026.
Vendor detector	A detector built into a humanizer that scores the humanizer's own output. This site never treats a vendor detector's reading as evidence of a pass, because the vendor sets both sides of the test.
Meaning retention	The share of an input's claims that survive the rewrite, with nothing added. This site's method has two editors blind-rate each output from 1 to 5, weighted 35 of 100 points.
Errors added	Grammar faults, wrong words and broken sentences a rewrite introduces, counted per 100 words by two editors. Weighted 20 of 100 points in this site's method; readability change is reported, not scored.
Detector pass rate	The share of outputs a named detector labels human, by its own threshold on the run date. This site pools five detectors, GPTZero, Originality.ai, Copyleaks, Pangram and ZeroGPT, weighted 30 of 100 points.
Run	One pass of every tool over the same inputs in the same week, checked on the same detectors and scored with the same rubric. A run has an ID, a date and downloadable raw files.

What is perplexity in AI detection, and why does it misfire?

Perplexity is the oldest statistic in the field and the one most often misread. A language model assigns a probability to each next word; text the model finds predictable has low perplexity, and early detectors treated low perplexity as a sign of machine authorship. Liang and colleagues showed what that measures in practice: the TOEFL essays that all seven detectors in their study flagged had significantly lower perplexity than the rest, and the detectors flagged 61.22 percent of non-native essays as AI while classifying US eighth-grade essays accurately (Liang et al. 2023, Patterns, read 9 November 2026). Limited vocabulary is also low perplexity.

Burstiness was added to patch that. GPTZero describes human writing as varying in perplexity from sentence to sentence, long surprising sentences beside short plain ones, where model text stays even, and states that its current model is an end-to-end classifier rather than the two statistics alone (GPTZero technology page, read 9 November 2026). Both statistics describe the surface of a text, not who wrote it.

Statistic	Low value suggests	Also low for
Perplexity	Model-written text	Non-native prose, plain instructions, lists, formulaic genres
Burstiness	Model-written text	Legal and technical writing with even sentence length
Probability curvature	Model-written text at a local probability peak	Text paraphrased by a second model; a different generating model

What does paraphrase attack mean for a humanizer buyer?

It means the product category has a research literature, and the literature agrees on one point. Krishna and colleagues built an 11-billion-parameter paraphraser and used it to drop DetectGPT's detection rate from 70.3 percent to 4.6 percent at a 1 percent false positive rate, while proposing retrieval over a store of generated text as a defence (Krishna et al. 2023, arXiv, read 9 November 2026). Sadasivan and colleagues then showed that recursive paraphrasing lowers detection rates across watermark, neural, zero-shot and retrieval detectors on passages of about 300 tokens, with only slight quality loss in many cases (Sadasivan et al. 2023, arXiv, read 9 November 2026).

A commercial humanizer is a paraphrase attack with a pricing page. What the studies measured was the detector's reading, not the reader's experience, and that gap is why this site's method weights meaning retention above detector pass rate. A rewrite that turns "no effect" into "a small effect" has lowered a reading and failed the reader. Detector makers answer the studies in kind: GPTZero describes a Paraphraser Shield against paraphrasing and homoglyph attacks, read 9 November 2026, and Turnitin states that its model is trained to flag AI-paraphrased text and names bypassers in its FAQ, read 9 November 2026 (Turnitin, AI writing detection FAQ). Each side updates, and a reading is true for its date.

How do the plan terms and the method terms connect?

The plan terms decide whether an input fits a tool at all; the method terms decide what happens to it once it does. A request cap below an input's length splits the input, and the halves are rewritten without sight of each other; a monthly pool decides how many inputs a month the plan holds; a yearly toggle decides what the headline price means. The method's 15 points for limits and access score the first two from the vendor's own pages on the run date, and price is reported, never scored.

Term group	What it decides	Where it is scored
Request cap, monthly pool, free allowance, sign-up	Whether a 350 to 500 word input fits, and how many a month	Limits and access, 15 points
Mode	Which rows a tool gets in a run: default, plus the strongest documented mode	Recorded per row, not scored
Meaning retention, errors added	What the rewrite did to the reader's text	35 and 20 points, two blind editors
Detector pass rate, threshold, baseline reading	What five named detectors said, against the input's own baseline	30 points, per detector and pooled
Vendor detector	Nothing; never used as evidence of a pass	Noted only

No detector score should be the only evidence in a misconduct case. Anyone writing for a class should check their institution's AI policy before using any tool in this glossary; a humanizer does not make undisclosed AI use acceptable.

Sources

Questions about this guide

What is perplexity in AI detection?

Perplexity is how surprised a scoring language model is by each word; low perplexity means the text is what a model would have written, and Liang and colleagues showed in 2023 that essays by non-native English writers also score low, which is why 61.22 percent of them were flagged.

What is burstiness?

Burstiness is the variation in predictability from sentence to sentence; GPTZero describes human writing as mixing long, surprising sentences with short plain ones, where model text stays even, and uses the measure alongside perplexity, read on 9 November 2026.

What does paraphrase attack mean?

A paraphrase attack rewrites AI text, by a tool or a second model, so that a detector's statistics change while the meaning stays; Krishna and colleagues' paraphraser cut DetectGPT's detection rate from 70.3 to 4.6 percent at a 1 percent false positive rate in 2023.

What is the difference between an AI humanizer and a paraphraser?

A paraphraser rewrites text for any purpose; a humanizer is a paraphraser sold on changing detector readings, and both substitute words, reorder clauses and vary sentence length, so the difference is the promise rather than the mechanism.

What is a false positive in AI detection?

A false positive is human-written text labelled AI; Weber-Wulff and colleagues found false positive rates from 0 to 50 percent across 14 tools on the same documents in 2023, and GPTZero states a rate of no more than 1 percent for its classifier, read on 9 November 2026.

What is a request cap on a humanizer?

A request cap is the most words a tool accepts in one submission, separate from the monthly pool: Walter Writes AI's Starter caps a request at 750 words, StealthGPT's Pro at 1,000 and Clever AI Humanizer at 3,000, all checked on 8 November 2026.