humanise.ai editorial mark

AI detection, explained -
how detectors work and where they fail.

Per-detector breakdown, what accuracy claims actually mean, where false positives concentrate, and how to interpret a score responsibly.

Graph paper with navy pencil dots and two overlapping curves, a printed page and a brass magnifying loupe
Read time: 12 min Published: 20 Apr 2026 Updated: 20 Apr 2026

A detector says your text is "98% AI." A different detector says the same text is "7% AI." A teacher treats the first number as proof. An institution builds policy around it.

Something is wrong with this picture, and the thing that's wrong is the assumption that AI detectors produce verdicts. They don't. They produce probabilistic classifier scores, and understanding what that means - what detectors measure, how they measure it, and where they fail - is the difference between using detector output well and treating it like a lie detector.

This guide walks through, in plain English, how the major detectors actually work, what their published accuracy claims mean (and don't mean), where false positives concentrate, and how to interpret a score responsibly.

How do AI detectors work?

AI detectors run text through a trained model that scores each sentence for the probability it was machine written, as GPTZero's technology page describes in October 2026. Turnitin's guide adds limits: it needs 300 words of prose and hides scores from 1 to 19 percent.

The short version of how detectors work

Every major AI detector is a classifier. You give it text. It outputs a probability that the text was generated by an AI model. Behind the probability is a statistical model - usually a machine-learning classifier trained on pairs of human-written text and AI-generated text - that's learned to distinguish them based on a handful of measurable properties of the prose.

The most commonly used properties are:

  • Perplexity - how "surprising" each next word is given the words before it, measured against a reference language model. AI text tends to have low, smooth perplexity because it was generated by picking the most likely next word at each step.
  • Burstiness - the variance of sentence length and complexity across a passage. Human writers burst: short sentence, long sentence, fragment. AI writers flatline.
  • Token-level patterns - frequencies of specific phrases and collocations that models overuse: "intricate," "multifaceted," "it is important to note," "a key insight."
  • Structural features - list density, transition-phrase density, paragraph-length uniformity.

Some detectors add further layers: an ensemble of classifiers, a language-model-based scorer, stylometric features, or a proprietary transformer trained specifically for AI/human discrimination. The branding differs. The underlying mechanism is more similar than the marketing suggests.

How AI detectors work, step by step

Vendors differ in detail, but published documentation and research describe the same broad pipeline.

  1. Tokenise and filter the input. The text is split into tokens and sentences, and anything the model cannot judge is set aside. Turnitin only scores "qualifying text", meaning prose sentences in long form writing, and needs at least 300 words; poetry, scripts, code and bullet lists are not reliably detected (Turnitin guide).
  2. Score predictability. A language model estimates how likely each token was. DetectGPT (2023) showed that machine text tends to sit in negative curvature regions of a model's log probability function, which makes predictability a usable signal.
  3. Measure variation. Detectors look at how that predictability changes across sentences. GPTZero's 2023 explainer on perplexity and burstiness says models "formulaically use the same rule to choose the next word", which leads to low burstiness.
  4. Run a classifier. A model trained on labelled human and AI text turns those signals into a probability. GPTZero describes a deep learning model trained on web, education and AI text from a range of models, with a sentence by sentence classifier (GPTZero technology).
  5. Apply a threshold. The vendor decides where a probability becomes a label. GPTZero says it calibrates to prefer false negatives over false positives, and Turnitin shows an asterisk instead of any score between 1 and 19 percent.
  6. Report the result. Turnitin reports the share of qualifying text that could be AI generated or AI modified. GPTZero reports one of three classes, human, AI or mixed, with confidence categories and highlights.
MethodWhat it measuresKnown weakness
Trained classifier (GPTZero)Patterns learned from labelled human and AI text, scored per sentenceOnly as fair as its training data; GPTZero reports a 1.1 percent false positive rate on TOEFL essays after bias reduction work
Turnitin AI writing model (Turnitin guide)Share of qualifying prose that could be AI generated or AI modifiedTurnitin reports a higher incidence of false positives between 0 and 19 percent; no score under 300 words
Probability curvature (DetectGPT, 2023)Whether small rewrites of a passage lower its log probability under a modelNeeds log probabilities from the model of interest plus many perturbed copies made with a second model
Two model contrast (Binoculars, 2024)How differently two closely related language models score the same textStrong lab results (over 90 percent of ChatGPT samples at a 0.01 percent false positive rate), but RAID found current detectors easily fooled by adversarial attacks and unseen models
Watermarking (Kirchenbauer et al., 2023)Overuse of a hidden "green" list of tokens, tested statistically with p valuesOnly works on text from a model that embedded the watermark; the paper sets out robustness and security as issues to analyse

How each of the main detectors actually works

GPTZero

GPTZero was the first widely-deployed public detector, built on perplexity and burstiness as its two main signals. It computes a perplexity score for the input text using a reference language model (originally GPT-2-based), computes the variance of perplexity across sentences, and outputs a classification and a per-sentence highlight.

Strengths: interpretable, fast, publicly documented. Weaknesses: it was trained primarily on pre-2024 text distributions; newer models produce text with higher perplexity than GPT-3 did, and older human writing (particularly formal or academic) has lower burstiness than it assumed. As of 2026, GPTZero publishes a false-positive rate of around 1% on their internal benchmark - but third-party evaluation puts it higher on non-native-English text.

Turnitin AI writing indicator

Turnitin doesn't publish its methodology in detail. What's known: it's a proprietary classifier trained on a large corpus of academic writing, integrated into the same workflow as plagiarism detection. It produces a percentage score for "AI-generated content."

Turnitin's public position is that their detector is "designed to achieve a false positive rate of less than 1%." In practice, third-party audits have found higher rates, particularly on:

  • Non-native English writers (false-positive rate estimated 2 - 5× higher)
  • Short passages (under 300 words)
  • Heavily edited or revised text of any origin

Turnitin's own documentation warns that their AI indicator "should not be used as the sole basis for any academic action." Many institutions use it that way anyway, which is a policy failure, not a detection failure.

Originality.ai

Originality.ai targets the content-marketing and SEO world rather than education. It produces a binary "AI" / "Original" score plus a confidence percentage, and markets itself on high accuracy against current-generation models.

Its classifier is regularly updated; they publish benchmarks showing high recall on GPT-4 and Claude-generated text. It is one of the detectors most sensitive to paraphrased AI text - standard synonym-swap humanisers rarely fool it, while rhythm-level rewrites often do.

The tradeoff is that Originality.ai's aggressive calibration produces more false positives on formal human writing, particularly marketing copy written in an editorial style. They do not publish a single headline false-positive rate; third-party evaluation places it in the 2 - 4% range on typical content, rising sharply on highly formal prose.

Copyleaks

Copyleaks positions itself across both education and enterprise, with a classifier that ensembles multiple signals (perplexity, stylometry, token patterns) and adds a "source similarity" layer that compares input text against known AI output patterns.

Copyleaks publishes a claimed false-positive rate of 0.2%. Third-party reproduction has been difficult to run at scale; the available published evaluations place its real-world false-positive rate closer to 1 - 3%, varying substantially by input domain.

Winston AI

Winston AI is a newer entrant, competing on speed and on an advertised near-perfect accuracy figure. The mechanism appears to be an ensemble classifier with a transformer-based scorer. It produces a percentage score and per-sentence highlighting similar to GPTZero.

The 99.98% headline is, like most detector accuracy claims, measured on their internal benchmark and not reproducible independently at that figure. Winston's scores on in-the-wild text are often more moderate and less deterministic than the marketing suggests.

What vendor accuracy figures actually mean (and do not)

Detector accuracy claims almost always come from internal benchmarks measured on balanced synthetic datasets - roughly equal amounts of human-written and AI-written text, often drawn from known sources. These are not the distributions detectors face in production.

In the real world:

  • Most submitted text is human-written. If 99% of submissions are human and the detector has a 1% false-positive rate on human text, roughly half of everyone flagged as AI is a human who was misclassified. The Bayesian math is against the detector, and most users don't run it.
  • Human prose distributions differ across populations. A classifier trained largely on native-English university writing will flag non-native English writers more often. The Stanford 2024 study quantified this and the effect is real.
  • AI text distributions are moving targets. GPT-4o, Claude 3.5, Claude 4, Gemini 2 all produce prose that's shaped differently from GPT-3 and each other. A detector trained on last year's outputs has degraded on this year's models.

The result: published accuracy numbers from detector vendors are internal, best-case measurements. Real-world performance is meaningfully worse, and the false-positive pattern is not random - it concentrates on the people least able to defend themselves against the accusation.

Why do different AI detectors give different results?

Different AI detectors give different results because each one uses its own model, training data, eligible text rules and thresholds, so the same essay can score very differently. The main reasons, with sources:

  • Different training data. Each vendor trains on its own corpus. GPTZero names web, education and multi model AI text (GPTZero); Turnitin does not publish its training set.
  • Different thresholds. Turnitin suppresses scores from 1 to 19 percent, while GPTZero tunes toward false negatives (Turnitin guide). A borderline text can show a number on one tool and nothing on another.
  • Different text counted. Turnitin ignores non prose and short submissions; GPTZero classifies sentence by sentence and can return "mixed".
  • Different robustness. The RAID benchmark (2024) tested 12 detectors on over 6 million generations and found them easily fooled by sampling strategies, repetition penalties and unseen models, so results swing with how text was generated.
  • Different bypass defences. Turnitin added AI bypasser detection on 27 August 2025, and GPTZero describes a "Paraphraser Shield". Rewritten text is judged differently depending on which defences a tool has.

For tool by tool detail, read the GPTZero guide, the Turnitin vs GPTZero comparison, our list of free AI detectors and the best AI detectors in 2026.

Where false positives land

Several categories of writing reliably look more AI-like to detectors than they are:

Non-native English writers. Detectors look for uniform cadence, moderate vocabulary, and predictable structure - features that often describe careful second-language writing. Multiple studies (most prominently the 2023 Stanford paper "GPT detectors are biased against non-native English writers") have documented false-positive rates several times higher on this population.

Formal or technical prose. Academic writing, legal drafting, policy documents, and scientific abstracts all use restrained, cadence-uniform styles because the genre demands it. These are precisely the features detectors flag.

Heavily edited text. Prose that's been through many rounds of copy-editing loses the burstiness that early drafts have. Polished writing flattens. Detectors see a flat sentence-length variance and classify it as AI-shaped.

Short passages. Under 300 words, detectors have too little statistical signal to classify reliably. Most vendors acknowledge this in their documentation. In production use, many institutions still run detectors on three-paragraph answers.

Text that was AI-assisted but then substantially rewritten. If the underlying ideas were AI-suggested but the prose was rewritten by a human, classifiers often flag the residual structural patterns even though the final text is, in any meaningful sense, the human's.

If you're in any of those categories, the detector score on your text has a higher-than-average chance of being wrong. That's not an opinion; it's the measured pattern.

How to interpret a detector score

A few rules that hold up across detectors:

  1. A single score is one data point. Run the text through two or three detectors. Meaningful disagreement between them - common at the margins - tells you the score is uncertain.
  2. The per-sentence highlights are more useful than the headline percentage. If the highlighted "AI-like" sentences are all list items, transition-phrase-heavy sentences, or short-generic-opener sentences, you've found the style problem, not evidence of AI authorship.
  3. False-positive risk goes up with formality and down with specificity. Text that names specific people, events, numbers, and places is harder to misclassify as AI than text that stays abstract.
  4. A "98% AI" score on a 200-word answer is unreliable. The statistical base is too thin. Demand longer samples before making judgments.
  5. A score is not a confession. No detector can prove authorship; they can only estimate probability. A high score is a signal to look more carefully, not a verdict to apply penalties against.

What institutions should be doing

The useful institutional response to AI detection is not "more detection." It's policy clarity.

  • Define what AI assistance is and isn't acceptable, in writing, per assignment or per context.
  • Make it normal to disclose AI assistance honestly, and make disclosure not-penalised when it's within the allowed range.
  • Use detectors as one input among many (alongside process evidence, drafts, oral defence, revision history) - never as a single basis for penalty.
  • Train graders on detector limitations, especially on non-native-English bias.
  • When flagging a submission as AI-generated, require evidence that holds up outside the detector: process artefacts, inconsistent knowledge, citation failures.

The institutions that handle AI well treat the tool as a teaching question, not a policing one. The ones that lean hard on detectors as arbiters tend to produce the worst outcomes - false accusations against the most vulnerable students, and no effect on actual bad-faith AI submissions.

What we do about all this

At humanise.ai, we test our rewrite quality against all five major detectors on every release. We publish the measured median pass-rate and stand behind it. We don't promise 100% detector bypass, because nothing honest promises that - detectors update, models update, and an arms race guarantees variance.

We also think the detector arms race is downstream of the real problem, which is that institutions haven't decided what they actually want to happen when AI is part of the writing process. Until that policy question is answered, detector scores will remain what they are: noisy signals being treated as clean verdicts, with human writers - especially the ones least able to push back - bearing the cost of the error rate.

TL;DR

  • Detectors are probabilistic classifiers, not proof. They measure perplexity, burstiness, and token patterns.
  • Vendor headline accuracy claims come from internal benchmarks on balanced synthetic datasets. Real-world performance is meaningfully worse.
  • False positives concentrate on non-native English writers, formal/technical prose, heavily-edited text, and short passages.
  • A single detector score is one data point. Run multiple detectors; look at per-sentence highlights, not just the headline percentage.
  • Institutions that treat detector scores as verdicts make policy errors that harm real people.
  • humanise.ai tests against GPTZero, Turnitin, Originality.ai, Copyleaks, and Winston AI on every release; our median pass-rate is >85% and we publish the number rather than promising miracles.

Next: try the humaniser on a draft you have now, or read Natural AI writing for the stylistic markers detectors look for and how to remove them without gutting the prose.

Frequently asked questions

How does GPTZero work?
GPTZero is a classifier built on perplexity and burstiness. It computes a perplexity score using a reference language model, computes variance of perplexity across sentences, and outputs a classification with per-sentence highlights.
What is Turnitin's AI detection false-positive rate?
Turnitin publicly claims under 1 percent, but third-party audits find higher rates, particularly on non-native English writers (2 to 5 times higher), short passages under 300 words, and heavily edited text.
Can AI detectors prove someone used AI?
No. Detectors are probabilistic classifiers that output a score, not proof. A high score is a signal to look more carefully, not a verdict.
What is perplexity in AI detection?
Perplexity measures how predictable the next word is, computed against a reference language model. AI-generated text has lower perplexity; human-written text has higher and more variable perplexity because humans occasionally reach for unexpected words.
What is burstiness in AI detection?
Burstiness is variance in sentence length and perplexity across a passage. Humans write in bursts: a short sentence followed by a long one, interspersed with fragments. AI-generated text is more uniform. Low burstiness is the single strongest signal detectors use.
Which AI detector is the most accurate in 2026?
On vendor-published figures Winston AI (99.98 percent) and Originality.ai (near-perfect) lead. On independent evaluation the gap narrows and all fail in similar patterns on paraphrased or non-native English text. Pick based on cost and integration rather than headline accuracy.
Are AI detectors biased against non-native English writers?
Yes. A 2023 Stanford HAI study of 7 leading detectors found more than 60 percent of essays by non-native English writers were flagged as AI-generated versus 5 percent for native speakers. 19 percent were flagged by all 7 detectors simultaneously. The bias is systemic to the perplexity-and-burstiness class.
Do AI humanisers defeat AI detectors?
Often, by changing the perplexity and burstiness signatures detectors read. No humaniser produces 100 percent pass-rate reliably because detectors update. Cadence-aware humanisers move the signals more than thesaurus-style synonym swappers.
Can AI detectors detect ChatGPT, Claude and Gemini equally?
Reasonably yes. Mainstream detectors read the shared perplexity-and-burstiness signature. ChatGPT, Claude and Gemini outputs are flagged similarly, with small per-model differences in connective-phrase density and hedging frequency.
Should teachers use AI detectors as evidence of misconduct?
No. Scores should be a flag that triggers investigation, not proof. Using them as evidence will systematically penalise non-native English writers, as the Stanford HAI 2023 research documented.