Every AI detector on the internet gives you a percentage. Almost none of them tells you what it measured, and the marketing around them claims a certainty the research does not support. This one works the other way around: it scores only surface patterns it can show you, highlights each one in your text, and this article explains what such a score is worth, with the actual numbers from the accuracy studies. Read it before you act on any detector score, including ours.
What the score is built from
The tool measures five families of patterns, all computed in your browser, none requiring a machine-learning model.
| Signal family | What is measured | Why it leans AI |
|---|---|---|
| Vocabulary | A weighted rate of words like delve, tapestry, pivotal, seamless, per 1000 words | Corpus studies measured "delves" at 28 times its pre-ChatGPT frequency in scientific abstracts |
| Phrases | Fixed templates: "it's important to note", "plays a pivotal role", "not just X, but Y" | Chat models reproduce these moulds at rates human writers do not reach |
| Rhythm | Sentence-length variance (burstiness), connective sentence openers, rule-of-three lists | Generated prose keeps an even beat; humans mix short and long sentences |
| Punctuation | Em-dash rate, the U+202F narrow no-break space | Current models set em dashes unusually often, and U+202F is a paste-from-a-chat artifact word processors rarely produce |
| Formatting | Bullet lists with a bolded lead-in, heading-and-bullet scaffolds | The default answer shape of chat interfaces |
Signals also run in the other direction. Informal markers ("tbh", "gonna"), a lowercase "i" as a pronoun and doubled punctuation subtract points, because assistant-style output essentially never produces them unprompted. Strongly varied sentence rhythm subtracts too.
Everything the score counts appears highlighted in the annotated view. That is deliberate, and it is the feature we think most detectors should have and do not: a number you cannot inspect is an accusation, a highlighted pattern is an observation you can check.
How the number is computed
Each signal contributes points scaled by how far it exceeds normal human rates, with a cap per signal so no single measurement dominates. The points run through a logistic curve to become the 0 to 100 score.
Two safeguards shape the result. A high score requires agreement between at least three independent signal families, so a human text that happens to be full of business jargon cannot score high on vocabulary alone. And below 80 words the verdict is capped at "too little text for a stable reading", because on short texts every method, including the commercial ones, is guesswork; OpenAI's own classifier declared itself unreliable under 1000 characters.
The score is a rate of style evidence. It is intentionally not labelled a probability, because turning style evidence into a probability of authorship requires knowing how often humans write in this style, and that base rate differs wildly between a Reddit comment, a thesis chapter and a press release. Tools that print "87% AI probability" are asserting a precision they do not have.
What a score can and cannot tell you
What a high score genuinely supports: the text matches the default house style of current chat models across several independent dimensions. For the most common real case, someone pasting unedited assistant output straight into a document, that is exactly the evidence you would expect to find, and the annotated view will be dense with highlights.
What it cannot support: any claim about a specific person or a specific text's origin. Every pattern here can be produced by a human, especially one writing formally, in a second language, or to a corporate template. And every pattern can be removed from model output by asking for a different style or making one editing pass. A low score therefore certifies nothing either; edited or deliberately styled model output is pattern-free.
There is also a moving-target problem this tool shares with every detector. The signals are calibrated to the model generation of 2023 to 2025; vendors are actively training the famous tics out, and the next generation will carry fewer of them. Word lists age. Statistical detectors age more slowly, but they age too, which the accuracy numbers below make concrete.
How accurate detectors really are
The gap between marketing and measurement is the defining fact of this product category.
| Source | Finding |
|---|---|
| Vendor claims | 98 to 99.98 percent accuracy, from internal test sets without published methodology |
| RAID benchmark (ACL 2024, 6M texts, 11 generators) | 62 to 85 percent accuracy at a fixed 5 percent false-positive rate; all 12 detectors "easily fooled" by sampling settings |
| Scribbr independent test, 2024 | Best paid tool 84 percent, best free tools 78 percent |
| Perkins et al. 2024, seven detectors | 39.5 percent on unmodified GPT-4/Claude/Bard text, 22 percent after trivial evasion |
| OpenAI's own classifier, withdrawn July 2023 | 26 percent detection, 9 percent false positives |
| Turnitin, per its own blog | Deliberately misses about 15 percent of AI text to keep false positives near 1 percent |
Note what the last two rows mean together: the vendor with the most training data in the world could not build an acceptable detector and shut it down, and the market leader in education openly trades recall for a false-positive rate that still, at Vanderbilt's scale of 75,000 submissions a year, implied roughly 750 wrong accusations. Vanderbilt disabled the feature in 2023, followed by Yale, Georgetown, Northwestern, Michigan State and others.
How model-based detectors work
Knowing the mechanics explains both the accuracy and the failures. Serious detectors use one of three approaches.
Perplexity methods run the text through a language model and ask how predictable each word was. Model output is predictable by construction, so low perplexity leans AI. Refinements compare perplexity across perturbations of the text (DetectGPT) or between two related models (Binoculars, which reaches over 90 percent detection of ChatGPT text at a 0.01 percent false-positive rate on benchmarks, yet drops to roughly coin-flip performance on Claude output it was not tuned for).
Trained classifiers are networks fine-tuned on millions of labeled examples. In their training distribution they are the strongest option; outside it they collapse. The canonical example: the RoBERTa detector OpenAI released for GPT-2 scores about 95 percent on GPT-2 text and 33 to 42 percent, near chance, on modern models. Commercial vendors survive by retraining within days of every model release, which is a subscription service, not a property of the model.
Watermarking biases token choice at generation time and detects the bias later with a secret key. It is real technology (Google's SynthID-Text runs in production) and irrelevant to third-party checking, since only the key holder can detect it, only the vendor's own models carry it, and open models carry nothing.
This page's heuristics are the visible shadow of the first approach: stock words, even rhythm and stock phrases are what low perplexity looks like to the naked eye. The trade is transparency for power. A well-tuned statistical detector beats word counting; word counting shows its work.
The false-positive problem
Every accuracy number above understates the real damage, because errors are not evenly distributed. Detectors flag predictable writing, and predictable writing correlates with groups that do not deserve flags: the Stanford study found 61 percent of non-native speakers' TOEFL essays falsely flagged on average across seven detectors, against near-zero for native eighth-graders. Documented cases cluster around ESL students, autistic writers whose structured style reads as generated, and formulaic genres. GPTZero rated the US Constitution as likely AI; ZeroGPT scored the Book of Genesis at 88 percent AI. Memorised, formulaic and translated text all trip the same wire.
The individual cases are worth reading before using any detector on another person's work: the Texas A&M-Commerce class failed on ChatGPT's own fabricated say-so, or the student put on academic probation because Grammarly edits tripped Turnitin, cleared only by Google Docs version history. In each case the process evidence, drafts and timestamps, settled what the score could not.
Why detection is easy to beat
The evasion numbers are not close. One pass through a paraphrasing model (DIPPER) cuts DetectGPT from 70.3 to 4.6 percent detection; a repetition-penalty flag at generation time costs up to 38 points on RAID; a one-line style prompt was enough in the Stanford study to collapse detection almost entirely. An entire "humanizer" industry automates this, and while its own 99-percent-bypass marketing is as inflated as the detectors' accuracy marketing, good paraphrasers beat most detectors most of the time.
The result is an asymmetry worth stating plainly: anyone trying to evade detection succeeds with seconds of effort, so the population that scores high is dominated by people who did not try, plus innocent writers whose style happens to match. That asymmetry, more than any accuracy percentage, is why a score must never function as an accusation.
Using a result responsibly
Where a pattern score earns its keep: triaging your own inbox of guest-post pitches, checking whether a contractor's deliverable is unedited model output you are paying human rates for, reviewing your own writing for tics that read as generated, or seeing concretely what "AI-sounding" means so you can edit it out. In all of these the annotated view matters more than the number, and the cost of an error is a second look, not a sanction.
Where no score, from this tool or any commercial one, is fit for purpose: grading, academic-misconduct decisions, hiring decisions, or public accusations. The tools that dominate that market print the same warning in their own documentation; Turnitin says its score "should not be used as the sole basis for adverse actions", and GPTZero shows a do-not-punish disclaimer on every scan. If you must evaluate authorship with stakes attached, ask for process evidence: version history, drafts, a conversation about the content. Ten minutes of talking to the author is more diagnostic than any classifier.
Reading the accuracy literature to build this left us with one conclusion: detection is a real but weak signal, useful in aggregate and on your own text, and unfit to judge individuals. We built the tool to show the evidence rather than to deliver verdicts, and if you paste text a person handed you, the fair reading of any score is a question to ask them, not an answer about them.
What detector scores can and cannot tell you
How accurate are AI detectors really?
Independent tests put the best detectors at 65 to 85 percent accuracy, against marketing claims of 98 to 99 percent. Scribbr measured 76 to 84 percent for the leading paid tools in 2024, the RAID benchmark (ACL 2024, 6 million texts) measured 62 to 85 percent at a fixed 5 percent false-positive rate, and a 2024 study by Perkins et al. across seven detectors found only 39.5 percent accuracy on unmodified GPT-4 and Claude output. Accuracy also collapses on short texts, edited texts and anything the detector was not trained on. The vendor numbers come from their own curated test sets; no major vendor has published an independently reproducible benchmark that supports a 99 percent claim.
How do AI detectors like GPTZero actually work?
Almost all of them combine two approaches. Statistical detectors measure how predictable the text is to a language model: perplexity (how unsurprising each word is) and burstiness (how much sentence rhythm varies), where low values lean AI. Classifier detectors are neural networks trained on millions of labeled human and AI texts to recognise the difference directly. Research systems like Binoculars (ICML 2024) compare the perplexity two related models assign to the same text and reach over 90 percent detection of ChatGPT text at a 0.01 percent false-positive rate on benchmark data, but even that method drops to roughly chance level on text from model families it was not tuned for. This page takes a third, simpler route: it counts the surface patterns those statistical properties produce, and shows them instead of hiding them.
What are perplexity and burstiness in AI detection?
Perplexity measures how surprised a language model is by each next word: predictable wording scores low, unusual wording scores high, and generated text is systematically low because models pick likely words by construction. Burstiness measures variation, usually of sentence length or of per-sentence perplexity: human writing mixes short, blunt sentences with long winding ones, while unedited model output keeps an even rhythm. Both are weak individually. Formulaic human writing (legal boilerplate, lab reports, non-native prose) also has low perplexity, which is precisely where detectors produce false accusations.
Why do AI detectors flag human-written text as AI?
Because they measure predictability, and some humans write predictably. The Stanford study by Liang et al. (2023) ran essays by non-native English speakers through seven detectors: on average 61 percent were falsely flagged as AI, one detector reached 97.8 percent, while essays by US eighth-graders passed almost untouched. Limited vocabulary and regular sentence structure produce exactly the low-perplexity signature detectors look for. The same mechanism hits autistic writers, legal and scientific boilerplate, and famously the US Constitution, which GPTZero rated as likely AI-written. If a detector has flagged your own work, that mechanism, not your honesty, is the likely cause.
What words does ChatGPT use most often compared to humans?
The best-measured list comes from Kobak et al. (Science Advances, 2025), who analysed 15 million PubMed abstracts: after ChatGPT launched, "delves" appeared 28 times more often than its pre-2022 baseline, and hundreds of style words spiked, led by delve, pivotal, intricate, underscore, showcase, boast, realm and tapestry. Phrase templates spike the same way: "it is important to note", "plays a pivotal role", "in today's fast-paced world", and the "not just X, but Y" contrast. None of these words is proof by itself; humans used them all before 2022. What is informative is the rate, which is what this tool scores.
Can AI detectors detect paraphrased or reworded AI text?
Mostly no. Krishna et al. (NeurIPS 2023) showed one pass through the DIPPER paraphraser drops DetectGPT from 70.3 percent detection to 4.6 percent at a 1 percent false-positive rate, and it defeats watermark detection and GPTZero as well. Perkins et al. (2024) found even manual tricks that take seconds, such as prompting for a different style or inserting small errors, cut detection by another 17 points from an already low baseline. The RAID benchmark showed a simple repetition-penalty setting during generation costs detectors up to 38 points. That asymmetry is the core problem of the whole category: determined cheaters get through, careful writers with a formulaic style get caught.
Is there an AI detector that is 100% accurate?
No, and the strongest evidence is who has given up. OpenAI shut down its own AI Classifier in July 2023 after six months, citing low accuracy: it caught only 26 percent of AI text while falsely flagging 9 percent of human text. Vanderbilt, Yale, Georgetown, Northwestern, Michigan State and a dozen other universities have disabled Turnitin's AI detection over false-positive risk; Vanderbilt calculated that even the advertised 1 percent false-positive rate would have wrongly accused about 750 of its students in one year. Detection is a moving target: every new model generation writes more like a human, and every detector trained on the last generation slowly goes stale.
How can I prove I wrote a text myself if a detector flags it?
Process evidence beats counter-scores. Google Docs and Word both keep full version history (in Docs: File, Version history) showing the text growing over hours, which is how the accused students at Texas A&M-Commerce cleared themselves in 2023. Git commits, draft files with timestamps, notes and browser research history all work the same way. Running the text through other detectors until one says "human" convinces nobody, since scores disagree with each other routinely. If the accusation is institutional, ask which tool produced it and cite the tool's own disclaimer: Turnitin and GPTZero both state their scores must not be the sole basis for action.
Do AI text watermarks exist, and can you check for them?
They exist, but nobody outside the vendor can check for them. Google ships SynthID-Text in Gemini (published in Nature, 2024): it biases which tokens the model picks, leaving a statistical signature that only Google's own detector, holding the secret key, can read. OpenAI built a text watermark it says is 99.9 percent effective and has deliberately not shipped it. Anthropic does not watermark Claude text, and open models can be run by anyone with no watermark at all. So a third-party checker like this page cannot test for watermarks even in principle, and no character-level marker exists to find: watermarks live in word choice, not in hidden characters.
Does Google penalize AI-generated content?
Not for being AI-generated. Google's published stance since February 2023 is that it rewards helpful content "however it is produced" and penalises content made primarily to manipulate rankings, whether humans or machines wrote it. The March 2024 core update wiped out sites publishing mass low-quality content, much of it AI-made, but the policy target is scaled content abuse, not authorship. Practically: publishing edited, accurate AI-assisted text carries no ranking penalty; publishing thousands of unedited generated pages is what gets sites deindexed.
Can ChatGPT tell you whether a text was written by AI?
No. A language model has no memory of what it or any other model generated, and OpenAI's help pages say directly that ChatGPT cannot recognise its own output. Asked anyway, it guesses fluently, which has already caused real harm: in May 2023 an instructor at Texas A&M-Commerce pasted his students' essays into ChatGPT, took its "yes, I wrote this" at face value and failed the class, and the claims did not hold up. Treat any chatbot verdict on authorship as noise.