How do AI detectors work?
An AI detector is a statistical scorer. It reads your text, measures how closely it matches the AI writing it was trained on, and returns a probability. It doesn't look your text up in a database, and it can't see who typed it.

- Perplexity: how predictable each word choice is. AI text tends to pick the likely next word, so it scores as predictable.
- Burstiness: how much sentence length and structure vary. People mix short and long sentences, and AI text tends to stay even.
- A trained classifier: a model taught on piles of labelled human and AI writing, which sorts new text into one pile.
- A watermark: a hidden pattern the AI model itself adds as it writes, which a matching tool can read back.
Most commercial tools combine the first three. GPTZero's own explainer lists machine learning, natural language processing, classifiers and embeddings, plus perplexity and burstiness. Watermarks work differently, and few tools can read them. Each method is covered below, with the evidence on how well it holds up.
What are perplexity and burstiness?
Perplexity measures surprise. A language model writes by picking likely next words, so text it generates is full of expected choices. A detector runs your text through its own model and asks how surprised that model is. Low surprise leans toward AI, and high surprise leans toward human.
Burstiness is the same idea at the level of the whole document. Human writing swings between a four word sentence and a forty word one. AI text tends to sit in a narrow band. GPTZero says too much of either measure can be a red flag, and that a minor change such as adding or removing a semicolon or em dash will not decide whether a text is flagged.
The trouble is that predictable writing is not always machine writing. A plain, careful style in a second language has few rare words. A textbook definition has to be predictable. OpenAI's own classifier page warns that very predictable text cannot be reliably identified, and gives the first 1,000 prime numbers as an example, since the right answer is always the same.
How do classifiers and watermarks differ from those two signals?
A classifier learns the difference from examples. Developers feed it large sets of labelled human and AI text, and it finds whatever patterns separate them, including ones no person would name. That makes it stronger than a single score, and also harder to explain. Vanderbilt said it saw no insight into how Turnitin's tool works, and listed that among its reasons for switching it off.
A watermark is a different bet. Instead of guessing afterward, the model marks its own output. Google DeepMind says SynthID adjusts the probability scores that decide each word the Gemini app writes, and adds an invisible mark to AI images and video. Its detector section lists images, video and audio, not text. A watermark only exists in output from a model that applied one, so it says nothing about text from other models. AI watermarks goes deeper.
An AI image detector follows the same split. Some tools are classifiers trained on known AI and real pictures, and some read a watermark such as SynthID. Neither the SynthID page nor the vendor pages I read give an accuracy figure for images, so treat any image verdict as a lead, not a result. What is a deepfake covers the harder cases.
Are AI detectors accurate?
Not reliably. Vendors report high accuracy on their own tests, and independent studies find more errors, especially false positives for some writers. The strongest evidence on the false positive side is a 2023 study by Weixin Liang and colleagues.
The researchers ran seven GPT detectors on 91 TOEFL essays written by real non native English speakers and on 88 essays by US eighth graders. The detectors were near perfect on the eighth graders. They misclassified over half of the TOEFL essays as AI generated, and the unanimously flagged essays had significantly lower perplexity. The authors conclude that perplexity based detection can penalize writers with limited vocabulary. The paper is a preprint, and the detectors it tested are from 2023.
A second study, Weber-Wulff and seven co authors in the International Journal for Educational Integrity, tested 14 tools, 12 public ones plus Turnitin and PlagiarismCheck. Its abstract concludes the tools are neither accurate nor reliable, that they lean toward calling text human written, and that disguising the text makes them worse. Read the paper for the figures, which I did not extract.
Three real cases show what that looks like:
- OpenAI's own classifier caught 26% of AI text and wrongly flagged 9% of human text on its challenge set. OpenAI says it withdrew the tool on 20 July 2023 because of low accuracy.
- Vanderbilt noted that Turnitin claimed a 1% false positive rate at launch and that the school sent 75,000 papers in 2022. It disabled the tool in August 2023 and wrote that it does not believe AI detection software is an effective tool that should be used.
- Liang and colleagues found the TOEFL essays above. Ordinary human writing from non native speakers was flagged most of the time.
Which AI detectors are there, and what do they claim?
Five names cover most searches: Turnitin, GPTZero, Originality.ai, ZeroGPT and the retired OpenAI classifier. The table uses each vendor's own page as of November 2026. A claim is the vendor's, not a test result, and a blank says not stated.
| Detector | Free use | Paid price | Accuracy the vendor states |
|---|---|---|---|
| Turnitin AI writing detection | Not stated on its FAQ page | Not stated | False positive rate under 1% for documents with over 20% AI writing; says it can miss some AI text |
| GPTZero | $0: 10,000 words a month, Basic AI Scan, 1 free Advanced Scan | Premium $12.99 a month billed annually, 300,000 words a month | 99% accuracy and a 1% false positive rate in its own explainer; also states 95.7% of AI text caught on the RAID benchmark |
| Originality.ai | $0: the page says 10,000 words once and also lists 60 credits and 3 scans a day, so it is unclear which applies | Pro $12.95 a month billed yearly or $14.95 monthly, 200,000 words a month | Calls its AI checker the most accurate; no figure on the pricing page |
| ZeroGPT | Free, no limits listed | Not shown | Calls its detector advanced and reliable; no figure |
| OpenAI AI text classifier | Retired 20 July 2023 | Not applicable | Caught 26% of AI text and flagged 9% of human text on its challenge set |
Read the gaps as much as the figures. The only paid prices in the table sit on GPTZero's and Originality.ai's pages. Turnitin sells to institutions and shows no price on the page I read. ZeroGPT shows plan names without amounts. The only vendor numbers that include a false positive rate come from Turnitin and GPTZero, and neither reports a result on the Liang test. Originality.ai's pricing page and GPTZero's pricing page are worth rereading before you pay, because plans change.
GPTZero's own explainer adds that no tool can honestly claim to be 100% accurate. That is the best sentence in the table.
How is an AI detector different from a plagiarism checker?
A plagiarism checker compares your text with a stored pile of existing documents and shows matches. An AI detector has nothing to match, because a chatbot writes new wording each time. It measures features of your text, such as predictability and sentence variety, and returns a probability. A clean plagiarism report says nothing about AI, and a high AI score does not mean anything was copied.
Why do different AI detectors give different results?
Each tool trains on different examples, weighs the signals differently and picks its own threshold, the score above which text is flagged. Originality.ai describes its checker as a classifier whose probability is compared with a chosen threshold. Move the threshold and the same text flips from human to AI. Length matters too: OpenAI said its classifier was very unreliable below 1,000 characters, and Turnitin needs at least 300 words. Mixed text, part human and part AI, is the hardest case, since detectors learn from clean examples of each. If two tools disagree about your essay, neither is necessarily wrong, and neither is proof.
Does Turnitin detect AI?
Yes, Turnitin offers an AI writing indicator inside its reports, and it says the result is data for an educator, not a verdict. Its FAQ says the false positive rate is under 1% "for documents with over 20% of AI writing." It also says scores below 20% show an asterisk with no percentage, and that a file needs at least 300 words of prose.
The same page says "there is a chance that we might miss some AI written text in a document," and that Turnitin "does not make a determination of misconduct." It adds that the percentage "should not be used as the sole basis for action." Turnitin says it tests on over 700,000 academic papers written before ChatGPT before each update, which is its own test, not an independent one. For the Grammarly question, see does Grammarly count as AI.
What AI detector do colleges use?
It varies by school, and I found no source that counts them. Turnitin is the name most students meet, because it sits inside the plagiarism checker many schools already license. Others use GPTZero, Originality.ai or another tool, and some use none.
Vanderbilt is the clearest public case of a school opting out. Its August 2023 post said it had decided to disable Turnitin's AI detection tool for the foreseeable future. Your own school's policy decides what counts, and your syllabus or student handbook is where to look. If it's silent, ask in writing. How to use ChatGPT for homework covers the rules most courses set.
What should you do if an AI detector flags your writing?
Treat the flag as a question, not a finding, and answer it with evidence of how you worked. The steps above put that in order. The strongest proof is a trail of dated drafts, notes and sources, because a detector score cannot show process and a document history can.
Two points help in the conversation. Turnitin tells schools not to act on the percentage alone, and the Liang results show the tools misfire on non native writing. Neither is a promise that your case will go your way. A school can still ask you to explain the work, so be ready to. If you used AI help your course allows, say so plainly, and see how to cite AI.
What can you ask an AI about this topic?
These prompts use a chatbot to understand detectors and prepare your case. None of them changes your text to dodge a detector. They work in Claude, ChatGPT, Gemini or Whizi.
- Explain a report: "Here is a detector report summary: [paste]. Explain in plain words what each number means and what it does not prove. Say which questions an instructor should ask next."
- Gather evidence: "I wrote [assignment] over [dates]. List the kinds of proof of my process I should collect, and a checklist to gather them. Add nothing I did not tell you."
- Draft a request: "Write a polite email to [instructor] that asks to see the detector report and the marked passages for [assignment], and offers to explain my work. Under 120 words, no accusations."
- Practise defending it: "Ask me five questions about my paper [paste] that only its author could answer, one at a time. After each reply, tell me where my answer was thin."
- Check the claim: "Search the web and give me the vendor's own stated false positive rate for [tool], with the page link, the date and the exact conditions. Say not stated if it is missing."
- Read the policy: "Here is my syllabus AI policy: [paste]. List what it allows, bans and leaves unclear. Quote the line behind each answer."
Where does Whizi fit?
Whizi isn't an AI detector, and it won't promise you a score. It is one app for the questions around the topic, instead of a pile of single purpose tools. Ask Claude, GPT and Gemini the same question about a detector's claims, with web search on, which is the default in a new chat, and read where they disagree.
The free plan gives 10 messages a day on Whizi AI, Auto and Claude Haiku, with no card. Pro carries Claude, GPT and Gemini together, side by side Compare puts two answers next to each other, and its first week is $1.99. See Whizi pricing. Dedicated detectors like GPTZero are the right tool if you only want a score, and Whizi won't give you one.
What can this page not tell you?
It can't say whether any one text is AI written, and no detector can either. Scores shift with the tool, the text length and the model update. The studies here use 2023 detectors, and vendors have retrained since then, so recent figures may differ.
It also can't predict how your school will treat a flag. Policies differ, and vendors update claims often, so reread the vendor page before you rely on a figure. Where a number came from one vendor's own page, the page says so.
- Read your syllabus AI policy before you submit
- Save dated drafts, notes and version history for every graded paper
- Treat a detector score as a question, never as proof
- Ask for the report and the marked passages if you are flagged
- Reread the vendor page before relying on any accuracy figure
Frequently asked questions
Does Turnitin detect AI?
Yes, Turnitin includes an AI writing indicator and says its false positive rate is under 1% for documents with over 20% AI writing. It also says the percentage should not be the sole basis for action. Scores below 20% show an asterisk, and a file needs 300 words.
Are AI detectors accurate?
Not reliably. Vendors report 99% on their own tests, but the Liang study found seven detectors flagged an average of 61.22% of human TOEFL essays as AI, and a 14 tool study judged the tools neither accurate nor reliable. Treat a score as a lead, not proof.
Is ZeroGPT or GPTZero more accurate?
The vendor pages cannot settle it. GPTZero states 99% accuracy and a 1% false positive rate in its own explainer, while ZeroGPT's pricing page calls itself reliable and gives no figure. Both claims are the vendor's own, as of November 2026.
Can AI detectors spot AI images?
Some can, by classifying pictures or reading a watermark. Google says SynthID adds an invisible mark to images and video made with its tools. The pages I read give no accuracy figure for image detectors, so a verdict is a lead and not proof.