AI text detector: check writing for AI signals
Estimate AI-writing signals in English or Arabic text on your device: stock phrases, sentence rhythm, personal voice and GPT-2 perplexity, sentence by sentence. Not proof of authorship.
Loading tool…
Read this first: signals, not proof
This tool estimates AI-writing signals. It cannot tell who wrote a text, and no detector can do that reliably. False positives are common for non-native English writers and for formal, technical, simple or short texts. Never use the result as the only evidence in an academic, hiring or disciplinary decision; talk to the writer and look at drafts and notes instead.
Paste a text and this tool measures the writing patterns that chatbot-generated text tends to have: stock phrases such as "it's important to note" or "delve into", sentences of very even length, no personal voice, tidy punctuation and, for English, wording that a language model finds highly predictable. It shows every measurement, explains what each one means, and highlights the sentences that carry the most signals.
Please read the limits before you use the result. No AI detector can tell who wrote a text, and this one does not try to. It gives an estimate of AI-writing signals in three bands: likely human, mixed, and likely AI-assisted. People who write in a second language, formal and technical writers, and anyone writing a short or simple text regularly get "AI-like" readings; a Stanford study of seven commercial detectors (Liang et al., 2023) found that, on average, they labelled more than half of the essays written by non-native English speakers as AI-generated. A person can also edit AI text until the signals disappear. Never treat the result as the only evidence in an academic, hiring or disciplinary decision.
Everything runs in your browser. The writing-style signals are computed instantly for English and Arabic. The optional language-model check runs GPT-2 small (OpenAI, MIT licence) with ONNX Runtime on your device to measure perplexity; it downloads about 144 MB once (the model plus the ONNX Runtime engine, which other on-device tools here share), then works from the browser cache. Your text is never uploaded, stored or used for training.
What the signals measure
Predictability (perplexity, English only): the average surprise of GPT-2 at each next word. Chatbots choose likely words, so their text is unusually predictable. Burstiness: how much that predictability changes from sentence to sentence. Stock phrases: a list of words and phrases chatbots overuse, in English and in Arabic ("في الختام"، "من المهم أن نلاحظ"، "تجدر الإشارة"، "بشكل عام"). Sentence-length variety: people mix short and long sentences; chatbots keep them even. Personal voice and informal punctuation: first person, contractions, colloquial words, brackets, exclamation marks.
Vocabulary variety (a moving-average type-token ratio), repeated phrasing and readability are shown for context but do not change the score, because on our test set they did not separate human from AI text. Every signal has a "why" note in the results table.
How well it works: a small, honest test
We checked the tool on 51 short texts (120-290 words each). The only genuinely human texts are 16 passages from public-domain books (Austen, Chesterton, Melville, Carroll and others, via Project Gutenberg). The "chatbot-style" texts were generated by an AI model, and the "human-style" texts were also written by an AI model imitating a blog post, an email, a forum question, a diary, a technical manual, a lab report and a learner's email, so they are not real human writing. The stock-phrase list and the samples come from the same source, which flatters the result. Treat the table as a sanity check, not an accuracy figure; real-world accuracy will be lower, and we do not quote a percentage.
Where it went wrong is the useful part: two simple children's stories from 1918 and the imitated technical manual and lab report were rated "mixed" once the language model was on, because plain, formulaic human prose is predictable too. An AI-written cover letter full of "I" statements scored "likely human" on writing style alone.
| Test texts | Count | Likely AI-assisted | Mixed | Likely human |
|---|---|---|---|---|
| English, chatbot-style (AI-generated) | 10 | 7 / 9 | 2 / 1 | 1 / 0 |
| English, public-domain books (human) | 16 | 0 / 0 | 0 / 2 | 16 / 14 |
| English, human-style imitations (AI-written) | 9 | 0 / 0 | 0 / 2 | 9 / 7 |
| Arabic, chatbot-style (AI-generated) | 8 | 8 | 0 | 0 |
| Arabic, human-style imitations (AI-written) | 8 | 0 | 0 | 8 |
Sources: Liang, Yuksekgonul, Mao, Wu, Zou: GPT detectors are biased against non-native English writers (Patterns, 2023) (checked 2026-09-22); Vanderbilt University: Guidance on AI detection and why we're disabling Turnitin's AI detector (2023) (checked 2026-09-22)
How to use it
- Paste or type the text (at least 80 words in English or 60 in Arabic), open a .txt, .docx or .pdf file, or try one of the two samples.
- Keep "Include language-model signals" on for English text if you can spare the one-time download of about 144 MB; turn it off on a phone or slow connection.
- Click "Analyze text". The writing-style result appears at once; with the language model on, it is refined when the model has scored every word.
- Read the band, then open each signal in the table to see why it leans one way, and tap highlighted sentences to see what was found in them.
- Copy the report or download it as a PDF, including the limits of the method, if you need to keep a record.
Frequently asked questions
Can this prove that ChatGPT wrote a text?
No. It estimates how many writing patterns typical of chatbot text are present. Some people naturally write that way, and AI text can be edited until the patterns disappear. A "likely AI-assisted" result is a reason to ask questions, for example about drafts and sources, never proof on its own.
Why was my own writing flagged?
False positives are common. Formal, technical, legal and academic writing, text written in a second language, simple writing for children, lists and short texts all tend to have even sentences, no personal voice and predictable wording, which are the same signals chatbots leave. That is why the page shows every signal instead of a single verdict.
Should teachers or employers use it to make decisions?
Not as the only evidence. Use it, if at all, to start a conversation: ask the writer to explain their reasoning, show earlier drafts or notes, or write a short piece in person. Some universities, Vanderbilt among them, have switched off AI-detection features altogether because of false positives.
How short can the text be?
At least 80 words in English or 60 in Arabic. Below that the tool still lists the measurements but gives no band, because a few sentences can look like anything. Longer texts, several paragraphs, give steadier signals.
Does it work for Arabic?
Yes, with the writing-style signals: an Arabic stock-phrase list, sentence-length variety, personal and colloquial voice, punctuation. The language model (GPT-2) is English-only, so Arabic text gets no perplexity signal, which makes Arabic results less reliable than English ones.
Is my text uploaded or saved?
No. The analysis and the language model run in your browser; files you open are read on your device. Nothing is sent to a server, stored or used for training. The model files come from this site and stay in your browser cache.
Which AI models can it detect?
It does not identify models. It looks for patterns common to chatbot answers in general (ChatGPT, Claude, Gemini and similar). Newer models, custom instructions and paraphrasing tools can change those patterns, so treat every result as uncertain.
Related tools
- Word counter and character counterCount words, characters (with and without spaces), sentences and paragraphs, estimate reading and speaking time, and see keyword density.
- Compare text (diff checker)Compare two texts line by line or word by word and see additions and deletions side by side or inline.
- Case converterConvert text to UPPER, lower, Title Case, Sentence case, camelCase, PascalCase, snake_case, kebab-case or CONSTANT_CASE.
- Find and replace textFind and replace text, as plain text or with regular expressions (capture groups, flags), with a count of replacements.
- Add Arabic diacritics (tashkeel) automaticallyAdd harakat (fatha, damma, kasra, sukun, shadda, tanween) to Arabic text automatically in your browser, then fix any letter with a manual tashkeel toolbar.
- Summarise a PDF, Word or text documentSummarise a PDF, Word or text file on your device: key points, action items, tables to Excel, details found and passage search with page numbers.