AI Detector

Free AI detector: check if text was written by ChatGPT, Gemini, Claude or other AI, with sentence-level highlights. Runs privately in your browser.

Analysed in your browser: your text and images are never uploaded.
0 words

AI detection gives a probability, not proof. False positives happen, especially with short, formulaic or non-native English text. Never use a score alone to accuse someone.

How to check if a text was written by AI

  1. Paste the text into Text to check, or click Upload file to load a .txt, .pdf or .docx document. The word counter shows how much text you have; 80 words or more gives the most reliable result.
  2. Click Detect AI. The first time, your browser downloads the detection model (about 126 MB). It is cached, so later checks start quickly.
  3. Read the overall score and verdict at the top, then scroll through Passages by AI probability to see which parts drove it.
  4. Use the Writing-style signals table as background information. These figures are not part of the score.

The model runs inside your browser. Your text and files are not sent to our server or anyone else.

How the detector works

The main model is TMR, an open-source RoBERTa classifier trained on the RAID benchmark. RAID contains text from 11 different language models, including text that was deliberately paraphrased or altered to slip past detectors.

Long texts are split into passages of roughly 160 words, and each passage gets its own AI probability. The overall score is the average of the passages, weighted by their length. Passages are highlighted by score: 80% and above, 50–80% and 30–50%; anything under 30% is left unmarked.

When the main score lands in the borderline zone, between 50% and 97.5%, a second model gives its opinion. It is a RoBERTa classifier trained on the HC3 dataset of human and ChatGPT answers (another download of about 126 MB, only when needed). The second model gets most of the weight in the final score, because it rarely flags informal human writing, where the main model is weakest. The result then shows both scores.

How accurate is it?

We tested the detector on 32 texts. The 24 human texts were Hacker News comments written between 2012 and 2017, before ChatGPT existed, and passages from classic novels. The 8 AI texts covered different styles: an essay, a casual review, a technical how-to, a short story, a Slack-style team update and a product review.

  • All 8 AI texts scored between 95% and 99%.
  • 23 of the 24 human texts scored under 50%.
  • One informal human comment was wrongly flagged as AI.

That is a small test, not a benchmark. Expect AI text in common styles to be caught, and expect a real person to be flagged now and then.

Reading the result

The verdict follows the overall score: 85% and above is "Very likely AI-generated", 60–85% "Likely AI-generated or heavily AI-edited", 35–60% "Mixed or uncertain", 15–35% "Likely human-written" and under 15% "Very likely human-written". Below the score you also see the estimated share of AI-like passages, useful for documents that mix human and AI sections.

The writing-style signals describe the text without judging it:

  • Burstiness measures how much sentence length varies. People tend to mix short and long sentences; chatbots often produce an even rhythm.
  • Vocabulary variety is a moving-average type-token ratio, which stays comparable between short and long texts.
  • Common AI phrases lists stock wording chatbots overuse, such as "delve", "tapestry", "it is important to note" or "in today's fast-paced…".
  • Em dashes are counted because frequent use is a well-known AI habit, although many careful writers use them too.

When the detector gets it wrong

Treat the score as a probability, never as proof. Results are less reliable for:

  • short texts (under about 80 words the result is marked as low confidence);
  • lists, bullet points, forms and other text with little running prose;
  • formulaic, legal or technical writing, which is predictable by nature;
  • non-native English, which tends to use simpler, more regular sentences;
  • AI drafts that a person has heavily rewritten;
  • languages other than English.

Detectors can also be fooled on purpose, for example by running AI text through a paraphrasing tool or asking the chatbot to write "like a human". A low score does not prove that a person wrote the text.

Tips

  • Checking a job application rather than an essay? The AI CV & Resume Detector adds buzzword and specificity signals made for CVs.
  • If the text includes sources, run it through the AI Hallucination Checker as well. Invented references are stronger evidence than any score.
  • Suspect copying rather than AI? The Plagiarism Checker compares the text with a source you provide.
Tip: If you are a teacher, ask the student to talk you through the argument or show earlier drafts. A conversation tells you far more than a percentage, and it is fair to a student who was wrongly flagged.

Frequently asked questions

Can it detect ChatGPT, Gemini and Claude?

The main model was trained on output from 11 language models, and the second model on ChatGPT answers, so it is designed to work with text from the common chatbots. No detector catches everything, especially text from newer models.

Is my text uploaded?

No. The models are downloaded to your browser and run there. Your text never leaves your device.

Why is the first check slow?

The first run downloads about 126 MB of model data, plus another 126 MB if a second opinion is needed. The models are cached for later checks.

Can I use the score to prove someone cheated?

No. A score is an estimate, and false positives happen, especially with informal, non-native or formulaic writing. Use it as a reason to look closer, never as the only evidence.

Copied to clipboard