AI Text Detector

About this AI text detector

This is a free browser-based detector that estimates how likely a piece of text was generated by an AI model such as ChatGPT, Claude, Gemini, or Llama. Instead of relying on a closed machine-learning classifier, it scores five well-known linguistic signals that AI-generated writing tends to exhibit and combines them into a single readability-style score.

It is meant for quick editorial judgement, classroom discussion, and content review before publishing. The result is a probability indication, not a definitive ruling, and is best used together with your own reading of the text.

The 5 signals this detector checks

1. AI phrase patterns. Looks for phrases that show up unusually often in AI output, such as "in conclusion", "it is important to note", "delve into", "tapestry of", and similar transitional clichés.
2. Sentence length uniformity. Human writing usually mixes short sentences and long sentences. AI output often produces sentences clustered around a similar length. Low variation here raises the AI score.
3. Vocabulary diversity. A measure of how often the same words and phrases are reused across the text. Repetitive vocabulary in a long passage can be a soft AI signal.
4. Transition word overuse. Heavy reliance on words like "moreover", "furthermore", "however", "additionally", and "therefore" is common in AI output, especially in formal explainers.
5. Burstiness. Human writers often vary sentence rhythm. AI output tends to produce more even, predictable rhythm. This is a statistical signal, not a stylistic verdict.

How to read the score

• 0–30%: Likely human-written or heavily edited.
• 30–60%: Mixed. Could be AI-assisted writing, light editing of AI output, or a formal human writer.
• 60–100%: Strong AI patterns detected. Worth a closer review.

None of these ranges is a guarantee. A skilled writer can hit a high AI score by accident, and a well-edited AI draft can score low. Always read the text and compare it with the patterns highlighted by the tool before making a final call.

When the detector works well

• Plain, unedited AI output from models like ChatGPT, Claude, or Gemini.
• Text samples of 200 or more words, where the patterns have enough room to appear.
• Single-author writing where the style is consistent throughout the sample.
• English text. Other languages may produce less reliable signals because the phrase lists are tuned for English.

When the detector may be wrong

• Heavily edited AI output. If a human carefully edits AI text, the AI patterns may be broken and the score may drop.
• Very short text. Below about 50 words, the statistical signals are too noisy to be useful.
• Highly formal human writing. Some academic, legal, and corporate writing naturally uses formal phrases and uniform sentence lengths. This can produce a high AI score even though a person wrote it.
• Mixed-source documents. If parts of the text are human-written and parts are AI-written, the score blends both.
• Translations. Translated text often loses the original rhythm and can score higher than the original would.

Practical use cases

• Editors and content managers doing a sanity check on submitted drafts.
• Teachers opening a conversation with a student about a suspicious assignment, with the score as one input among many.
• Bloggers reviewing their own draft to check if it reads too "machine-like" before publishing.
• Marketing teams auditing copy from freelancers or agencies for AI-assisted output.
• Job applicants reviewing their cover letters and About sections so they sound more like a real person.

Privacy and how your text is handled

All analysis runs locally in your browser using JavaScript. Your text is not uploaded to a server, not stored, and not used to train any model. You do not need to create an account or accept extra tracking to use the tool. Close the tab and the text is gone.

Frequently asked questions

Is this AI detector 100% accurate?

No. No public AI detector is. This tool gives a heuristic probability based on five statistical signals. Use it as a starting point, then read the text yourself before drawing a final conclusion.

Can it detect output from ChatGPT, Claude, or Gemini specifically?

It is model-agnostic. The detector looks at writing patterns rather than a specific model fingerprint, so it works the same way regardless of which model produced the text.

Why does it need at least 50 words?

Statistical signals like sentence variation and vocabulary diversity need enough text to be meaningful. With fewer than 50 words, the score can swing dramatically based on a single sentence and is not reliable.

Does the tool work for non-English text?

English produces the most reliable results. Other languages may still flag rough patterns but the phrase list and transition list are tuned for English writing.

Can I use this to grade student assignments?

Use it as a discussion starter only. Detection scores should never be the sole basis for an academic decision. Treat the score as one signal, then have a conversation with the student about their process before drawing a conclusion.