Estimate whether a text reads like AI writing

The four signals this AI detector measures

• Sentence length variation (40% of the score). People mix short punchy sentences with long winding ones. Language models tend to produce sentences of similar length. The tool divides the spread of sentence lengths by the average; values below roughly 0.45 are marked as uniform.
• Typical AI phrases (35%). Stock expressions such as "delve into", "it is important to note", "in today's fast-paced world", "a testament to" or "seamless" appear far more often in AI drafts. The detector lists the ones it found, in English, Spanish, German, French and Portuguese.
• Vocabulary variety (15%). The share of different words in each 100-word window. Very repetitive vocabulary nudges the score up.
• Punctuation pattern (10%). How many of eight marks the text uses (! ? ; ( … " — :) and how often long dashes appear per sentence. Narrow punctuation or heavy dash use counts as an AI-like habit.

The weighted total is shown as a percentage between 2 and 98. From 65 up the label reads "likely AI-generated", from 38 to 64 "mixed or edited", and below 38 "likely written by a human".

Why no detector is proof, including this one

Every signal above also appears in human writing. Formal reports, legal text and academic abstracts have uniform sentences by design. People writing in a second language often rely on safe, common phrases. Templates, style guides and heavy editing flatten variation. On the other side, lightly edited AI text can pass as human easily. Commercial detectors that use machine-learning models have the same basic problem: they produce probabilities, and independent tests have repeatedly reported false positives, including on texts written entirely by people.

That is why this page never says "AI-written" with certainty and why the score must not be used on its own to accuse anyone of cheating, to reject a job applicant or to penalize a writer. A score is a reason to look closer, not a verdict.

Responsible ways to use the result

For teachers and editors, the useful approach is conversation, not detection: ask the writer about their process, look at drafts and version history, or ask them to explain a passage in their own words. For your own writing, the metrics are a style mirror. If your text is flagged for uniform sentences and stock phrases, it probably reads as flat to humans too, whoever wrote it.

If you used AI assistance legitimately and want the result to sound more natural, rewrite it in your own voice; the AI humanizer can help by replacing the most robotic phrases, which you should then review. Always follow the disclosure rules of your school, employer or publication.

Tips for a more reliable reading

• Analyze one author and one piece at a time; mixing quotes, tables or reference lists into the text distorts sentence statistics.
• Remove headings, bullet lists and code before pasting, since they are not normal sentences.
• Compare against a known sample: run a text the same person wrote before AI tools existed and look at the difference, not the absolute number.
• Treat results between 38 and 64 as inconclusive. That band covers most edited and collaborative writing.

Want to write better prompts so AI drafts need less fixing in the first place? Try the AI prompt generator.