AI detector
Understands Spanish from Spain and from Latin America. No translating, no detour through English.
0/300 words today
Free: 300 words/day without an account, 500 with one · Pro from $7.49/month · Unlimited with a 3-day trial → See plans
What we measure
- Sentence rhythm
- People alternate long and short sentences irregularly; models drift toward a flat cadence. It is the field's most robust finding and the one that best survives a rewrite.
- The extremes
- Writing produces four-word sentences and forty-word sentences. Generated prose clusters between fifteen and twenty-five.
- Paste residue
- The spaced em dash, invisible characters, curly quotes, Markdown that survived the copy. Text typed by hand does not carry these.
- The slips that are missing
- Flawless spacing across two thousand words, without a single fossil correction, says something. We only count it in long texts, where it means anything.
Why you will not see a percentage
Tools advertising "92% AI-generated" have no way to stand behind that figure: there is no calibrated model that turns these signals into a real probability, and presenting one as though there were invites serious decisions on a fragile measurement. You get a band — no indications, some indications, clear indications — and underneath it the list of what we found, with examples. It is less impressive and it is more honest.
The mistake we refuse to make
This tool ends up pointed at students, and English AI detectors have a documented record of flagging people who did the work — non-native writers most of all, whose prose is measurably more regular. Missing generated text and accusing a person are not errors of equal cost, so the thresholds sit deliberately high, and the calibration corpus includes careful non-native English on purpose. If a change started flagging it, our test suite fails before the change ships.
Under two hundred words we say nothing
Variation measures need a sample. Below two hundred words or eight sentences the result would be noise wearing the costume of data, so we answer grey and tell you why. Paste residue is still reported at any length: an invisible character is an invisible character in twenty words as much as in two thousand.
Frequently asked questions
How does this AI detector work?
It measures two things. The rhythm of the text: how much sentence length varies, how many sentences fall at the extremes, whether each sentence echoes the length of the last, how regular the paragraphs are. And the material traces of a copy-paste out of a chat: spaced em dashes, invisible characters, leftover Markdown, curly quotes. We score sliding windows of four sentences rather than the whole text, so a generated paragraph inside written work does not wash out in the average.
Is it reliable? Can I use it to accuse someone?
No. No AI detector, ours included, is evidence. They all produce false positives and false negatives: careful formal writing by a person can show indications, and generated text that was then rewritten by hand may show none. The result is orientation about style, never a verdict about authorship.
Why no percentage, when everyone else gives one?
Because it would be made up. "87% AI-generated" implies a statistical precision no current technique has. We give a qualitative band and, underneath, exactly what supports it, so you can look at the actual sentences and judge for yourself.
Does it penalise non-native English?
That is the failure mode we calibrate against. Second-language English tends to be more regular in rhythm, which is precisely what naive detectors read as machine-like. Careful non-native prose is part of the corpus the thresholds are set on, and it is pinned in the automated tests, so a change that starts flagging it breaks the build.
How many words does it need?
At least two hundred words and eight sentences. Below that we answer grey instead of handing you a meaningless number. You can analyse up to 300 words a day without an account.
Do you keep the texts I analyse?
No. The analysis runs on our server against the text you send and is not stored: we only record the word count needed to apply your plan's limits.
Reads as AI and you want to fix it? Try the humanizer, see the plans.