The Detect tower · floor plan

AI detectors: what they can and cannot prove

ai detector is searched 8.289.000 times a month across six English-speaking countries — more than ai image generator, more than vpn, more than any other tool category we measured. Behind that number are two people with opposite problems: a teacher who wants to know if a text was written by a machine, and a student who has been told it was. This page is about what the answer is worth.

What people are searching for

The ten words this floor is built on, measured before a line of it was written:

Also on this floor, measured and kept for the guides below: plagiarism checker 1.700.100 · word counter 1.873.000 · grammar checker 1.866.200 · ai content detector 58.900 · turnitin ai detection 38.400 · ai plagiarism checker 34.300 · ai writing detector 29.780 · ai text humanizer 20.280. Brands are not hub words — people searching quillbot (673.000), turnitin (301.000), gptzero (201.000) or zerogpt (135.000) want that site, not a guide. They belong in reviews, not here.

The geography matters. ai detector is searched 2.240.000 times a month in the Philippines and 550.000 in India, against 4.090.000 in the United States — a market ten times smaller than the American one producing half its volume. That tells you who is being tested, and it is exactly the population the research below is about.

Floor one · What a detector actually measures

No detector can see where a text came from. It never watches you type and it has no record of what a model produced. What it does is statistical: it measures how predictable the text is. Language models are trained to choose likely next words, so their output tends to be smoother and less surprising than human writing, which wanders, repeats itself, and makes odd choices. A detector scores that smoothness and turns it into a percentage.

This has an obvious consequence that vendors rarely put on the front page: anything written in plain, careful, unadventurous English looks like a machine. A student writing in their third language, a scientist following a rigid template, a nervous applicant polishing a cover letter for the twelfth time — all of them produce low-surprise prose. So does ChatGPT.

Floor two · The false positive problem

Three facts, all published by the people best placed to know.

OpenAI withdrew its own detector

OpenAI built a classifier to spot AI-written text in January 2023 and took it offline six months later. Their notice reads: As of July 20, 2023, the AI classifier is no longer available due to its low rate of accuracy. The company that makes the text could not reliably recognise it.

The research found the bias, and named who pays for it

A study by Weixin Liang, Mert Yuksekgonul, Yining Mao, Eric Wu and James Zou, published in the journal Patterns in 2023, tested seven commercial detectors against essays by native and non-native English writers. Its conclusion, quoted from the abstract: These detectors consistently misclassify non-native English writing samples as AI-generated, whereas native writing samples are accurately identified. (arXiv:2304.02819, read 20 August 2026.)

The biggest vendor publishes its own error rate

Turnitin, used by most universities in the English-speaking world, states in its FAQ, read on 20 August 2026: We strive to maximize the effectiveness of our detector while keeping our false positive rate — incorrectly identifying fully human-written text as AI-generated — under 1% for documents with over 20% of AI writing. In other words, we might flag a human-written document as AI-written for one out of every 100 fully-human written documents.

Under 1% sounds excellent. Put it in a university: a mid-sized institution processing a hundred thousand submissions a year, at that rate, wrongly flags up to a thousand innocent papers. Turnitin is transparent about the trade-off too — to keep that rate low, it deliberately misses AI writing: if we identify that 50% of a document is likely written by an AI tool, it could contain as much as 65% AI writing.

And it says plainly what the score is not: Turnitin does not make a determination of misconduct; rather, it provides data for the educators to make an informed decision… the percentage on the AI writing indicator should not be used as the sole basis for action or a definitive grading measure by instructors.

Two more limits from the same page, worth knowing before you argue with a number: a submission must be at least 300 words and no more than 30.000, under 100 MB, in English — and scores between 1% and 19% are not shown at all, because that is where the errors live.

Plagiarism and AI detection are not the same problem

plagiarism checker is still searched 1.700.100 times a month, and it is worth separating the two, because they are often sold in the same box. A plagiarism checker compares your text against documents that exist: the web, journals, a database of past submissions. When it finds a match it can show you the source, and you can go and look at it. That is evidence you can argue with.

An AI detector has nothing to compare against. There is no document to point at, because the text it suspects was never copied from anywhere — it was generated. All it can produce is a probability, and a probability cannot be inspected the way a matched paragraph can. Same interface, same percentage sign, completely different epistemic value.

Why students check themselves first

ai checker — 4.240.000 searches a month, half of them from the Philippines and India — is mostly people running their own work through a detector before handing it in, hoping for a green light. It is a rational reaction to a system that treats the number as a verdict, and it does not work as well as they hope: detectors are trained differently, weighted differently and updated on different schedules, so the same essay can come back at 4% on one and 68% on another. A clean score on the tool you chose says nothing about the tool your institution uses.

Floor three · Humanisers, and the line we do not cross

ai humanizer is searched 1.193.000 times a month. These tools rewrite machine text to raise its unpredictability — swapping words, varying sentence length, adding the small irregularities a detector reads as human. Technically they are paraphrasers pointed at a score.

We will explain what they are, because 1.193.000 people a month deserve an honest answer rather than silence. We will not publish a guide to using one to pass off machine writing as your own. That is not squeamishness: submitting AI text as your work is misconduct at every institution we know of, and a site that helps with it is not an independent guide, it is an accomplice. The same line runs through the whole island — we explain how something works and how to protect yourself; we do not write the how-to for the thing that gets someone expelled.

Worth saying anyway: the humaniser and the detector are the same trick pointed in opposite directions, and both are guessing. Nobody in that arms race can promise you an outcome, whatever the landing page says.

Floor four · If a detector has flagged you

A percentage is not evidence, and the largest vendor in the field says so in writing. If you are on the receiving end of one, these are the things that actually help:

• Keep the process, not just the result. Version history in Google Docs or Word, drafts, notes, search history, the messy first attempt. Provenance beats probability.
• Ask which tool, which score, and which policy. An institution should be able to name all three. If the score is below the vendor's own reporting threshold, that matters.
• Quote the vendor back. Turnitin's own documentation says the indicator "should not be used as the sole basis for action". That sentence was written for exactly this moment.
• If English is not your first language, say so. The peer-reviewed finding above is about you, and it is citable.

If you are the one running the check, the mirror image applies: treat the number as a reason to have a conversation, never as the conversation's conclusion.

What is not on this page yet

Which detector is best. We have not run GPTZero, Copyleaks, Originality.ai, Turnitin and the rest side by side on the same corpus for two weeks, so we are not going to rank them — and we would rather publish nothing than publish a ranking we have not earned. When we do, it will carry the date, the texts we used and the versions we tested.

Want to write it? Claim this floor.