AI detector

In plain English

Almost every tool sold under this name works the same way underneath: it measures how predictable the writing is and reports a number. It does not know who wrote anything.

The category label hides real differences, though, and the differences decide what a result is worth.

A worked example

One paragraph, submitted to three products all described as AI detectors. The outputs are not different opinions about the same question — they are answers to different questions.

  tool A   "38% AI"
           one figure for the document. no passage named.

  tool B   sentence 2: high · sentence 5: high · rest: low
           same measurement, reported per sentence.

  tool C   "no watermark detected"
           did not read the writing at all. checked for a
           signal the generator would have planted.

A and B differ only in resolution, and B is the more useful of the two because a reader can go and look at sentences 2 and 5. Neither is more accurate than the other.

C is doing something categorically different. It made no judgement about style and its "no" means only that no watermark was found — which is the normal state of almost all text, including everything a person wrote.

Read C's output as "not proven machine" and you have understood it. Read it as "proven human" and you have made the mistake this entry exists to prevent.

Why it matters for AI detection

Because "the AI detector said" is a sentence that hides which of these produced the result, and the three carry very different weight. A watermark hit is strong evidence about origin; a document-level percentage is a statistical impression.

The category name also implies a capability none of the style-based tools have. Detecting requires something to detect, and what these measure is predictability — a property of prose that plenty of human writing has, which is the whole false-positive problem in one sentence.

For anyone reading a report: establish which kind of tool produced the number before arguing about the number. The useful follow-up question is not "how accurate is it" but "what did it actually measure".

Commonly confused with

Plagiarism checker

A plagiarism checker compares your text against a database of documents and can show you what it matched. An AI detector compares it against a model's expectations and has nothing to show you.

Watermark reader

Sold under the same name and doing an unrelated job. It looks for a planted signal rather than judging style, so it can confirm origin and can never establish human authorship.

Read next

• How AI detectors work
• Document-level detection
• How accurate AI detectors are
• ZeroGPT vs GPTZero