Are built-in AI detectors accurate? Vendor checkers vs independent ones

Each company states its own figure on its own page. The table lists the claim, the test it rests on where the page names one, and the date read. A homepage claim without a named test set is still a claim; it is simply one that cannot be checked.

Walter Writes AI and Phrasly both ship a detector next to their humanizer; no claim from either is quoted here, and their checkers are reviewed on the same fields on the Walter Writes AI detector review and the Phrasly AI detector review.

The difference is what happens after the flag. An independent detector's business ends when it reports a label. A humanizer vendor's detector reports a label on a page that also sells the fix.

Undetectable AI's wording is precise and worth taking at face value: its checker predicts what "major checkers" would say. That is a model of other models, offered by a company whose humanizer is sold on the same page. The prediction may be good or bad on any given day, and nothing on the page lets a reader tell which.

Weber-Wulff and colleagues tested 12 public detectors plus Turnitin and PlagiarismCheck on 54 documents in 2023, 756 tests in all, with the ground truth known for every document. The headline finding was that no tool reached 80 percent accuracy, and the detail is more useful than the headline (Weber-Wulff et al. 2023, International Journal for Educational Integrity, read 9 November 2026).

Two of the tools in that study, GPTZero and ZeroGPT, are among the five independent detectors this site uses, and both were tested as 2023 models. GPTZero now describes a "Paraphraser Shield" against paraphrasing and homoglyph attacks, read 9 November 2026; whether that changes the 2023 pattern is a question for a dated test, not for a product page. Weber-Wulff also recorded tools that stalled, returned server errors or gave different results on the same text at different times, which is why the method on this site screenshots every reading with its date.

Because they measure different things on different texts. A vendor's false positive rate is measured on the vendor's own human corpus at the vendor's own threshold; Turnitin's is measured only on documents already over its 20 percent display floor. An independent study measures every tool on one shared set with one definition of correct. Weber-Wulff's human-text figure of 96 percent is close to what vendors claim; the gap opens on edited and paraphrased AI text, which is exactly the text a humanizer produces and exactly what a humanizer vendor's own detector is asked to judge.

That is why a 99 percent claim and an under-80 percent finding can both be true statements about different tests, and why neither is a statement about your document today. No detector score should be the only evidence in a misconduct case, and the detector makers say so themselves: Turnitin states its percentage "should not be used as the sole basis for action", and GPTZero states it should be used "as a conversation starter, and not as the final verdict", both read 9 November 2026.

As one more vendor statement, recorded and dated, never as the measure of the humanizer that ships it. The published method runs every output through five detectors that are independent of every humanizer in the test set, GPTZero, Originality.ai, Copyleaks, Pangram and ZeroGPT, reads each one's own label on its own published threshold on the run date, and publishes the screenshot and the raw CSV. Where a humanizer also ships a detector, the first run records that detector's reading of the same output in a separate column, so a reader can set the vendor's checker beside the independent five and draw their own conclusion.

A built-in detector is a useful thing to look at and a poor thing to rely on. The individual checkers are reviewed on their own pages, and the rest of this series sits in the guides hub.