Phrasly AI detector review

Phrasly's AI detector is free with no sign-up, requires at least 30 words, and claims 99.8 percent accuracy on its detector page, read 9 November 2026, with no public test behind the figure. This Phrasly AI detector review covers what the vendor says the checker measures, what the page leaves out, and why the humanizer company's own checker cannot be the judge of its own humanizer. The paid rewrite tool is reviewed separately in the Phrasly review, and the combined tools sit in the humanizer and detector list.

What does Phrasly publish about its method?

The detector page is unusually talkative about method for a humanizer vendor, read 9 November 2026. It states that the models are proprietary, trained on more than 1 million human-written articles rather than on AI output, and that the detector looks for four things: variation in sentence length and structure, generic versus specific phrasing, how ideas connect across a document, and unnaturally consistent tone. It returns a percentage and, for signed-in users, highlights the sentences that drove it.

The accuracy claim is 99.8 percent, repeated in the meta description, the header and three FAQ answers. The same FAQ states that no method of detection is 100 percent accurate and that false positives are inevitable, read 9 November 2026. No dataset, date, threshold or third-party test for the 99.8 figure appears anywhere on the page, so it stands as a claim the vendor makes about itself. GPTZero's technology page, by comparison, states a false positive rate of no more than 1 percent and a 1.1 percent false positive rate on TOEFL essays (GPTZero, How GPTZero works, read 9 November 2026); those are also self-reported, but they are rates with a named test set rather than a single headline number.

What does the research say about paraphrased text and detectors?

Phrasly sells a humanizer as its main product, and the detector is the free door to it. The published research explains why a rewrite moves detector readings. Weber-Wulff and colleagues tested 14 AI text detectors and found every one below 80 percent on their accuracy measure, with accuracy falling further on paraphrased text (Weber-Wulff et al. 2023, International Journal for Educational Integrity, read 9 November 2026). Krishna and colleagues showed that a paraphrase model dropped DetectGPT's detection rate from 70.3 percent to 4.6 percent at a 1 percent false positive rate (Krishna et al. 2023, arXiv, read 9 November 2026).

A detector that is good at catching rewritten text would, by the same token, catch its own company's rewrites. Phrasly's page states that the checker is designed to identify rewritten and obfuscated AI content. Whether that applies to Phrasly's own humanizer output is a question the page does not answer, and could not answer credibly. No detector score should be the only evidence in a misconduct case.

Does Phrasly treat its own detector as independent evidence?

No, and that is to its credit. The detector page carries a table from a vendor-run benchmark of four humanizers dated September 2026, and the caption states that every output was scored with Pangram and that "those scores come from Pangram, not from the checker on this page", read 9 November 2026. The vendor chose an outside detector to grade its own humanizer because a reading from its own checker would persuade nobody. That is the same reason this site's published method scores every tool on five detectors no humanizer vendor owns, and never on a vendor's own checker.

What does the Phrasly AI detector report?

A percentage for the whole document, a red highlight on the sentences that look AI-generated to the model, and no label threshold. The page's instructions say to check the highlighted parts, rewrite them and test again, read 9 November 2026. It also states that the model aims to tell apart fully AI-written text, human writing edited with AI, and human writing that happens to match AI patterns, without saying how those three are shown in the result.