Is Phrasly AI Detector Accurate? It Caught 3 of 16 AI Texts

Is Phrasly AI detector accurate? On human writing, yes: Phrasly's free AI detector labelled all 12 of our human texts human. On AI writing, no: only 3 of our 16 Claude texts scored 50% or more, and 4 of them were labelled "by a human", on a page that claims 99.8% accuracy.

I've spent nearly 20 years in AI development and digital marketing, and Phrasly is the eighth AI detector I've run through the same 28 texts. I tested it on November 10, 2026, logged what the page sent and received on each check, and compared its "Third Party AI Scores" with the real results those tools gave us on the same texts.

My Main Points:

• Phrasly labelled all 12 human texts human, at 3% to 14% AI, and flagged no sentence in any of them.
• Only 3 of our 16 Claude texts scored 50% or more. The average was 33%, and an Opus 5.5 research introduction that GPTZero and Originality.ai both flag got 9% and the label "by a human".
• Phrasly's Third Party AI Scores are its own predictions. It predicted Originality.ai would flag 3 of our 16 AI texts; Originality.ai flagged all 16.
• The free detector stopped us after 3 checks with a sign up window, although its page says no signup is needed.
• All four AI texts rewritten by Grammarly's humanizer scored 16% AI or less on Phrasly.

Below you'll find every score, how Phrasly compares with the tools in our AI detector accuracy roundup, and what its Turnitin prediction is worth.

Phrasly AI Detector Results on 28 Texts

Phrasly's free AI detector cleared every human text and missed most AI texts on November 10, 2026. The 12 human texts scored 3% to 14% AI, an average of 9.8%. The 16 AI texts averaged 33%: the Opus 5.5 texts 36% and the Fable 5.1 texts 30%.

Phrasly labelled 6 AI texts "by AI", 6 "with a mix of AI and human" and 4 "by a human". Only the two argument essays (100% each) and an Opus 5.5 student essay (63%) crossed 50%, and every text at 31% or more got the "by AI" label.

Text group	Texts	Average AI score	Scored 50% or more	Labelled "by AI"
Human texts	12	9.8%	0	0
Claude Opus 5.5 texts	8	36%	2	4
Claude Fable 5.1 texts	8	30%	1	2
AI texts rewritten by Grammarly's humanizer	4	9%	0	0

Phrasly flagged no sentence in any human text and 86 of the 289 sentences in the AI texts, or 30%. Its scores were stable: a human health paper got 14% twice, and an Opus 5.5 argument essay got 100% twice. Every check went to Phrasly's own scoring service and came back with a score, so Phrasly is a working detector, unlike the free JustDone detector, which gave us random numbers from our test location.

How We Tested Phrasly's AI Detector

Phrasly got the same 28 texts as every detector in our series: 10 plain AI texts (5 from Claude Opus 5.5, 5 from Claude Fable 5.1), 6 AI texts disguised with a casual voice and typos, and 12 human texts written before 2022. The human set is four peer reviewed introductions from PubMed Central, three developer doc sections for Python, PostgreSQL and Django, two personal blog posts, and three 2006 essays by English learners. Texts run 280 to 390 words.

We used Phrasly's free AI detector, labelled Model 7.0, May 2026, on November 10, 2026. The first 3 checks ran without an account. After the sign up window appeared, we created a free account with Google and ran the rest. For each text we pasted it in, clicked Check for AI, saved a screenshot and logged what the page sent and received. That is 28 checks, four extra checks on Grammarly humanizer rewrites and two repeats: 34 checks in all.

All 16 of our AI texts come from Anthropic's Claude models, which Phrasly's page names among the models it detects. We did not test Phrasly's paid plans, Phrasly did not see these results before publication, and no money changed hands.

Phrasly's Third Party AI Scores vs Real GPTZero, Originality.ai and ZeroGPT Results

Phrasly's Third Party AI Scores panel shows a tick, a warning or a cross for Pangram, Originality.ai, ZeroGPT, Turnitin, Copyleaks and GPTZero. The panel's tooltip says Phrasly uses its own AI to predict how those detectors would judge the text. Our network log matches that: the panel's results came from a Phrasly address, and the page sent no request to any of the six services during the checks we logged.

We ran the same 28 texts through GPTZero, Originality.ai and ZeroGPT in October, so we could check the predictions. Counting "AI" and "Likely AI" as a predicted flag and a real score of 50% or more as a real flag, Phrasly's prediction matched the real result on 16 of 28 texts for GPTZero, 15 of 28 for Originality.ai and 18 of 28 for ZeroGPT.

Phrasly predicted GPTZero would flag only 5 of our 16 AI texts and Originality.ai only 3, when in reality both flagged all 16. Phrasly also predicted GPTZero would flag our human health paper, which GPTZero scored 0%.

Is Phrasly AI Detector Accurate for Turnitin?

Phrasly's Turnitin result is the same kind of prediction, made by Phrasly's model without a Turnitin scan. We don't have Turnitin access, so we could not score those predictions directly. Phrasly's predictions for GPTZero and Originality.ai missed 12 and 13 of our 28 texts, so we would not read its Turnitin tick as a forecast of your school's result.

On our texts, Phrasly predicted a Turnitin flag for 5 of the 16 AI texts and for 1 human text: a 2006 essay by an English learner, which Phrasly's own detector scored 14% AI.

Phrasly AI Detector vs the Grammarly Humanizer

Phrasly scored all four Grammarly humanizer rewrites at 16% AI or less: 7%, 16%, 10% and 3%. Three got the "by a human" label and one got "mix". Their originals had scored 9%, 100%, 31% and 27%, so the humanizer took the 100% Opus 5.5 argument essay down to 16%.

On the same four rewrites, Originality.ai and GPTZero caught all four, as we reported in our Grammarly accuracy test. Phrasly predicted that both would call all four human.

Phrasly's Claims vs What We Saw

Phrasly's AI detector page makes four promises that our November 10, 2026 test can check, and on two texts the score it showed differed from what its own model returned.

Claim or feature on Phrasly's AI detector page	What we saw
99.8% accurate	3 of 16 AI texts scored 50% or more; 15 of 28 texts right at a 50% cutoff
Free, no signup, no scan limits	Our fourth guest check was blocked: the server answered "Free limit reached" with a limit of 3, and a sign up window opened
No tracking	During our checks the page also sent requests to Microsoft Clarity and four other data collection addresses
Your text is never stored	With a free account, every result showed "Saved to your history"
The AI score shown	Two texts that Phrasly's model scored 85.2% and 87.5% were shown as 100%

Microsoft Clarity is Microsoft's free tool for session recordings and heatmaps. We don't know what data Phrasly sends it. Phrasly's own blog review, which ranks for Phrasly searches, reports 98.7% accuracy on its own tests, a figure our test did not come near.

Phrasly AI Detector vs GPTZero, Originality.ai, Solvely, ZeroGPT, Grammarly and Scribbr on the Same Texts

Phrasly caught 3 of our 16 AI texts at 50% or more, second lowest in our series after Scribbr at 1. GPTZero, Originality.ai and Solvely caught all 16, ZeroGPT caught 10 and Grammarly 8. None of these seven tools flagged any of our 12 human texts at 50% or more.

Detector	AI texts at 50% or more (of 16)	Average on AI texts	Highest human score	Humanized rewrites caught (of 4)
GPTZero	16	100%	0%	4
Originality.ai	16	93.4%	5%	4
Solvely	16	94.5%	27.2%	not tested
ZeroGPT	10	65.8%	24.9%	0
Grammarly	8	49.8%	15%	0
Phrasly (free)	3	33%	14%	0
Scribbr (free)	1	7.8%	0%	0

Count every score of 50% or more as an AI verdict and Phrasly got 15 of 28 texts right, or 54%. A tool that answered "human" for every text would get 43% on this set. Counting Phrasly's own "by AI" label instead, it caught 6 of 16 AI texts and still flagged no human text.

Is Phrasly Legit? What to Do With a Phrasly Score

Phrasly is a working AI detector run by Phrasly LLC, and it also sells an AI humanizer. When the detector scored a text 63% AI or more, a pop up asked whether to humanize it and listed passing all major AI detectors as a benefit. After we signed up for the free account, Phrasly opened a checkout for a $2 three day trial that then costs $39.99 a month. The free account was enough for this test.

A Phrasly "human" result tells you little about AI use: 4 of our 16 AI texts got that label, and 6 more were called a mix. A high Phrasly score means more, since no human text in our set went above 14%. Phrasly's habit of passing AI text as human matches research on detectors in general: a 2023 study of 14 AI detection tools led by Debora Weber-Wulff found that the tools leaned toward labelling text as human written instead of catching AI.

For students checking their own drafts, Phrasly is unlikely to flag human writing. For catching AI, GPTZero and Originality.ai did far better on the same texts, and Phrasly's predictions of what those tools will say are wrong too often to replace running them.

FAQ

Is Phrasly AI detector accurate?

Only on human writing in our November 2026 test. Phrasly's free AI detector labelled all 12 of our human texts human, but only 3 of our 16 Claude texts scored 50% AI or more, and 4 of them were labelled human.

Is Phrasly AI detector accurate for Turnitin?

Phrasly's Turnitin result is a prediction made by Phrasly's own model, not a Turnitin scan, and we could not test Turnitin itself. Phrasly's predictions for GPTZero and Originality.ai matched the real results on only 16 and 15 of our 28 texts.

Is Phrasly AI detector free?

Partly. Phrasly's page says no signup is needed, but our fourth check as a guest was blocked with a free limit message and a sign up window. A free account then ran our other 31 checks with no limit.

Is Phrasly legit?

Phrasly is a working detector: every check went to Phrasly's own scoring service and gave the same score on repeat. Phrasly also sells an AI humanizer, and after a high AI result the detector offers to humanize the text so it will pass AI detectors.

Does Phrasly store your text?

Phrasly's page says your text is never stored or shared. As a guest, the server's reply marked each check as not kept in history; with a free account, every result showed the line Saved to your history.

Final Thoughts

Is Phrasly AI detector accurate? On human writing it was safe in our test, and on Claude writing it caught 3 of 16 texts at 50% or more. Its Third Party AI Scores are predictions that missed up to 13 of our 28 texts, its free tier stops guests after 3 checks, and a humanizer pop up followed every result of 63% or more. Use it on your own draft if you like, and use GPTZero or Originality.ai when you need to catch AI.

If you need AI tools tested on your own content before you trust them, our custom AI solutions team can run that kind of test for you.