Does Phrasly bypass Pangram? Nobody has published a result — but Phrasly’s own blog calls Pangram “the most accurate detector available” and says it is “specifically designed to catch humanized AI text”, which is a strange thing to publish if your product beats it.
Key Takeaways
- Pangram’s own paper rates Phrasly L3 — the worst of three tiers: “Poor quality sentences. Much worse in some texts rather than others.”
- That tier got Phrasly excluded from every detection experiment in the paper, which narrowed its tests to L1 tools only.
- Phrasly’s marketing says its models are “ahead of every AI detector on the market” with a “99% Detector pass rate”.
- Phrasly’s own blog says Pangram is “the most accurate detector available” and is “specifically designed to catch humanized AI text”.
- Its own three-detector self-test sweeps ZeroGPT, Originality.ai and Copyleaks — and leaves Pangram out.
- Its Terms of Service state in capitals that it “CANNOT GUARANTEE” output will pass detectors. Refunds require zero usage, ever.
- The one thing nobody has published, including Phrasly: an actual Phrasly-versus-Pangram result.
Want to bypass Turnitin in 2026? Grab the free prompt pack.
Get the exact text-humanization prompts I use to drop an AI score by hand — copy, paste, submit. Free, straight to your inbox.
Send me the free prompts →Does Phrasly bypass Pangram?
There is no test. Not from Pangram, not from Phrasly, not from an academic, not from a competitor, and not on Reddit — where a search for the two names together returns nothing at all, ever.
What makes this one worth writing anyway is that Phrasly has answered the question by accident. The company runs a large content-marketing blog that reviews rival tools and detectors, and one of those reviews is of Pangram. Read it alongside Phrasly’s own product pages and the two documents disagree with each other.
This page reviews the published record and reports no first-party test. Every claim is dated and attributed, and where a source has a commercial stake, we say so.
What did Pangram’s own paper find?
The DAMAGE paper (Masrour, Emi and Spero of Pangram Labs, January 2025) audits humanizing tools for output quality and sorts them into three tiers. Phrasly.ai is listed under “Humanizers”, and its verbatim assessment reads: “Poor quality sentences. Much worse in some texts rather than others.”
It is graded L3 — the bottom tier, described in the paper as tools that “often added nonsensical phrases, words, and characters, constructed incorrect and uninterpretable sentences, and often distorted the meaning of the text.” On the paper’s fluency measure, L3 tools won 2.67% of comparisons against 26.0% for L1.
Then the same thing that happened to StealthWriter happened here, only more so. Setting up its detection experiments, the paper states it will “narrow our scope to L1 humanizers”. Phrasly is L3. It contributed zero data to any detection figure in the paper.
What we can say is narrower and firmer: Pangram looked at Phrasly, judged its prose worst-tier, and never measured whether it beats their detector.
Why isn’t Phrasly in Pangram’s benchmark?
Pangram’s August 2025 benchmark names 20 tools with a detection accuracy for each — Undetectable AI 90.3%, StealthGPT 95.6%, QuillBot 100.0%. Phrasly appears nowhere in it. It was in the January paper and gone by August, with no explanation.
Eleven months later the disclosure closed entirely. The Pangram 4 technical report (July 2026) tests 13 commercial humanizers — selected as “a manually collected set of commercial humanizers chosen by top web traffic in early 2026” — and labels every one of them “Commercial A” through “Commercial M”. No tool is named. Whether Phrasly is in that set cannot be established from outside.
| Pangram publication | Date | Phrasly present? | Detection figure for Phrasly? |
|---|---|---|---|
| DAMAGE paper | Jan 2025 | Yes — audited, rated L3 | None — L3 tools excluded from all experiments |
| Humanizer benchmark, 20 tools | Aug 2025 | No — dropped without explanation | None |
| Pangram 4 technical report | Jul 2026 | Unknowable — all 13 anonymised | None attributable |
What does Phrasly say about itself?
Its product pages are confident. The technology page claims Phrasly’s model is “ahead of every AI detector on the market” and posts a “99% Detector pass rate”. Its humanizer page advertises a “99.7% Quality Score — Tested across 100,000+ documents”, and lists the detectors it says it regularly tests against: Turnitin, GPTZero, Copyleaks, Originality.ai, Winston AI and Sapling. Pangram is not on that list.
The pricing FAQ goes further, claiming “independent audits have confirmed we consistently stay compliant with all major AI detection systems, including TurnItIn.” No auditor is named, no report linked, no date given.
What does Phrasly say about Pangram?
This is where it gets interesting. In May 2026 Phrasly published a review of Pangram on its own blog. It is not a hit piece.
“AI-assisted detection is what sets Pangram apart from the majority of tools… it detects when AI was used to help edit, polish, or rewrite. The gray area that every other detector gets fooled by. This matters because most humanizers only swap words at the surface level. Detectors like Pangram are built to catch exactly that.”Phrasly’s own blog, Pangram AI Detector Review, May 2026
Elsewhere in the same review: Pangram “has one of the lowest false positive rates in the market”, and in its FAQ — “Can Pangram catch reworded AI content? Yes! This is Pangram’s strongest feature.” Two other Phrasly posts describe Pangram as “the most accurate detector available” and as “specifically designed to catch humanized AI text”.
The review runs a live test of Pangram — on ChatGPT output, on human text, on lightly rewritten text. It does not run Phrasly-humanized text through it. That is the one test the article’s own mid-page call-to-action invites, which reads: “Wondering if your humanized content would pass Pangram’s detection?”
Phrasly’s separate self-review does run a detector sweep — ZeroGPT, Originality.ai and Copyleaks, reported as a clean sweep. Pangram is not in it.
What do independent tests show?
Against detectors other than Pangram, the record is mixed and mostly unflattering. A November 2025 test by Originality.ai — a rival detector, so read it adversarially — found its own tool flagged Phrasly’s output at 100% AI both before and after humanizing, Copyleaks likewise, and GPTZero falling only from 56% to 31%. Its author’s summary: “they’re mixed at best.”
A competing humanizer’s 14-tool benchmark, which discloses its conflict up front, placed Phrasly 13th of 14 on evasion against ZeroGPT — while also recording that Phrasly broke zero documents, one of only four tools to manage that. Neither test touched Pangram.
The Reddit picture adds nothing either way. Phrasly’s footprint there is overwhelmingly seeded — near-identical listicles from different accounts, weeks apart, same rankings, scores of zero. The one organic thread we found is a German student comparing checkers, where Phrasly’s detector scored their own thesis at 100%.
What should you take from this?
The honest verdict is that nobody knows, and that Phrasly is the party best placed to find out and has chosen not to publish it. A company confident of beating Pangram would run that test; it invites the question in a call-to-action and then answers a different one.
Set against that: an L3 quality rating in Pangram’s own audit, exclusion from their experiments, removal from their benchmark, a Terms of Service that guarantees nothing, a refund policy voided by any use at all, and a $1,500-per-video influencer programme shaping much of what you will read elsewhere.
If your real concern is being wrongly flagged on writing you did yourself, keep your drafting history — it costs nothing and beats any rewriting tool as a defence. It matters most for the writers detectors treat worst: see ESL writers and AI detection and false positives for neurodivergent students. Check before you submit with the best pre-submission check and the Turnitin self-check, and strip the obvious tells by hand using what to remove before using an AI humanizer.
For the Pangram cells where evidence does exist, see Undetectable AI, StealthGPT, GPTHuman, WriteHuman, QuillBot and StealthWriter, plus the overview at bypassing the Pangram AI detector, the Pangram review and our Phrasly review.
Get the free prompt pack instead
The manual rewriting prompts that lower AI signals without a subscription — and without betting a submission on a number nobody has reproduced. Free, no card.
Frequently asked questions
Does Phrasly bypass Pangram?
No published result exists, from Phrasly or anyone else. What does exist points against it: Pangram’s own paper rates Phrasly worst-tier for output quality, Phrasly’s own blog describes Pangram as the detector built specifically to catch humanized text, and Phrasly’s Terms of Service decline to guarantee that its output passes any detector.
What does the L3 rating mean?
L3 is the bottom of three quality tiers in Pangram’s DAMAGE paper. The verbatim note on Phrasly reads: “Poor quality sentences. Much worse in some texts rather than others.” It measures prose quality, not evasion — the paper says the tiers reflect “faithfulness and fluency, not… effectiveness at bypassing AI detectors.”
Doesn’t worse writing make a tool harder to detect?
Sometimes, and this is the honest complication. Pangram itself observed that “the ‘good’ humanizers in terms of fluency are more detectable, and the ‘bad’ humanizers are less likely to be detectable.” So an L3 rating is not automatically bad news for evasion — but the trade is text a human reader can see is broken.
Why was Phrasly excluded from Pangram’s detection tests?
On quality grounds. The paper narrows its experiments to L1 humanizers because “their subtle changes are the hardest to detect by eye… making them most relevant in real-world adversarial attacks.” Phrasly is L3, so it contributed no data to any detection figure in that paper.
What does Phrasly itself say about Pangram?
Its product pages never mention Pangram, and the six detectors it says it regularly tests against do not include it. Its blog, however, published a Pangram review calling it “the most accurate detector available” and stating that catching reworded AI content is “Pangram’s strongest feature.”
Is there any independent test of Phrasly?
Against other detectors, yes. A rival detector’s November 2025 test found Originality.ai flagged Phrasly’s output at 100% AI both before and after humanizing, with GPTZero dropping from 56% to 31% — results its own author called “mixed at best.” A competitor’s 14-tool benchmark ranked Phrasly 13th on evasion. Neither used Pangram.
What does Phrasly cost, and can I get a refund?
A 3-day trial is $2.00 and explicitly non-refundable. The Unlimited plan displays at $10.99/month billed annually. Refunds are 14 days but only if there has been “absolutely no usage on your account whatsoever” — any use at any point voids eligibility.
Should I trust Phrasly reviews on YouTube?
Check for a disclosure. Phrasly’s affiliate programme pays 20% recurring and advertises an influencer tier of “up to $1,500/video” for YouTube or TikTok content. Assume a video review sits inside that programme unless it says otherwise.
