Does StealthWriter bypass Pangram? No test exists. Pangram’s own researchers studied StealthWriter, rated it mid-tier for writing quality, then excluded it from every detection experiment they ran — and have never published a StealthWriter number since.
Key Takeaways
- Pangram’s own researchers studied StealthWriter in their DAMAGE paper, rated it L2 (middle tier) — then excluded it from every detection experiment they ran.
- That L2 rating measures writing quality, not evasion. The paper says so: the tiers are based on “faithfulness and fluency, not on their effectiveness at bypassing AI detectors.”
- StealthWriter is absent from Pangram’s August 2025 benchmark of 20 named tools, with no explanation for the removal.
- In July 2026, Pangram 4 anonymised all 13 tools it tested as “Commercial A” through “Commercial M” — making the question permanently unanswerable from outside.
- Even the best evader among those 13 was still caught 92.78% of the time as AI, and 95.28% as AI-or-Mixed.
- StealthWriter names zero detectors in its own marketing and publishes no bypass percentage — unusual in a category built on naming Turnitin.
- Its refund window is 24 hours and voids the moment you use the product, with no performance guarantee of any kind.
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 StealthWriter bypass Pangram?
Unknown, and unusually for this question the gap is not an accident of who owns what. Pangram Labs has actually had StealthWriter in its hands. It is named in Pangram’s own research paper. It was simply left out of the part that would have answered this.
We searched for any published StealthWriter-versus-Pangram result from any party — Pangram, StealthWriter, academics, competing vendors, Reddit. There is nothing. Not a weak test, not a dated one. Nothing.
This page reviews the published record and reports no first-party test. Every claim below is dated and attributed, and where a source has a commercial stake, we say so.
What did Pangram’s own paper actually do with StealthWriter?
The DAMAGE paper (Masrour, Emi and Spero of Pangram Labs, January 2025) audits a set of humanizing and paraphrasing tools. StealthWriter.AI is in it, listed under “Humanizers” rather than “Paraphrasers”, and given a written assessment: “Reduces quality of the text. There are grammar, punctuation, capitalization issues, generally one per paragraph.” It is graded L2.
L2 is the middle of three quality tiers. And the paper is explicit about what those tiers do not measure:
“It is worth noting we only make these classifications based on faithfulness and fluency, not on their effectiveness at bypassing AI detectors.”DAMAGE, Pangram Labs — arXiv:2501.03437, January 2025
Then comes the sentence that closes the door. Setting up the actual detection experiments, the authors write that they “narrow our scope to L1 humanizers” — the top tier — because those produce the subtlest, most fluent output. StealthWriter is L2. It was studied, tiered, and then dropped before anything was measured about detection.
Why isn’t StealthWriter in Pangram’s benchmark?
Pangram has not said, and the omission is specific. Its August 2025 humanizer benchmark scores 20 tools by name, with a detection accuracy for each — Undetectable AI at 90.3%, StealthGPT at 95.6%, QuillBot and Grammarly at 100.0%. StealthWriter does not appear anywhere in it.
Several tools that were in the DAMAGE paper are missing from that table, while several that were not in the paper appear in it. No selection criteria are published for either the removals or the additions. The benchmark also states no sample size and no methodology.
Then the disclosure narrowed again.
What changed with Pangram 4?
Pangram 4 shipped on 29 July 2026. Its technical report evaluates 13 commercial humanizers, selected, in its words, as “a manually collected set of commercial humanizers chosen by top web traffic in early 2026”.
Every one of those 13 is labelled “Commercial A” through “Commercial M”. Not a single tool is named, anywhere in the report.
StealthWriter had meaningful traffic in early 2026, so it may well be one of them. It may not be. There is no way to establish which, and Pangram has arranged its disclosure so that nobody outside the company can check. For a question like “does this specific tool beat Pangram”, that is the end of the road.
| Pangram publication | Date | Tools named? | StealthWriter present? |
|---|---|---|---|
| DAMAGE paper | Jan 2025 | Yes — all named individually | Yes, rated L2 on quality — excluded from detection tests |
| Humanizer benchmark | Aug 2025 | Yes — 20 named, each with a score | No — dropped, no reason given |
| Pangram 4 technical report | Jul 2026 | No — “Commercial A–M” | Unknowable |
What does StealthWriter itself claim?
Strikingly little, and that is a finding rather than an oversight. We searched its homepage, refund page, changelog and affiliate page for every major detector name — Pangram, Turnitin, GPTZero, Originality, Copyleaks, Winston, ZeroGPT. Zero mentions of any of them. No bypass percentage appears either.
In a category where competitors put “beats Turnitin” in the page title, that silence stands out. Its own FAQ asks “What if AI detectors get updated — will StealthWriter still work?” and answers that the tool is continuously updated so content stays “natural and high-quality”. The question is about evasion; the answer is about prose.
Pricing is a free tier of 10 humanizations a day up to 1,000 words, then $20, $50, $100 and $400 a month. Its models are branded Ghost 5.2 Mini and Pro, with Ghost 5.1 and 4.6 as legacy options. Our full StealthWriter review covers the product.
What do other detectors say about StealthWriter?
The nearest real data point is Turnitin, not Pangram. When Turnitin’s AI bypasser detection shipped in August 2025, an independent academic re-scored a set of humanized documents; StealthWriter’s came back in an “unknown” band between 1% and 19%, with no indication of which passages were flagged.
That reads like a win and mostly is not. A low, unspecified band is Turnitin declining to reach a verdict, and the tester’s framing was that weak scores show the detector struggling rather than the humanizer succeeding — he called totally accurate AI detection “a myth”. In the same run StealthGPT drew a 72% and GPTHuman a 31%.
A competing detector also published a StealthWriter review with a single sample of text. One document is not a measurement. We track the wider disagreement in how much AI detectors disagree.
What should you take from all this?
Three things, in order of confidence. First, there is no evidence StealthWriter beats Pangram, because there is no evidence either way — and that void is now structural, not temporary. Second, Pangram’s own published ceiling for the category is low: across 13 unnamed commercial tools it reports catching humanized text 97.69% of the time, with the best single performer still caught 92.78% of the time. Third, a vendor that declines to name a single detector, publishes no percentage and voids refunds on first use is not behaving like one holding a winning hand.
If your actual worry is being wrongly flagged on your own writing, a rewriting tool is the wrong instrument. Keep your drafting history — it costs nothing and is the strongest defence available, and it matters most for the writers detectors treat worst, covered in ESL writers and AI detection and false positives for neurodivergent students. Check before you submit, using the best pre-submission check and the Turnitin self-check, and strip the obvious tells by hand first with what to remove before using an AI humanizer.
For the tools where Pangram evidence does exist, see Undetectable AI, GPTHuman, WriteHuman and QuillBot, plus the overview at bypassing the Pangram AI detector and the Pangram 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 StealthWriter bypass Pangram?
Nobody knows, because nobody has tested it. We found no published test of StealthWriter against Pangram from any party — not Pangram, not StealthWriter, not an academic, not a competitor. The structural evidence points against it, but no measurement exists.
Didn’t Pangram test StealthWriter in its research?
Partly, and the part that got skipped is the important one. Pangram’s DAMAGE paper lists StealthWriter.AI among the humanizers it studied and rates it L2 for output quality. It then narrows the detection experiments to L1 tools only, so StealthWriter never appears in a detection result.
What does the L2 rating mean?
It is a writing-quality tier, not a detection score. DAMAGE describes L2 as text where quality degrades but intent survives, and its note on StealthWriter reads: “Reduces quality of the text. There are grammar, punctuation, capitalization issues, generally one per paragraph.” For comparison, StealthGPT was rated L1.
Why isn’t StealthWriter in Pangram’s benchmark?
Pangram has not said. It appeared in the January 2025 DAMAGE paper and was absent from the August 2025 table of 20 named tools. Several other DAMAGE tools were dropped at the same time and several new ones added, with no stated selection criteria for either change.
What is Pangram 4 and why does it matter here?
Pangram 4 shipped in July 2026. Its technical report tests 13 commercial humanizers chosen by web traffic — and labels every one of them “Commercial A” through “Commercial M”. No tool is named. So StealthWriter may or may not be in that set, and there is no way to find out.
What is the realistic ceiling for any humanizer against Pangram?
On Pangram’s own published figures, low. Across those 13 anonymised tools it reports detecting humanized text as AI 97.69% of the time and as AI-or-Mixed 98.83%. The single best performer, Commercial H, was still caught 92.78% of the time.
Does StealthWriter claim to beat detectors?
Not by name. Its site names no detector at all — no Pangram, no Turnitin, no GPTZero — and publishes no bypass percentage. Asked in its own FAQ what happens when detectors update, it answers that output will stay “natural and high-quality”, which is not the same promise.
What does StealthWriter cost, and can I get a refund?
A free tier gives 10 humanizations a day up to 1,000 words. Paid plans run $20, $50, $100 and $400 a month. The refund policy allows 24 hours and only if the account has not been used — one humanization voids it — and refunds are at the company’s sole discretion.
