Last updated: July 2026
No AI humanizer reliably bypasses Pangram in 2026. Pangram detects 93.66% of high-quality humanized text against GPTZero’s 34.53%, and an independent University of Chicago audit found its false-negative rate stays near zero even against StealthGPT.
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 →Key Takeaways
- Pangram’s humanizer-trained model detects 93.66% of high-quality humanized text, versus 73.07% for its own baseline model — Pangram Labs, Jan 2025
- GPTZero caught just 34.53% of the same humanized text; the open-source Binoculars detector caught 29.73% — Pangram Labs
- An independent audit of 1,992 passages by University of Chicago researchers found Pangram’s false-negative rate is robust to humanizers including StealthGPT, while GPTZero’s rose to 50% or higher — Jabarian & Imas, BFI Working Paper 2025-116
- Pangram 3.2 (Feb 2026) reports a 4x improvement in humanizer detection over 3.1 on its internal evaluation set — a vendor self-report with no absolute numbers published — Pangram Labs
- Pangram CEO Max Spero told Nature that upgraded versions are being designed specifically to detect humanizer use — Nature, 7 July 2026
- Pangram’s own marketing claims 99.98% accuracy; that figure describes raw AI output under lab conditions, not humanized text
- Claims that a specific humanizer “beats Pangram” typically rest on a single unreplicated test of one passage — not a methodology you should bet an academic record on
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 →What is Pangram, and why does it keep coming up?
Pangram is a commercial AI-text detector built by Pangram Labs, used by universities, publishers and recruiters to flag machine-written text. It matters to anyone using a humanizer because it is the one detector that independent researchers have repeatedly found to survive humanization.
That is a different claim from the one most detectors make. Turnitin and GPTZero both advertise high accuracy on raw AI output. Pangram advertises the same, but the evidence suggests it also holds up once text has been rewritten to look human, which is the scenario that actually matters if you are running your draft through a rewriter before submitting.
Screenshot: Pangram (pangram.com), captured July 2026. The homepage advertises 99.98% accuracy — a figure that describes raw AI output, not humanized text.
Does any AI humanizer actually bypass Pangram?
On the published evidence, no — not reliably, and not in a way you could depend on. Pangram’s humanizer-trained model detected 93.66% of high-quality humanized content in the company’s own evaluation, and the only independent audit of the question reached a compatible conclusion using a different corpus and different humanizer.
That does not mean detection is certain. It means a humanizer that clears Pangram once has roughly a one-in-fourteen chance of doing so, on the vendor’s numbers. If you are weighing an academic-integrity hearing against those odds, the odds are bad. Our coverage of bypass rates after humanization and humanizers that beat Turnitin shows how differently the same tools perform against weaker detectors.
What does the independent evidence say?
The strongest evidence is not from Pangram. Brian Jabarian and Alex Imas of the University of Chicago audited four detectors — Pangram, Originality.ai, GPTZero, and an open-source RoBERTa baseline — across a 1,992-passage corpus spanning six genres, with AI text generated by GPT-4.1, Claude Opus 4, Claude Sonnet 4 and Gemini 2.0 Flash. They then ran the AI passages through StealthGPT to test humanizer resistance.
Their finding, in their words: Pangram’s false-negative rate “is robust to the use of current ‘humanizers’ and remains low even when AI-generated passages are modified using tools such as StealthGPT.” GPTZero, by contrast, “largely los[t] its capacity to detect AI-generated text, showing false negative rate scores around 50% and above across most genres and LLM models.”
This is the part worth sitting with. A humanizer that makes your text invisible to GPTZero may leave it fully visible to Pangram. The bypass is detector-specific, and marketing copy rarely says which detector it was tested against.
How does Pangram compare to GPTZero on humanized text?
The gap is not marginal. On Pangram’s published comparison, both detectors handle raw AI text competently; the divergence appears only once a humanizer has been applied.
| Detector | Raw AI text detected | Humanized text detected | Source |
|---|---|---|---|
| Pangram (humanizer model) | 100.00% | 93.66% | Pangram Labs |
| Pangram (baseline model) | 100.00% | 73.07% | Pangram Labs |
| GPTZero | 95.60% | 34.53% | Pangram Labs |
| Binoculars (open source) | 94.40% | 29.73% | Pangram Labs |
One caveat you should apply yourself: this table is the detector vendor grading its own competitors. It is directionally corroborated by the independent UChicago audit, which is why it is here at all. Treat the exact decimals as vendor-reported, and the ranking as evidence-backed.
Why do humanizers fail against Pangram specifically?
Because Pangram trained on them. In January 2025 the company audited 19 publicly available humanizers, built a classification system ranking their output quality, and retrained its production model with high-quality humanized text in the training corpus. The technical work was presented at the Workshop on Detecting AI-Generated Content at COLING and published as arXiv:2501.03437.
That is the structural problem for the humanizer market. A humanizer applies a finite set of transformations — strip em dashes, break the “not just X, but Y” construction, vary sentence length. Those transformations are themselves a fingerprint. Once a detector trains on the output, the very edits meant to hide the text become the signal that identifies it. Our explainer on how humanizers work and why detectors still flag text covers the mechanism in full.
Pangram 3.2, released February 2026, claims a further 4x improvement in humanizer detection over version 3.1 on its internal evaluation set. The company published a model card but no absolute humanized-detection figure for 3.2, so the 93.66% above remains the last number with a published denominator.
What happened with the academic humanizer in July 2026?
On 20 June 2026 a machine-learning researcher at the University of Minnesota, Jie Ding, released an “academic humanizer” aimed at research papers and grant proposals. Nature covered it on 7 July. Its stated core principle was to “strip the AI tells without casualizing”, with rules including avoiding “not just X, but Y” and removing em dashes.
Max Spero, Pangram’s CEO, called the tool “not sophisticated” and said that in initial tests Pangram caught most — though not all — of the humanized language. He added that upgraded versions of Pangram are being designed specifically to detect humanizer use. After Nature put the deception question to Ding, he rewrote the tool’s GitHub description from “removes the usual AI tells” to “sharpens clarity and voice,” and added an ethics note.
The sequence is the whole story of this market in miniature: a humanizer ships, a detector announces it has already trained on that class of tool, and the humanizer’s description quietly softens.
What about the claims that WriteHuman or Undetectable.ai beat Pangram?
They exist, and they are thin. The most-cited version is a single blog test in which one passage run through WriteHuman returned “100% human” on Pangram while the same passage through Undetectable.ai returned “100% AI.” One passage, one run, no reported methodology, no repetition.
Detectors are probabilistic. A single passing result tells you almost nothing about the distribution — a tool with a 6% pass rate will pass sometimes, by construction. Set that against a 1,992-passage audit and the asymmetry is obvious. We maintain full reviews of WriteHuman, Undetectable.ai and StealthGPT, and none of them should be read as Pangram-proof.
What should you actually do about it?
If an institution you answer to runs Pangram, treat humanizers as a liability rather than a shield. The rational move is not a better rewriter — it is documentation. Keep version history, keep drafts, keep the authorship trail. Our guides on building an authorship packet and the safe AI writing workflow for college drafts exist for exactly this scenario.
And be clear-eyed about the other error. Detectors are not harmless. The same UChicago audit that vindicated Pangram’s recall found open-source detectors misclassifying up to 78% of human text as AI, and detector false positives fall hardest on second-language writers — the reason a growing list of universities has dropped AI detection entirely. A detector being hard to fool is not the same as a detector being fair to rely on. Both things are true, and we cover the false-positive side and the ESL bias evidence in depth.
Frequently asked questions
Can any humanizer guarantee it will bypass Pangram?
No. Pangram’s own evaluation puts humanized-text detection at 93.66%, and no vendor publishes a replicated study showing consistent evasion. Any tool guaranteeing a Pangram bypass is making a claim its own category cannot support.
Is Pangram more accurate than Turnitin or GPTZero?
On humanized text, the published evidence favours Pangram heavily — 93.66% versus GPTZero’s 34.53% on the same set, with an independent audit finding GPTZero’s false-negative rate climbing above 50% once a humanizer is applied.
What is Pangram’s false positive rate?
Pangram self-reports roughly 0.19%. The University of Chicago audit found it the only detector meeting a stringent 0.5% false-positive policy cap without losing detection power. Self-reported laboratory figures should still be read cautiously.
Does StealthGPT bypass Pangram?
Not in the one study that tested it directly. Jabarian and Imas used StealthGPT as their humanizer and found Pangram’s false-negative rate remained low, while the other detectors degraded.
Why did one test show WriteHuman passing Pangram?
Because detectors are probabilistic and a single passage is not a sample. On Pangram’s numbers roughly 6% of high-quality humanized text goes undetected, so isolated passes are expected and prove nothing about reliability.
Did Pangram 3.2 change any of this?
It claims a 4x improvement in humanizer detection over Pangram 3.1 on an internal evaluation set, released February 2026. No absolute humanized-detection percentage was published for 3.2, so the older 93.66% figure is the most recent one with a stated basis.
Are AI detectors reliable enough to accuse a student?
Most institutions now say no on their own. Turnitin’s guidance states its scores should not be the sole basis for adverse action, and several universities have discontinued detection software over accuracy and fairness concerns.
Vlad runs DetectionDrama, where he has tested and reviewed more than 60 AI humanizers and detection tools against Turnitin, GPTZero, Originality.ai and Pangram. He publishes AI-detection research at linkedin.com/in/vladivanovseo.
