Does Undetectable AI bypass Pangram? In independent testing, no — Undetectable AI output was flagged 100% AI-generated by Pangram Labs and scored 72% AI-probability on Quetext, while a competing tool, GPTHuman, passed the same Pangram test at 100% human-written.
Vlad Ivanov — Creator of Words At Scale (26K+ YouTube subscribers). I test AI detectors and humanizers hands-on every week for DetectionDrama, and I don’t take money from the tools I review. Last updated: July 2026.
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 matter for AI humanizers?
Pangram is an AI-detection platform that analyzes perplexity, burstiness, and repetitive phrasing patterns to flag AI-generated text. The company claims roughly 97% accuracy, and independent testing by researchers at the University of Chicago Booth School of Business found it hit near-100% accuracy on unmodified AI text, never dropping below 99.8% across most models tested.
That matters here because Pangram is increasingly the detector humanizer tools get measured against. It’s also the detector at the center of this month’s Nature news story about an academic AI humanizer — Pangram’s CEO said the company is building upgraded versions specifically to catch humanizer use, not just AI-generated text in general.
Does Undetectable AI actually bypass Pangram?
Based on independent testing, not reliably. In a head-to-head comparison, Undetectable AI’s humanized output was run through Pangram Labs and came back 100% AI-generated — not flagged at a lower confidence, not partially caught, a full miss. The same output scored 72% AI-probability on Quetext, a separate detector, meaning the failure wasn’t unique to one tool’s methodology.
That’s worth sitting with if you’ve seen Undetectable AI marketed as a Pangram-proof option. Our own Undetectable AI review covers where the tool holds up and where it doesn’t in more depth.
How did other humanizers perform against Pangram?
In the same class of tests, GPTHuman returned a 100% human-written result against Pangram Labs — a pass where Undetectable AI failed. That doesn’t make GPTHuman a universal answer either; bypass rates vary by text type, length, and which detector you’re checking against, which is exactly why we don’t publish a single “best humanizer” number without a source attached.
| Tool | Pangram Labs result | Quetext result | Source |
|---|---|---|---|
| GPTHuman | 100% human-written (passed) | — | Gold Penguin |
| Undetectable AI | 100% AI-generated (failed) | 72% AI-probability | Anangsha Alammyan testing roundup |
If you’re comparing the full field rather than just these two, our Undetectable AI vs WriteHuman vs StealthGPT comparison and StealthGPT review and WriteHuman review break down the rest of the affiliate-stack tools we track.
How accurate is Pangram in the first place?
More accurate than most detectors on unmodified AI text, less consistent once text has been humanized. Pangram’s false-positive rate averages roughly one in ten thousand academic essays (about 0.004%) in the Chicago Booth testing, and it correctly identified 100% of human-written texts in that same study. But research comparing detectors against heavily humanized text found that no detector consistently identifies heavily humanized AI text — Pangram held up longer than most, but its effectiveness “significantly decreases” once text is run through an AI humanizer repeatedly.
For more on how humanizers and detectors interact in general, see how AI humanizers work and why detectors still flag text and our general guide to bypassing Pangram.
Does today’s bypass rate matter next month?
Less than you’d think, and that’s the point. Pangram’s CEO told Nature this month that upgraded versions of Pangram are being built specifically to detect humanizer use — not just AI-generated text broadly, but the patterns humanizers themselves leave behind. Any bypass percentage you read, including every number in this article, was measured against a moving target. A tool that passed Pangram in June testing is not guaranteed to pass the version Pangram ships in September.
This is also visible on the detector side of the ledger: Turnitin has already added humanizer-specific detection, and our Turnitin AI detection statistics page tracks how those numbers have moved over 2026. If you’re relying on a specific bypass number to make a decision with real stakes attached — a graded paper, a published article — treat it as perishable.
Which AI humanizer should you use if bypassing Pangram matters?
Based on the testing above, GPTHuman is the stronger option against Pangram specifically, and Undetectable AI is not one we’d recommend if Pangram is the detector you’re worried about. Neither claim should be treated as permanent. If you’re combining a humanizer with a detector check before submitting anything, our humanizer-plus-detector combo guide walks through a workflow that doesn’t rely on trusting one vendor’s marketing page.
Methodology note: the test results cited in this article come from third-party testing roundups (Gold Penguin, Anangsha Alammyan’s comparison series, University of Chicago Booth School of Business, Phrasly), not from DetectionDrama’s own lab — we’ve linked each claim to its source rather than restating vendor numbers as fact.
Frequently asked questions
Does Undetectable AI bypass Pangram in 2026?
Not reliably, based on independent testing. Undetectable AI’s output was flagged 100% AI-generated by Pangram Labs and scored 72% AI-probability on a second detector, Quetext, in the same test round.
What AI humanizer passed the Pangram test?
GPTHuman returned a 100% human-written result against Pangram Labs in the same independent testing round that flagged Undetectable AI.
How accurate is Pangram compared to other AI detectors?
Pangram claims about 97% accuracy and was independently verified at near-100% on unmodified AI text by University of Chicago Booth researchers. Its accuracy drops once text has been heavily humanized, which is true of most detectors.
Is Pangram building new detection specifically for humanizers?
Yes. Pangram’s CEO said in July 2026 that upgraded versions of Pangram are being built specifically to detect humanizer use, following coverage of a new academic AI humanizer tool in Nature.
Will Undetectable AI’s bypass rate change over time?
Likely, in both directions. Detector makers update their models in response to bypass tools, and humanizer companies update their output in response. Any bypass percentage published today is a snapshot, not a permanent rating.
Should I trust a single bypass test result?
Treat one test as a data point, not a verdict. Results vary by text length, topic, and which detector you check against — the comparison table above uses two different detectors on the same output for that reason.
