On the published evidence, Pangram is the most accurate commercial AI detector that has been independently tested, and in several studies the margin is wide. Independent studies have measured false positive rates near zero on everyday human writing, and in a 2026 test on long academic papers it was the only one of four detectors that caught AI-written work. That is unusual in this category and it deserves to be said plainly, even by a competitor.
It is not infallible, and the evidence for that is just as specific. A 2026 Notre Dame study found that AI text run through a commercial humanizer slipped past Pangram 96.6% of the time, and that lightly AI-edited human writing was flagged at high rates. Below: what Pangram claims, what each study measured, and where the limits are.
Start with the vendor's own numbers and what they rest on, because Pangram is one of the few vendors that publishes enough to check.
Pangram's homepage says its detector is "over 99% accurate" and puts its false positive rate at "currently 1 in 10,000." The company was founded by Max Spero and Bradley Emi, both Stanford-trained machine-learning engineers, and it sells to individuals, teams and API customers.
Unlike most vendors, Pangram published a technical report (Emi & Spero, arXiv:2402.14873, first posted February 2024, revised July 2024). It describes a benchmark of 1,976 documents across 10 domains and 8 language models, a training method the authors call "hard negative mining with synthetic mirrors," and results on three non-native English essay sets: 0% false positives on 91 TOEFL essays, 0 on 3,907 ELLIPSE essays and 0.09% on 5,600 ICNALE essays.
Two caveats matter. The report is the vendor's own work and is not peer reviewed. And the ESL numbers, which are the most striking in it, have not been reproduced by an independent team that we could find.
The technical report is direct about how the tool should be used: "We strongly discourage the use of our classifier as a sole arbiter of academic integrity and plagiarism checking. All AI detection tools have a nonzero false positive rate, and should be used in conjunction with other evidence." Pangram's dashboard currently requires at least 50 words per scan.
Sources: pangram.com homepage, pricing and knowledge-hub pages, and Emi & Spero, "Technical Report on the Pangram AI-Generated Text Classifier" (arXiv:2402.14873), all read October 2026.
Four independent research teams have published Pangram-specific numbers. Three are strongly in its favour. The fourth found the failure modes.
Brian Jabarian and Alex Imas built a corpus of 1,992 human-written passages across six everyday genres (news, blogs, novel excerpts, Amazon and restaurant reviews, résumés), matched each with AI text from GPT-4.1, Claude Opus 4, Claude Sonnet 4 and Gemini 2.0 Flash, and ran Pangram, Originality.ai, GPTZero and an open-source RoBERTa model across a range of thresholds. The authors declare no conflicts of interest.
Further reading: Is ZeroGPT Accurate? (Lincoln test), on Substack.
| Detector | False positive rate across thresholds | Missed AI text after StealthGPT humanizing |
|---|---|---|
| Pangram | 0.10% to 0% | 0% to 5% |
| Originality.ai | 0.27% to 0.11% | 2% to 21% |
| GPTZero | 0.71% | 26% to 77% |
| RoBERTa (open source) | 97.7% to 89.9% | not comparable |
Their conclusion: "Pangram is the only tool to satisfy a strict cap (FPR ≤ 0.005) without sacrificing accuracy." Two limits on reading that: it is a working paper, not yet peer reviewed, and the humanizer test used one humanizer, StealthGPT, at its default setting. Pangram was not best on every measure: GPTZero missed less GPT-4.1 text than Pangram did. The study also scored passages under 50 words through Pangram's API, where Pangram again did best overall, though it missed 22% of short Claude Opus 4 news passages (GPTZero missed 44% of the same set).
Source: Jabarian & Imas, "Artificial Writing and Automated Detection," NBER Working Paper 34223, 2025. Tables 3, 4, 6 and B.4.
Jenna Russell, Marzena Karpinska and Mohit Iyyer compared human expert readers with automated detectors on 300 articles. Pangram caught 98.0% of AI articles overall with a 2.0% false positive rate, against GPTZero's 85.3% and 0.7%. On articles generated by OpenAI's o1-pro and then humanized, Pangram caught 90.0% and GPTZero 46.7%. The samples are small (30 human and 30 AI articles per configuration, so a single error moves a cell by about 3 points), and the authors thank Pangram and GPTZero for providing API credits. Originality.ai was not in this study.
Researchers in Brussels ran 160 academic papers of 4,000 words or more through Pangram, Copyleaks, GPTZero and Turnitin: 40 human master's theses, 40 papers written by ChatGPT's Deep Research mode, 40 hybrids and 40 humanized hybrids. All four tools cleared all 40 human theses. On the fully AI papers, Pangram placed 39 of 40 in an AI band (26 in the top band), while Copyleaks, GPTZero and Turnitin detected none. On hybrid and humanized papers Pangram's accuracy was 92.5%, against 60.0% and 50.0% for Turnitin, the next best.
Source: Van Vlasselaer, Van Droogenbroeck & Spruyt, "Who wrote this? Evaluating the reliability of AI detection tools in higher education," International Journal for Educational Integrity 22:16 (2026), Tables 4 and 5.
The most recent study is also the most sobering. Jonathan Karr and colleagues ran 642 academic abstracts through Pangram 3.2 and GPTZero. On abstracts written in 2013 to 2015, before modern chatbots, neither tool flagged anything. But:
Their conclusion is that detector scores "should not serve as standalone misconduct evidence," and that honest AI-assisted editing can carry a higher sanction risk than deliberate evasion.
Source: Karr, Khvatskii, Hua & Chawla, "Why AI Detection Fails for Academic Integrity," arXiv:2608.11256 (2026), ACM AI Summit 2026.
A fair audit says what the evidence supports, even about a competitor.
This is the number that matters most to a student or a freelancer, and on it Pangram has one of the best independent records in the category. The Chicago team measured false positive rates at or near zero on everyday writing, and the Brussels team found no false positives on 40 long human theses.
The Brussels study is the only independent test of long, carefully structured AI research papers we know of, and Pangram was the only one of four tools that caught them.
A technical report, a minimum input length, a sole-evidence warning in writing, and an API that academic teams can test. Being checkable is what made the independent studies possible in the first place.
On unedited output from current models, and against at least one popular humanizer, Pangram's miss rates in the Chicago and Maryland studies were low single digits.
Four cases where the "over 99%" figure does not carry over. If your writing sits in one of them, treat any score as provisional.
The Chicago study found Pangram robust against StealthGPT. The Notre Dame study found it nearly blind after Undetectable AI, at a 96.6% miss rate. "Robust to humanizers" is a claim about specific humanizers on specific dates, not a property of the detector.
If you wrote a draft yourself and let an assistant tidy the wording, Pangram may well flag it: 64% to 80% of lightly refined abstracts were flagged in the Notre Dame set. Whether that counts as a "false positive" depends on your institution's policy, which is exactly why the score alone cannot settle it.
The Notre Dame team found flag rates on 2023 to 2025 abstracts that varied by discipline, reaching 26.6% in theology. Some of that may be real AI use. Some of it may not, and the score cannot tell you which.
The Chicago team found that every detector's accuracy drops on shorter passages. Pangram held up best on passages under 50 words, but still missed 22% of one category of short news text. Its dashboard now requires at least 50 words, and a single paragraph is weak evidence on any detector.
A detector score is a reason to look closer, not a finding. This is the sequence that works for people who wrote their own work and were flagged anyway.
Get in writing which detector produced the score, which model or mode it ran, and what score the institution or client treats as actionable. If no threshold is documented, that is worth knowing early.
Two detectors built on different signals disagreeing on the same passage is meaningful and easy to show. Run the identical text through a second tool and keep both results side by side, with timestamps.
Version history is the strongest evidence a writer has, and it is the one thing no detector can manufacture: Google Docs revision history, Word autosaves, notes, outlines, browser history on your sources. A document that visibly grew over many sittings did not arrive from a chatbot.
Most people making these decisions have never read a detector evaluation. Cite the published record rather than arguing about your own score. We keep an appeal-letter template that does this properly.
We sell a detector, so read this section with the scepticism it deserves.
We do not publish a single headline accuracy percentage. One number across every length, genre and register is the move this page has spent its length examining, and it would be no more meaningful coming from us.
What we do publish is one false-positive measurement with the data attached: 1,180 human-written academic abstracts from 2018, before modern chatbots existed, so every AI flag in that set is a false positive by construction. Our detector flagged 5.85% of them, and the per-document results are downloadable so you can check the arithmetic. That figure covers academic abstracts only. It does not tell you how our detector, or Pangram's, behaves on student essays, blog posts or fiction, and we do not claim it does.
The detector is English-only, and we have no independent third-party evaluation yet. Pangram has four. On the evidence that is a real difference, and we would rather say it here than have you discover it.
Sources: our false-positive benchmark (dataset and reproduction script linked on the page) and our accuracy methodology.
Not on its own, and Pangram says so itself: its technical report "strongly discourage[s]" using the classifier "as a sole arbiter of academic integrity." Its independent record is the strongest in the category, but a 2026 Notre Dame study found it flagged 64% to 80% of lightly AI-edited human abstracts and missed 96.6% of humanized AI text. Treat a score as a reason to look closer, alongside drafts and version history.
Pangram advertises "1 in 10,000." Independent measurements depend on the writing. A University of Chicago study of 1,992 everyday human passages measured between 0.10% and 0% across thresholds. A University of Maryland study measured 2.0% on 150 human articles. A Notre Dame study measured 0% on pre-2016 academic abstracts but a 14.9% flag rate on 2023 to 2025 abstracts, which the authors note are not all confirmed false positives.
In the independent studies that tested both, Pangram caught more humanized AI text and had an equal or lower false positive rate on everyday writing, and in a 2026 Brussels study it detected 39 of 40 long AI-written papers where GPTZero detected none. In the University of Chicago study GPTZero missed less GPT-4.1 text than Pangram, and in the Notre Dame study GPTZero flagged fewer recent academic abstracts (8.9% against 14.9%). Neither tool is better on every measure.
Only one independent study has tested both. The University of Chicago NBER paper measured lower false positive rates for Pangram (0.10% to 0%) than for Originality.ai (0.27% to 0.11%), a far lower miss rate on unedited AI text at fixed score cut-offs (under 4% against 11% to 42%), and a lower miss rate on StealthGPT-humanized text. Pangram's own technical report also reports Originality at a 9.24% false positive rate, but that is the vendor's benchmark, not an independent one.
It depends on the humanizer. Against StealthGPT, the University of Chicago study measured Pangram missing 0% to 5% of humanized text. Against Undetectable AI, the Notre Dame study measured it missing 96.6%. No detector has shown it can catch every humanizer, and results from one tool on one date do not carry over to the next.
Pangram's dashboard requires at least 50 words. In the University of Chicago study, which scored passages under 50 words through Pangram's API, Pangram was the most accurate of the detectors tested on them, with false positive rates of 2.5% or less, though it missed 22% of one category of short news text. For a reliable reading, scan a full document rather than a single paragraph.
Pangram says it offers twenty free checks a day after you sign up for its free plan. Its paid Individual plan is listed at $20 a month for 300,000 words, after a 7-day free period, with Professional and Team plans above that. Prices were read from pangram.com in October 2026 and may change.
If you need a different price point or sentence-level evidence, the alternatives compared.
See the alternative →The same audit for the detector most universities reference.
Read the audit →The detector built for publishers, checked against the same studies.
Read the audit →Why human writing gets flagged, and which kinds of writing carry the risk.
Read the guide →1,180 human-written abstracts, our flag rate, and the data to check it.
Check our numbers →Cite the published studies instead of arguing about a percentage.
Use the template →Scan the same passage Pangram scored and compare the two readings yourself. 3 checks a day, no account, and the one false-positive benchmark we quote is published with its data.