How accurate is Pangram? What it claims and what independent tests show

How accurate is Pangram? Very, by the standards of AI detection: in several independent studies it caught nearly all AI-written text while flagging almost no human writing, and it held up better than rival tools against humanizers. It can still be beaten. A 2026 study found that one commercial humanizer pushed its detection rate below 4%, and it flags lightly AI-edited text often, which matters wherever that kind of help is allowed.

What does Pangram claim about its accuracy?

In an October 2026 post on its current model, Pangram 4, the company sets a target false positive rate of 1 in 10,000 (0.01%). In its internal tests on 2,046,273 human-written documents, it reports doing better than that: about 41 false positives per million, or 0.0041%. On 519,993 AI-generated documents it reports a false negative rate of 0.3396%. Pangram counts a document as a false positive if any human-written segment in it is flagged, which is a strict definition.

The same post lists limits. Pangram is built for prose of 50 words or more; at 50 to 99 words it reports a 0.02% false positive rate and a 1.29% false negative rate. Poetry had the highest false positive rate of any domain, at 0.078%. Pangram says it is deliberately tuned to prefer missing AI text over flagging human text, and that it "leans human" on mixed documents. It names out-of-scope content, including code, machine translations, mathematical notation and reference lists, and advises double-checking PDF results. It supports English and 23 other languages.

These are vendor figures on vendor test sets. They are useful for understanding how Pangram behaves, but they aren't independent evidence. The studies below are. We describe how each was set up in are AI detectors accurate?, so this page keeps to what they found about Pangram.

How did Pangram do in the University of Chicago study?

Brian Jabarian and Alex Imas's 2025 working paper for the Becker Friedman Institute is the broadest test of Pangram so far, with human passages written before 2020 in six genres and AI text from four OpenAI, Anthropic and Google models. Their findings for Pangram:

• On medium-length and long passages, Pangram made essentially no mistakes in either direction.
• Its error rates stayed low on shorter passages, including stubs of under 50 words.
• When the AI passages were rewritten with the StealthGPT humanizer, Pangram still detected nearly all of the longer ones.
• It was the only detector that met a strict cap of 0.5% false positives without losing its ability to catch AI text.

The authors declared no conflicts of interest. It is a working paper rather than a peer-reviewed article, and the authors note that results like these won't stay fixed as models and humanizers change.

How did Pangram compare with expert human readers?

In the ACL 2025 study by Jenna Russell, Marzena Karpinska and Mohit Iyyer, Pangram caught 98.0% of AI-written non-fiction articles with a 2% false positive rate. A newer Pangram model trained on humanized text caught 99.3%, with 2.7%. The study's expert panel, five frequent AI users voting together, caught 99.3% with no false positives, and Pangram was the only automatic detector to keep up with them.

On the hardest set, articles OpenAI's o1-Pro wrote while following a guide to avoiding AI tells, the base Pangram model caught 90.0% and the humanizer-trained model 96.7%. GPTZero fell much further on the same set, as covered in how accurate is GPTZero. The authors thanked both Pangram and GPTZero for API credits.

How does Pangram do on long academic papers?

Researchers at Vrije Universiteit Brussel published a study in June 2026 testing Pangram, Turnitin, GPTZero and Copyleaks on 160 academic papers of at least 4,000 words. Pangram was the only tool of the four that performed well:

• On 40 papers written entirely by ChatGPT's Deep Research mode, it scored 65% as clearly AI and most of the rest as partly AI. Only one was misclassified.
• On 40 hybrid papers, part human and part GPT-4o, and on 40 humanized versions, its estimate fell in the correct range for 37 papers in each group.
• It flagged none of 40 human papers written by non-native English graduate students before ChatGPT.

The authors then ran Pangram on 1,163 real master's theses from their faculty, and it flagged 45.5% for some AI content. With no ground truth for those theses, the authors stress that this figure can't be read as a confirmed rate of AI use. For how Turnitin fared on the same papers, see does Turnitin detect AI?

Where does Pangram fall short?

The clearest counter-evidence comes from a 2026 study by Jonathan Karr and colleagues at the University of Notre Dame, presented at the ACM AI Summit. They took 642 published abstracts in four fields, had Gemini 3 Flash rewrite or write them, and scored everything with Pangram 3.2 and GPTZero.

• Humanizers can beat it. After the AI versions went through the Undetectable AI humanizer, Pangram missed 96.6% to 98.5% of them. GPTZero fared no better.
• Light AI editing gets flagged. Pangram flagged 64% to 80% of abstracts that the AI had only refined. That is a correct detection of AI involvement, but a problem if your rules allow AI polishing.
• Field matters. Pangram flagged 14.9% of unmodified abstracts published in 2023 to 2025, from 2.1% in chemistry to 26.6% in theology. Some of those may contain undisclosed AI help, so the authors don't call them confirmed false positives. Abstracts from 2013 to 2015 were never flagged.

Two smaller points belong here too. Pangram's own research on humanizers, a 2025 paper studying 19 humanizer and paraphrasing tools, was written by Pangram Labs staff, so it is vendor research. And the independent studies tested earlier Pangram versions than the Pangram 4 model behind the company's latest figures.

How accurate is Pangram overall, and can you rely on it?

For screening text, it is among the strongest detectors the evidence supports. Its errors cluster where you'd expect: very short or formatted text, writing that mixes human and AI work, and text deliberately run through a humanizer. Because it leans human on mixed documents, a "human" label doesn't rule out AI help. And like any detector result, a Pangram score is best used to open a conversation with the writer.

At the time of writing, a free Pangram account gets 20 credits a day, which Pangram says is enough for about 2,000 words. Our own detector is free, needs no account, has no word limit and keeps your text in your browser, which suits a first look at long or private documents such as a full manuscript. Its report adds a sentence map and an AI-likeness score, and how it is tested is set out in our open benchmark. Our list of Pangram alternatives compares the options.

Sources

1. Pangram: Is Pangram accurate? (16 October 2026), checked October 2026
2. Pangram: What do I get with a free account?, checked October 2026
3. Jabarian and Imas, Artificial Writing and Automated Detection, BFI Working Paper 2025-116, checked October 2026
4. Russell, Karpinska and Iyyer, People who frequently use ChatGPT for writing tasks are accurate and robust detectors of AI-generated text (ACL 2025), checked October 2026
5. Van Vlasselaer, Van Droogenbroeck and Spruyt, Who wrote this? Evaluating the reliability of AI detection tools in higher education (International Journal for Educational Integrity, 2026), checked October 2026
6. Karr, Khvatskii, Hua and Chawla, Why AI Detection Fails for Academic Integrity (ACM AI Summit 2026), checked October 2026
7. Masrour, Emi and Spero, DAMAGE: Detecting Adversarially Modified AI Generated Text (Pangram Labs, 2025), checked October 2026
8. GPTZero: re-evaluation on the Chicago Booth benchmark (12 January 2026), checked October 2026