ProofreaderPro.ai
AI Text Humanization

Pangram AI Detector Review 2026: Accuracy, False Positives

Is Pangram's AI detector accurate? Its 99% and 1 in 10,000 claims checked against the Chicago, RAID and ACL studies, plus short-text limits and pricing.

Dana - Author at ProofreaderPro.aiDana|Oct 6, 2026|Last updated on |28 min read
pangram ai detector - ProofreaderPro Blog

Pangram is an AI detector from Pangram Labs, and the company says it's more than 99% accurate. It also says it wrongly flags about 1 in 10,000 human texts. Its current model, Pangram 4.1, came out on October 19, 2026. Pangram points to outside studies from the University of Chicago and the University of Maryland that it says found similar results.

Independent tests back most of Pangram's accuracy claims for long English text. A University of Chicago working paper found near-zero errors on medium and long passages. That test used an older Pangram model and no academic writing, though, and about 2,000 human passages are too few to confirm a rate as low as 1 in 10,000.

What is Pangram?

Pangram is an AI detection tool built for schools, publishers, law firms and other organizations that need to tell human and AI writing apart. You paste in text or upload a file. Pangram returns a prediction of Human, AI-Generated or AI-Assisted, plus a score for how much of the document falls into each label.

Pangram's AI detector tool showing the paste-text box, free-credit and 3rd-party-verified badges, and the Check for AI button

Pangram Labs started in Brooklyn, New York in 2023. Its founders worked on AI research at Tesla and Google before starting the company. Next to the text detector, Pangram sells a Similarity Checker for plagiarism and an AI image and video detector called Pangram Image 1.0. It also has a browser extension, an MCP server and an API. It also works inside Gmail, Google Docs and learning management systems such as Canvas, Moodle, Google Classroom and Brightspace.

Pangram 4.1 needs at least 50 words of input and works best on complete sentences. Short replies and bullet points give it less to work with. Each result includes a confidence level of High, Medium or Low for every segment, along with highlighted sections that mark which sentences the model labels AI-Generated, AI-Assisted or Human.

An AI-Assisted label means Pangram thinks AI edited or polished writing that a person started. If you used a grammar tool or an AI rewriter on parts of your text, that label can appear even when the ideas and structure are yours. Check what your school or employer allows for AI editing tools before you treat the label as a problem.

How to use Pangram's AI detector

  1. Create a free account at pangram.com. The free plan gives you 20 checks a day.
  2. Paste your text, or upload PDF, DOCX or RTF files, up to 100 at a time. Pangram recommends raw text or .docx files, because converting a PDF can add errors to the text.
  3. Remove headers, footers, instructions and the reference list before you scan. Pangram treats reference lists, tables of contents and other front and back matter as out of scope.
  4. Click Check for AI, then read the label, the percentage and the highlighted segments.
  5. Open your scan history to download the report or compare it with a later scan.

Pangram's Similarity Checker

Pangram's Similarity Checker is its plagiarism tool, and it runs in the same scan as the AI detector. It compares your text with billions of webpages, academic papers, news articles and books. It flags direct matches and lists their sources next to the AI result. Similarity checks are included on every scan in the paid plans, and the free plan covers AI detection only. Teachers can also run it inside Canvas and Google Classroom.

The two checks answer different questions. Pangram's AI detector looks only at the text you submit. It doesn't compare it with other submissions or the web, and it doesn't record how the text was typed. The Similarity Checker is the part that looks for copied text.

Pangram 4.1: what changed

Pangram 4.1 is the successor to Pangram 4, released on October 19, 2026. Pangram says it improves overall accuracy and how the model behaves when it's unsure. The inputs, training and labels are the same as in Pangram 4.

Three changes affect what you see in a result:

  • Predictions change less after small edits. Across 29,103 pairs of texts that differed by a small edit, the share of verdicts that flipped fell from 1.08% with Pangram 4 to 0.15% with Pangram 4.1.
  • Each segment now also returns class probabilities, the model's estimate for Human, AI-Assisted and AI-Generated.
  • The dashboard shows less certainty for text under 150 words.

Pangram also changed its main test. It cleaned its English test set of spam, advertising and machine-made pages, which makes the set harder. On the cleaned set of 781,981 human documents, Pangram 4.1 had 70 false positives, or 0.0090%. That's close to the company's 1 in 10,000 target.

BenchmarkDocumentsPangram 4.1 result
English human documents, false positives781,9810.0090%
Human documents in 104 languages, false positives1,000,0000.0033%
English AI documents from 26 models, missed520,0000.3652%
AI documents in 18 languages, missed190,1491.99%

These are Pangram's own figures from its Pangram 4.1 model card. Pangram also re-ran the full-length set from the University of Chicago study with Pangram 4.1. On those 15,936 passages it reports no false positives and no misses.

What "accurate" means for an AI detector

A detector makes two kinds of mistakes, and one headline percentage hides the difference. A false positive is human writing labelled as AI. A false negative is AI writing labelled as human. The false positive rate (FPR) is the share of human texts wrongly flagged, and the false negative rate (FNR) is the share of AI texts missed.

Every detector has a threshold, a score above which a text counts as AI. Raising it lowers false positives and raises false negatives. Lowering it does the opposite. A study that reports only "accuracy" at one unknown threshold can't tell you which mistake the detector prefers to make.

The size of the human sample matters as much as the rate. Suppose a test uses about 2,000 human passages and flags none. The statistical "rule of three" then puts the true false positive rate below about 0.15%, or about 1 in 670. A claim of 1 in 10,000 needs hundreds of thousands of human documents, which is why Pangram quotes tests that large and independent labs mostly can't repeat.

Writers care about the false positive rate most, because a false positive is an accusation. Institutions that scan thousands of submissions care about both. Keep that split in mind for every number below.

The evidence at a glance

Six sources make up most of the evidence. They differ in who ran the test, what text was used and how recent the AI models were.

SourceWho ran itDateWhat it tested
Pangram technical report (Emi and Spero)PangramFebruary 2024, revised July 20241,976 documents, 10 domains, 8 language models, plus three non-native English datasets
RAID shared task overview (Dugan et al.)Shared task organizers, with Pangram taking partJanuary 2025AI text caught at a 5% false positive rate, with and without attacks
Russell, Karpinska and Iyyer, ACL 2025University of Maryland researchersPosted January 2025, revised May 2025300 articles read by people and by five detectors
Jabarian and Imas, NBER Working Paper 34223University of Chicago BoothSeptember 20251,992 human passages, 4 AI models, 4 detectors, StealthGPT output
Pangram 4 technical report and model cardPangramJuly 2026Over a million human documents, 519,993 AI documents, 104 languages
Pangram 4.1 model cardPangramOctober 2026781,981 English human documents, 520,000 AI documents, 104 languages

Two of the six are independent of the vendor, and only one of those is a controlled comparison of several commercial detectors on the same text. The Chicago paper is a working paper, and its own cover note says working papers "have not been peer-reviewed."

What Pangram claims about its accuracy

Pangram's FAQ states the detector is over 99% accurate, with a false positive rate of about 1 in 10,000. It says that figure is calculated across public datasets containing tens of millions of documents. Pangram also benchmarks its model against 26 AI systems, and its homepage reports 99.6% accuracy on those AI-generated samples as a group. It retests after every major model release.

Pangram publishes accuracy figures for individual models too:

AI modelPangram's reported accuracyTested
Claude Haiku 5.599.70% (2,365 of 2,372)November 2026
Gemini 4 Argon99.02% (2,318 of 2,341)November 2026
GPT-6.1 Sol98.78% (2,188 of 2,215)November 2026
Grok 4.399.3%October 2026
Claude Opus 599.8%July 2026
GPT-5.699.6%July 2026
Gemini 399.5%July 2026

These are Pangram's own figures, measured on its own test sets. Its model card breaks the number down further. On 520,000 English AI-generated samples from 26 models, Pangram 4.1 missed 0.37% of them, close to the headline figure. The model card also lists a 1.99% miss rate on a multilingual set covering 18 languages, so Pangram detects AI text a little less reliably outside English.

Pangram's newest figures

Pangram's accuracy page says its target false positive rate is 0.01%, or 1 in 10,000, and that it tunes the model to make false negatives before false positives. Its academic writing test has a false positive rate of 0.019% on 62,971 human documents and a false negative rate of 0.000% on 48,443 AI documents. Pangram also reports false negative rates for the newest models on smaller benchmarks:

AI modelPangram's reported false negative rate
GPT-6 Astra0.41%
Claude Fable 5.10.36%
Claude Opus 5.50.30%
Claude Sonnet 5.50.29%

In the Pangram 4.1 model card, the false negative rate by model runs from 0.130% for Claude Sonnet 5 to 0.720% for Llama 3.3 70B and Grok 4.3. Those are Pangram's figures, reported on its own pages in October 2026.

Pangram's 2024 report also compared Pangram with GPTZero, Originality.ai and DetectGPT on its own benchmark, and it left out Claude models because of prompting problems it describes in the report. Rival detectors have been retrained since then, so those 2024 comparisons describe older versions on both sides.

Pangram's false positive rate, by domain and language

Pangram's FAQ gives a single overall false positive rate of 1 in 10,000. Its model cards report more specific numbers. Pangram 4.1 had 70 false positives in 781,981 English documents from the FineWeb dataset, or 0.0090%, about 1 in 11,000. Pangram 4 had 41 in 1,000,000 documents, or 0.0041%, on the earlier version of that set, before Pangram removed spam and machine-made pages. On 1,000,000 documents across 104 languages, Pangram 4.1 had 33 false positives, or 0.0033%.

The rate changes by domain. Pangram's model card lists these false positive rates:

DomainHuman documentsFalse positive rate
Academic writing (English)62,9710.019%
Biomedical research papers65,0530.002%
News (multilingual)100,1990.003%
How-to articles (multilingual)166,1940.016%
AWS documentation11,6520.017%
Creative writing10,4950.000%
Amazon reviews (multilingual)10,4250.000%
Speeches1,0580.000%
Movie scripts9,9890.010%
Poetry12,7690.078%
Recipes22,4210.049%

For academic writing in English, the false positive rate is 0.019%, or about 1 in 5,000 documents. That's higher than Pangram's overall figure, and it's the number that matters most for students and researchers. Formal academic prose follows set patterns, and formulaic text is where Pangram's own data shows more false positives.

Pangram's own blog post on false positives says the rate rises on text that's short, not written in full sentences, or highly formulaic, such as recipes and poetry. It says the rate stays low on long, original prose such as essays, reviews and reports, because that kind of writing gives the model more to work with.

Language changes the rate too. Pangram 4.1 officially supports 24 languages. They are English, Spanish, French, Portuguese, Arabic, Chinese, Japanese, Korean, Russian, Turkish, Hungarian, German, Dutch, Swedish, Romanian, Ukrainian, Polish, Italian, Czech, Greek, Hindi, Persian, Urdu and Vietnamese. On Pangram 4.1's multilingual test, false positive rates by language run from 0.0000% for Chinese, Dutch, Greek, Japanese, Persian, Romanian, Swedish and Urdu up to 0.0157% for Portuguese. All of those are far below 1 in 1,000. Pangram says it can also work on languages outside that list and invites requests to evaluate a new one.

Missed AI text rises outside English. Pangram 4.1's false negative rate is 1.99% across 18 languages, against 0.37% for English. In the Pangram 4 model card, misses were highest for Urdu (5.32%) and Persian (3.17%). Machine translations, including Google Translate output, are out of scope according to Pangram.

Pangram also publishes comparisons with other detectors. On product reviews from the University of Chicago research, it reports false positive rates of 0.5% for itself, 1.7% for Originality.ai and 2.4% for GPTZero. These are Pangram's figures, from its own site.

Short texts

Pangram is designed for prose of 50 words or more, and its false negative rate rises as the text gets shorter. Pangram 4.1's dashboard also shows less certainty for text under 150 words. Pangram's accuracy page gives these figures by length.

LengthFalse positive rateFalse negative rate
50 to 99 words0.02%1.29%
100 to 249 words0.009%0.41%
250 to 499 words0.004%0.21%

Most short-text errors are misses. That is the direction Pangram tunes for, so a short passage that scores human is less certain than a long one.

The Chicago paper tested passages under 50 words, below Pangram's own stated scope. Pangram's false positive rate there was mostly zero and at most 2.5%, in restaurant reviews. The detector still scored every stub, and its AUROC ranged from 0.963 to 1.000. Originality.ai rejected some stubs with a length filter error, so the authors could not score it on that test.

For a researcher, the practical point is that an abstract or a one-paragraph answer is close to Pangram's lower limit. The model card also lists reference sections, tables of contents, code and math-heavy text as outside its main scope or more error-prone. A pasted reference list is a poor test.

Non-native English writers

For writers who learned English as a second language, the older evidence comes from other detectors. Liang et al., writing in the journal Patterns in 2023, found that seven detectors flagged about 61% of essays by non-native English speakers as AI. Pangram's 2024 report tested 91 of the same TOEFL essays and found no false positives, and it found none in 3,907 ELLIPSE essays by 8th to 12th graders. On 5,600 ICNALE essays by undergraduates at Asian universities, the rate was 0.09%. Those are Pangram's tests, and the report states that the TOEFL essays were held out of training.

Pangram's newest test is larger. In October 2026 it scanned 15,420 essays and papers by undergraduate and graduate students who speak English as an additional language. They came from nine research corpora, collected at universities in 16 countries, and the writers had 51 different first languages. Pangram 4.1 flagged none of them, and Pangram 4 also had zero false positives on the same set. The texts ran from 52 to 10,916 words. They include literature reviews by graduate students at Canadian universities and A-graded papers by students at the University of Michigan.

Pangram also reports zero false positives on 25,996 school essays from the PERSUADE corpus, which records the writers' disability status, ethnicity, economic background and English learner status. A separate study of 143,755 documents, including 85,887 by autistic writers, found a false positive rate of 0.0011% for Pangram 4 on 89,903 documents written before 2022. Pangram says that study was independent.

The Chicago corpus has no academic or non-native writing category, so Pangram's own tests are the main evidence for this group. Our post on why AI detectors flag non-native writers explains the mechanism behind the older detectors' failures.

How accurate is Pangram AI detection in independent tests?

In a University of Chicago working paper, Pangram had essentially zero false positives and false negatives on medium and long passages. It was the only one of four detectors to meet a strict 0.5% false positive cap without losing accuracy.

The paper, a 2025 University of Chicago Booth study, found that leading AI detectors flagged more than 90% of unedited AI text. Most of them flagged fewer than half once the text went through a humanizer. Pangram says it was the only one that kept its false positive rate under 0.5% while still detecting AI text.

Brian Jabarian and Alex Imas tested Pangram, Originality.ai, GPTZero and an open-source detector called RoBERTa on 1,992 human passages written before 2020. They matched each one with AI text from GPT-4.1, Claude Opus 4, Claude Sonnet 4 and Gemini 2.0 Flash.

The genres were news, blogs, Amazon reviews, restaurant reviews, novel excerpts and resume passages. Novels and resumes counted as long, news and blogs as medium, and the reviews as short. The paper did not include academic writing.

The results at each detector's best threshold, chosen the same way for all of them, ranged like this across genres and models.

MeasurePangramOriginality.aiGPTZero
False positives on human text0% to 0.75%0% to 2.18%0% to 2.38%
Missed AI text, unedited0% to 2%0% to 4.5%0% to 7%
Missed AI text after StealthGPT0% to 5%2.3% to 20.7%25.5% to 77.3%
Passages under 50 words (AUROC)0.963 to 1.000Not scored0.778 to 1.000

All three detectors made their most false positives on short reviews. Pangram's highest false positive rate was 0.75%, on restaurant reviews. GPTZero's was 2.38% on Amazon reviews, and Originality.ai's was 2.18% on restaurant reviews.

On unedited AI text, the gap between Pangram and the two commercial rivals was small. GPTZero in particular missed very little long AI text, and the paper puts its overall false positive rate at about 0.7% across thresholds. The larger differences appear in short passages and in humanized text.

The paper's cost comparison is a side finding. It calculated the fee per correctly flagged AI passage, and Pangram came out at about US$0.023, against US$0.042 for Originality.ai and US$0.057 for GPTZero. That matters to institutions running thousands of scans and not to a single writer.

What the Chicago paper can't tell you

Three limits apply, and the authors state the first one themselves. The AI models were mid-2025 releases, and the authors write that they do not expect the results to stay constant, since detectors, language models and humanizers keep changing. The detector version was Pangram's API from May 2025, which Pangram says was Pangram 3.

The second limit is the human sample. About 2,000 passages can show that a detector is very good, but they can't separate a false positive rate of 1 in 10,000 from 1 in 1,000. Pangram's own large-scale figures fill that gap, and they come from the vendor.

The third is the text type. A journal article, a thesis chapter or a methods section differs from a news story or a novel excerpt. For that, Pangram's own academic writing figure of 0.019% is the closest guide.

Pangram's call for new independent tests

Pangram says the Chicago study is still its largest third-party evaluation, and that it was run more than a year ago on Pangram 3. On November 6, 2026, Pangram asked researchers for new independent tests. It listed the questions it wants answered, such as how detectors handle the newest AI models, mixed human and AI writing, other languages and technical writing. It offers credit grants with no publishing terms, so researchers can publish their results whether or not they favor Pangram.

Pangram's 4.1 model card also lists a 2026 study by Van Vassalaer and colleagues. In its public subset of 39 fully AI-generated documents, Pangram 4.1 missed none.

The Vrije Universiteit Brussel study

Pangram also cites a June 2026 study from Vrije Universiteit Brussel. It tested Pangram, GPTZero, Turnitin and Copyleaks on 160 academic papers by non-native English speakers, split between human, AI-generated, hybrid and humanized text. Pangram says none of the four tools flagged the human-written papers in that study.

The RAID benchmark

RAID is the benchmark Liam Dugan and colleagues presented at ACL 2024, and Pangram's result comes from the follow-up shared task that used its test set. The shared task overview, posted January 15, 2025, ranks detectors by the share of AI text caught at a fixed 5% false positive rate. Pangram scored 99.3% without adversarial attacks, second to a Leidos entry at 99.4%. With adversarial attacks it scored 97.7%, tied for the top score with Leidos.

Two details change how to read that. First, the 5% false positive rate is a benchmark setting that lets detectors be compared on one scale. A school would not accept flagging 1 in 20 honest students. Second, the result shows how well a detector separates AI from human text at that setting. The false positive rate a user sees at the vendor's default threshold is a different number. The same table lists the Binoculars research baseline at 79.0% and 71.3%.

The RAID result for Pangram supports one narrow claim. Its classifier ranks AI text above human text very reliably, including under attack.

Rewrite the AI-flagged part in your own voice

Our academic humanizer rewrites AI-assisted sections while leaving citations, technical terms and numbers exactly as you wrote them.

Try ProofreaderPro.ai Free

A 2025 study: Pangram against expert human readers

A 2025 paper by Jenna Russell, Marzena Karpinska and Mohit Iyyer was published at ACL 2025. It tested whether people who often use AI writing tools can detect AI-generated text as well as automatic detectors do. Five annotators who said they frequently used LLMs for their own writing read 300 non-fiction English articles. The set mixed human-written text with text from GPT-4o, Claude and o1-Pro, including some articles run through a humanizer. The majority vote of those five annotators misclassified only 1 of the 300 articles.

The paper compared that panel with five automatic detectors: Pangram, GPTZero, Binoculars, Fast-DetectGPT and RADAR. The researchers ran two Pangram models, its base detector and a separate model called Pangram Humanizers that's trained to detect text run through a humanizer. Pangram Humanizers was the only detector to match the expert panel, with an average true positive rate of 99.3% and a false positive rate of 2.7% across the test set. Pangram's base model reached a 98.0% average true positive rate with a 2% false positive rate in the same test.

A 2.7% false positive rate is higher than the 1 in 10,000 figure Pangram reports from its own evaluations. This test used a much smaller set of articles, 300 against the millions Pangram cites elsewhere. The paper also scored Pangram using only its Human and AI-Generated labels, without giving it credit for its neutral "Possibly AI" label. Pangram and GPTZero both gave the researchers API credits to run the test.

The humanized subset is where the detectors separated. For o1-Pro text run through a humanizing prompt, Pangram detected 90.0% of articles with a 6.7% false positive rate, and its humanizer-aware variant detected 96.7% with 10%. GPTZero detected 46.7% in that subset, with a 3.3% false positive rate, though its overall rate was only 0.7%. Each condition had 30 human and 30 AI articles, so one article equals 3.3 percentage points, and the false positive rates are noisy. The study also covered only American English non-fiction under 1,000 words.

Humanized and edited text

On text run through the StealthGPT humanizer, Pangram missed 0% to 5% of AI passages in the Chicago test, and GPTZero missed 25.5% to 77.3%. The authors ran every AI passage through StealthGPT's default rewrite endpoint and rescored it. They say GPTZero largely lost its ability to detect AI text after that rewrite, with miss rates around 50% and above in most cases. Originality.ai was partly affected, with misses reaching about 21% on shorter text.

GPTZero's own results on unedited text were strong in the same paper, so the drop applies to one specific kind of input. The paper tested one humanizer, one default setting and mid-2025 versions of every tool. Our post on whether AI humanizers work covers what the wider research says.

Light edits and heavy edits

Pangram's accuracy page gives results for text that started as human writing and was edited with AI.

Type of editDocumentsLabelled humanLabelled mixedLabelled AI
Proofreading or polishing11,36399.45%0.54%0.0088%
Substantial AI edits14,99041.37%55.01%3.62%
WildChat-derived edits4,82627.85%65.17%6.98%

Light proofreading should usually keep a Human label, according to Pangram, while suggested wording changes generally earn "AI-Assisted" and large rewrites can earn "AI-Generated". The same page says the model errs toward Human when a text is mixed. The 41.37% in the middle row also shows that substantial AI edits are often missed, which is a false negative by Pangram's own definition.

Does Pangram have a humanizer?

No. Pangram sells detection only, and it doesn't offer a tool that rewrites text. Its detector looks for humanized text instead. For every segment, the Pangram 4.1 output includes a humanizer score and a flag for whether the segment looks humanized. Pangram's homepage says it can detect AI-generated text after it has been run through a humanizer.

The outside evidence for that claim is the two studies above. In the Chicago paper, Pangram missed 0% to 5% of AI passages after StealthGPT rewrote them. In the Maryland study, a Pangram model trained on humanized text matched a panel of expert human readers. Both tests used mid-2025 versions of the tools, and humanizers change as often as detectors do.

How Pangram detects AI writing

Pangram says it doesn't use perplexity or burstiness, the two signals most older AI detectors rely on. Perplexity measures how predictable each word is to a language model, and burstiness measures how much that predictability changes across a document. We explain both in our own post on perplexity in AI detection.

Pangram argues that perplexity-based detectors misfire on any text a language model has memorized during training, including the Declaration of Independence and large parts of Wikipedia. Models learn to predict memorized text with very low perplexity, regardless of who wrote it. Pangram also says perplexity-based methods have a higher false positive rate on non-native English writing, since simpler sentence structures are also more predictable. We cover that same false-positive pattern for non-native English writers in a post that looks at other detectors.

Instead, Pangram trains a deep learning model directly on large sets of human and AI text. It uses a method it calls hard negative mining to find human writing that happens to resemble AI output. Pangram processes full documents at the token level. It labels each segment Human, AI-Assisted or AI-Generated, with a continuous score for how much of the segment looks AI-written. Pangram says the model improves as it's trained on more data.

What a Pangram result means

A Pangram percentage is the share of the document that looks AI-generated or AI-assisted. Pangram's explainer, dated March 17, 2026, says a 30% score on a 10-page paper suggests about three pages show typical AI patterns. The score is built from segment-by-segment predictions across the document.

Pangram uses three labels. Human is human-written text. AI-Assisted is human text edited with AI. AI-Generated is text produced by an AI model. Each segment also gets a confidence level of High, Medium or Low, and the model card says that level is not a calibrated probability. Pangram 4.1 adds class probabilities for each label, which show how sure the model is about each one.

Pangram's explainer gives examples of what makes a document hybrid. They include using Grammarly, asking a language model to smooth a paragraph and translating from a native language into English. It describes scores below 20% as often standard digital writing assistance and scores above 60% as the range that often warrants a conversation about authenticity. It also states that no universal number requires action, and that a reviewer should set a threshold to match the institution's own AI policy.

By Pangram's own description, an AI-Assisted label on a polished draft is an expected output. It describes how the text was edited and says nothing about misconduct.

Pangram vs GPTZero, Copyleaks and Turnitin

Pangram, GPTZero and Copyleaks are open to anyone who signs up. Turnitin only works through a school or publisher account, and an individual test-taker typically sees a Turnitin score only if an instructor passes it along.

DetectorAccessVendor's own false positive claimFree plan
PangramOpen signupAbout 1 in 10,000 overall20 checks a day
GPTZeroOpen signupAbout 1%10,000 characters per scan
CopyleaksOpen signupAbout 0.2%Limited free scans
TurnitinSchools and publishers only0.51% on academic writingNot available to individuals

We've covered each of these detectors separately: see is GPTZero accurate, does Copyleaks detect AI and can Turnitin detect humanized AI text. The false positive figures in the table are each company's own claims, as Pangram and the other vendors state them.

How Pangram compared in the same studies

Pangram came out ahead in the Chicago paper on short text and on humanized text, and the two commercial rivals were close to it on long unedited text. In the Russell paper Pangram led on the humanized subset. In the RAID shared task it was second and then tied for first. Winston AI and Sapling weren't in those studies, and our posts on Winston AI detector accuracy and Sapling's AI detector cover them.

All of these detectors have been evaluated mostly on English prose, and none of them can prove who wrote a specific document. Pangram's explainer says the same about its own score, and that is the right standard to hold every detector to.

Pangram vs Grammarly's AI detector

Grammarly's AI detector is a free scanner inside Grammarly's writing suite, and Grammarly Pro, at US$12 a month, adds an AI Detector agent that marks passages and offers a rewrite. Grammarly says its detector ranks first for quality on the RAID benchmark. Pangram is a standalone detector with a Similarity Checker, LMS integrations and an API, so it suits teachers and institutions that scan many documents. Grammarly's detector suits a writer who already uses Grammarly for grammar and wants a quick check in the same place.

Pangram is a standalone detector. Grammarly and Scribbr offer AI detection as part of their writing tools, and we've looked at Grammarly's AI detector and Scribbr's AI detector separately.

Pricing and free plan limits

Pangram's free plan gives you 20 checks a day after you sign up, including detection of AI-assisted text, with no payment method. Plagiarism checks and the higher limits need a paid plan. These were Pangram's prices on November 8, 2026:

PlanBilled monthlyBilled yearlyWhat it adds
FreeUS$0US$020 checks a day
IndividualUS$20 a monthUS$15 a month (US$180 a year)300,000 words, 600 images or 50 minutes of video a month, a similarity check on every scan, file upload with OCR, Google Docs and Gmail, US$10 of API credits a month
ProfessionalUS$65 a monthUS$45 a month (US$540 a year)5 times the scans of Individual, plus US$200 of API credits a month
TeamUS$40 a month for 2 seatsUS$30 a month for 2 seats (US$360 a year)Individual features for each seat, an admin dashboard and one bill

The Individual and Team plans start with a 7-day free trial, and Professional is billed when you buy it. Developers can also pay as they go through the API. Pangram 4.1 text costs US$0.05 per 100 words, images cost US$0.25 each and video costs US$0.05 a second, with 20% off bulk text. Schools, publishers and other organizations have separate pricing under the Organizational tab on Pangram's pricing page.

Who uses Pangram

Schools and universities make up Pangram's largest customer base. Its LMS integrations connect directly to Canvas, Moodle, Google Classroom and Brightspace. A teacher can see a Pangram result next to a student's submission without leaving the platform they already use to grade.

Publishers, law firms and HR teams use Pangram for different reasons: checking submitted manuscripts, reviewing documents in a case, or screening applicant materials. The API lets any of these organizations build Pangram's detection into their own internal tools. The Gmail and Google Docs integrations let one person check text without leaving their inbox or a document.

For a student, this usually means a Pangram score appears inside the same dashboard your teacher already uses for grades and feedback. It isn't a separate report you request yourself. If your school runs Pangram through an LMS, ask your instructor or department what score counts as a concern and what happens next.

If Pangram flags your own work

Collect your drafting history first, because a record of your own drafting process counts for more than arguing about the score alone. Version history from Google Docs or Word, earlier drafts, notes, a reference manager library and tracked changes all show how the text developed. Pangram's own explainer tells reviewers to ask the writer to explain their drafting process. Our guide on process evidence as proof lists what to keep.

Then check which label you got. An AI-Assisted result on writing you edited with a grammar tool matches what Pangram says the label means. If your institution allows that kind of help, say so. Our post on AI-use disclosure statements for manuscripts has wording for the common cases.

If the result is AI-Generated on text you wrote yourself, ask for the full report with the highlighted segments and the confidence level, not only the percentage. Check whether the flagged parts are short, formulaic or made of references and templated text, since those are the weak spots in Pangram's own scope notes. Our appeal playbook has the steps for a formal request.

If AI helped draft part of your paper, our academic humanizer rewrites that section in your own voice. Citations, technical terms and numbers stay exactly as you wrote them, and the tool is built to bring AI scores down on the sections it touches.

Frequently asked questions

Q: Is Pangram accurate?

Pangram says it's over 99% accurate, and it cites outside studies with similar results. In the University of Chicago working paper, which tested an earlier Pangram model, its false positive rate was 0% to 0.75% across genres on medium and long English passages. Its own model card shows lower accuracy on texts under 50 words and on some languages other than English.

Q: What is Pangram 4.1?

Pangram 4.1 is Pangram's current text detection model, released on October 19, 2026. It replaced Pangram 4, gives more stable results after small edits and adds class probabilities to each segment. Pangram reports a 0.0090% false positive rate for it on 781,981 English documents.

Q: What is Pangram's false positive rate?

Pangram's FAQ states an overall rate of about 1 in 10,000. Its Pangram 4.1 model card reports 0.0090%, or about 1 in 11,000, measured on 781,981 English documents, with the rate varying by domain and language. The Chicago test measured 0% to 0.75% by genre on about 2,000 human passages, which is too small a sample to confirm a rate as low as 1 in 10,000.

Q: How does Pangram compare to GPTZero?

Pangram says its false positive rate is about 1 in 10,000, and GPTZero says about 1% on the RAID benchmark. Both have free plans and open signup. We cover GPTZero's own numbers in our GPTZero accuracy review.

Q: How does Pangram compare to Turnitin?

Turnitin only works through a school or publisher account, while anyone can sign up for Pangram directly. Pangram cites a 0.51% false positive rate on academic writing from Turnitin's own whitepaper and reports a lower rate for itself.

Q: How does Pangram compare to Copyleaks?

Copyleaks says its false positive rate is about 0.2%, and Pangram reports a lower rate for itself. Both offer free scans, and we cover Copyleaks in does Copyleaks detect AI.

Q: How does Pangram detect AI writing?

Pangram uses a deep learning model trained on human and AI text. It doesn't use perplexity and burstiness, the measures many older detectors rely on. It labels each segment of a document Human, AI-Assisted or AI-Generated and gives a confidence level for the result.

Q: Is Pangram's AI detector free to use?

Yes. The free plan gives you 20 checks a day after you sign up, with no payment method. Paid plans start at US$20 a month, or US$15 a month billed yearly, and add 300,000 words a month and plagiarism checks.

Q: Does Pangram have a humanizer?

No. Pangram only detects AI text. Its model scores each segment for signs of humanizer use, and Pangram says it detects AI text after it has been humanized.

Q: Does Pangram check for plagiarism?

Yes, on paid plans. Pangram's Similarity Checker runs in the same scan as the AI detector and compares your text with billions of webpages, academic papers, news articles and books.

Q: Does Pangram detect text run through a humanizer?

Pangram says it does, and it points to the 2025 Chicago Booth study as support. In that test, it missed 0% to 5% of AI passages after StealthGPT rewriting, against 25.5% to 77.3% for GPTZero and 2.3% to 20.7% for Originality.ai. Detectors and humanizers both update often, so results on humanized text can change over time.

Q: What did the University of Maryland study find about Pangram?

A 2025 University of Maryland paper found that Pangram's humanizer-aware model, Pangram Humanizers, was the only automatic detector to match a panel of five expert human readers. Both reached a 99.3% average true positive rate across 300 test articles. Pangram's false positive rate in that specific test was 2.7%, well above the 1 in 10,000 rate it reports from its own larger evaluations.

Q: Who makes Pangram?

Pangram Labs is the company behind Pangram, founded in Brooklyn, New York in 2023 by former AI researchers from Tesla and Google. The company also sells a Similarity Checker, an AI image and video detector and API access alongside its text detector.

Q: Does Pangram work on short text?

Yes, with higher miss rates. Pangram is designed for prose of 50 words or more, and its false negative rate is 1.29% at 50 to 99 words and 0.21% at 250 to 499 words. The Chicago test scored passages under 50 words and found a false positive rate of at most 2.5%.

Q: Is Pangram biased against non-native English writers?

Pangram's own tests found no bias. Pangram 4.1 had zero false positives on 15,420 essays and papers by university students who speak English as an additional language. Pangram's 2024 report found none on 91 TOEFL essays. These are the vendor's tests.

Q: What does an AI-Assisted label mean in Pangram?

It means human-written text that has been edited with AI. Pangram's model card defines it that way, and its explainer lists Grammarly, asking a model to smooth a paragraph and translating into English as examples. It describes how the text was edited and doesn't say the writer generated it.

Q: Does Pangram work in languages other than English?

It supports 24 languages, and Pangram 4.1's false positive rates by language run from 0.0000% to 0.0157%. It misses more AI text outside English, with a false negative rate of 1.99% across 18 languages. Machine-translated text is out of scope.

Q: Should one Pangram score decide a grade or a decision?

No. Pangram's explainer says no universal number requires action, and it describes the score as a diagnostic tool. A reviewer should look at the highlighted segments, the confidence level and the writer's drafting history.

Q: What should I do if Pangram flags my own writing?

Gather your drafts, version history and notes, then request the full report and ask how the institution treats the AI-Assisted label. Our appeal guide covers the steps, and the process evidence guide covers what to keep.

Academic AI Humanizer

Rewrites AI-assisted academic drafts in your own voice, leaving citations, technical terms and numbers exactly as you wrote them.

Share onXLinkedInFacebook
Dana - Author at ProofreaderPro.ai
DanaContent Creator

Dana is a content creator at ProofreaderPro, where she runs the daily blog and writing operations. She writes the articles on how the online editing platform works, and she handles customer messages every day, with a five-star satisfaction score to show for it.

More Articles

Keep Reading

Try Text Humanizer Free

Join researchers from 200+ universities worldwide. Free to start, no credit card required.

Get Started Free
Proofreader Pro AI
Refine your research with ProofreaderPro.ai, the world's leading AI-powered proofreader, tailored for academic text.
ProofreaderProAI, Greenleaf Ave, Staten Island, 10310 New York
© 2026 ProofreaderPro.ai. A leading academic proofreader, editor & humanizer. Made with ❤️ and linguistically sound syntax 🌳s