The short answer
Pangram is an AI text detector from Pangram Labs. Its own July 2026 benchmark reports a 0.0041% false positive rate, and independent studies of earlier versions also found very few false flags. Plagino has not tested Pangram. Choose Pangram for wide language and LMS coverage, Plagino for academic .docx reports with source matching.
What is Pangram AI?
Plagino makes a competing AI detector, so read this comparison with that in mind. Every Pangram figure below is linked to where it came from and dated, and we say which numbers Pangram measured itself.
Pangram is a commercial AI detection tool that labels writing as human-written, AI-assisted or AI-generated, and highlights which passages fall into each group. People use it through a web dashboard, a browser extension, a Google Docs add-on, an API and LMS integrations such as Canvas and Google Classroom.
Its paid plans also include a Similarity Checker for plagiarism, so one scan can return an AI result and a source-match result. Pangram's home page markets it to schools, universities and businesses.
Who makes Pangram and how is it funded?
Pangram Labs was founded in Brooklyn in 2023 by Max Spero, the chief executive, and Bradley Emi, the chief technology officer. According to Wikipedia's entry on the company), it launched as Checkfor.ai in October 2023 and took the Pangram Labs name in April 2024. Both founders studied computer science at Stanford. Spero previously worked as an AI engineer at Google and Emi as a machine learning engineer at Tesla.
The company announced $3.98 million in pre-seed and seed funding in June 2025, with the seed round led by ScOp Venture Capital. In July 2026 it raised a further $9 million led by Menlo Ventures, bringing total funding to almost $13 million. That round launched Pangram 4, the current detection model.
Pangram the AI detector is not a pangram sentence
The word "pangram" also means a sentence that uses every letter of the alphabet, such as "The quick brown fox jumps over the lazy dog." Most searches for the single word "pangram" are about that word game. This page covers only the detector made by Pangram Labs.
How does Pangram detect AI writing?
Pangram uses a trained classifier, a type of neural network, rather than a fixed rule about word choice. According to Wikipedia's summary of how it works), it learns patterns from a large set of human documents paired with AI text matched on topic, length and tone. Human examples that it wrongly flags are fed back into training.
Pangram's FAQ says it does not rely on perplexity, the word-predictability measure many detectors use, because predictable human writing can trip it. It reads only the submitted text. It does not check the web for matches as part of the AI score, and it does not use writing-process data such as copy-and-paste history.
What changed in Pangram 4?
Pangram 4, released on 29 July 2026, scores small runs of text and then groups them into passages. This lets it mark one paragraph as AI-generated and the next as human. It also sorts AI assistance into degrees, so a draft lightly edited with a grammar tool can be labelled differently from one written by a chatbot.
Reviewers who read the model card say Pangram 4 does not score text under 50 words. It is built for prose, so code, maths, reference lists, tables of contents and templated text fall outside what it is designed to judge. Pangram 3 was supported until 30 September 2026, so results from before October 2026 may come from the older model.
Is Pangram accurate? Claims against independent tests
Pangram has stronger independent evidence than most AI detectors, but the smallest error rates it advertises come from its own tests. The independent studies found very low false positive rates on human writing. They also found real misses on humanized text and on AI text written to imitate a specific author.
What Pangram says about its own accuracy
Pangram's home page says the detector is "over 99% accurate" and gives a false positive rate of about 1 in 10,000. The Pangram 4 launch post goes further. It reports a false positive rate of 0.0041%, about 1 in 24,000 documents, and a false negative rate of 0.3396%, down from 1.99% for Pangram 3. It also says Pangram 4 found AI involvement in 98.83% of output from 13 commercial humanizer tools.
These are internal benchmarks. A review by eesel that read the technical post says the 0.0041% figure was measured on 1,000,000 English web samples. The same review says the headline numbers apply to English, long-form, complete-sentence text.
What independent researchers found
The strongest outside test is a National Bureau of Economic Research working paper by Brian Jabarian and Alex Imas of the University of Chicago, published in September 2025. It reports near-zero false positive and false negative rates for Pangram across genres, AI models and humanizer tools. Pangram was the only detector that met a strict false positive cap of 0.5% without losing accuracy. GPTZero and Originality.ai were also tested.
A University of Maryland study by Jenna Russell, Marzena Karpinska and Mohit Iyyer, presented at ACL 2025, compared detectors with human annotators on 300 English articles. Pangram caught 98.0% of AI articles and flagged 2% of human ones overall. On articles from o1-Pro run through a humanizer, it caught 90.0% and flagged 6.7%. A majority vote of five experienced human readers misclassified only 1 of the 300 articles.
Where does Pangram get it wrong?
Pangram misses more AI text when the writing has been reworked to avoid detection. Wikipedia's summary of the research) lists several 2026 tests. Epoch AI found that Pangram missed 10.1% of AI passages written to imitate named authors. A Reuters Institute test labelled 86% of 588 evasive AI samples as human. A Notre Dame preprint found it missed more than 96% of abstracts rewritten with a humanizer, at that study's threshold.
Pangram's own limitations matter too. Scores are smoothed across neighbouring text, so a paragraph checked alone can score differently from the same paragraph inside a full essay. Users on review sites report the same thing. In 2026, several publishing disputes turned on Pangram scores, and the writers involved denied using AI.
Pangram's accuracy claims, source by source
This table separates what Pangram measured from what others measured. The last column shows the most each test can prove about false positives. If a test finds no false flags in a sample, the 95% upper limit on the true rate is roughly 3 divided by the sample size.
| Claim or finding | Who measured it | Version and sample | What it shows | What it cannot show |
|---|---|---|---|---|
| 0.0041% false positives (about 1 in 24,000) | Pangram Labs | Pangram 4; 1,000,000 English web samples (per eesel) | Very few false flags on English web text | Rate on student essays, non-English writing or short text |
| 0.3396% false negatives; 98.83% of humanized text caught | Pangram Labs | Pangram 4; 13 humanizer tools | Strong catch rate on its own test sets | How it does on humanizers or prompts it has not seen |
| Near-zero errors; only tool under a 0.5% false positive cap | Jabarian and Imas, University of Chicago (NBER, Sept 2025) | Earlier version; six genres, four AI models | Independent support for a low false positive rate | Anything about Pangram 4 |
| 98.0% caught, 2% false flags; 90.0% caught on humanized o1-Pro | Russell, Karpinska and Iyyer, University of Maryland (ACL 2025) | Earlier version; 300 articles | Close to expert human readers | A precise rate: each condition used a small sample |
| 0 false flags on 495 human passages | Epoch AI (2026) | Pangram 3.3.2 | False positive rate likely under about 0.6% | A rate as low as 0.0041% |
| Missed 10.1% of author-imitation AI passages | Epoch AI (2026) | Pangram 3.3.2 | AI text in a named author's style is harder | How often that happens in real coursework |
To confirm a rate as low as 0.0041% with zero errors, a test would need about 73,000 human documents. The independent tests Plagino found were all far smaller, and none tested Pangram 4.
How do Pangram and Plagino compare side by side?
The two tools solve overlapping problems for different buyers. Pangram is a broad detection platform with an API, browser tools and image detection. Plagino is a narrower checker for academic drafts. The table compares what each vendor publishes, as of October 2026.
| Pangram | Plagino | |
|---|---|---|
| Main use | AI detection for schools, publishers, platforms and content teams | Checking an academic draft before you submit it |
| Free use | Free plan with a daily allowance (see pricing below) | 3 scans a month, no card |
| Entry paid plan | Individual: $20 a month, 300,000 words | Pro: $19 a month, 20 scans a day |
| Teams and schools | Team from $40 a month for 2 seats; Institutional by quote | Team: custom pricing |
| How you submit | Paste text; upload PDF, DOCX or RTF, up to 100 files at once | .docx upload in the web app |
| Result | Human, AI-assisted or AI-generated, with segment highlights | PDF with sentence-level AI highlights and matched sources |
| Plagiarism check | Similarity Checker on paid plans | Open-access academic literature via CORE only |
| Languages | 20+ languages listed | Benchmark covered English only |
| Integrations | Chrome and Firefox extensions, Google Docs, API, Canvas, Moodle, Brightspace, Google Classroom | Web app |
| Published accuracy | Internal Pangram 4 benchmark plus outside studies of earlier versions | March 2026 benchmark of 1,000 documents, with intervals |
| In Plagino's benchmark? | No | Yes |
How much does the Pangram AI checker cost?
Pangram's pricing page lists the Individual plan at $20 a month. It includes up to 300,000 words a month, 100 image scans, a Similarity Checker on every scan, file upload and a 7-day free trial. The Professional plan costs $65 a month for five times the words and image scans, plus $200 of API credit each month. Team starts at $40 a month for two seats, with extra seats at $20 each.
The free plan is harder to pin down. Pangram's FAQ says it allows twenty free checks a day, while two August 2026 reviews describe it as 2,000 words a day. Plagiarism checks need a paid plan. The Pangram 4 launch post moved billing to units of 100 words and set the API at $0.05 per 100 words. That is cheaper for short scans but up to ten times dearer for long documents.
Pangram AI detection Chrome extension and other integrations
The Pangram AI detection Chrome extension checks text as you browse. Highlight a passage, right-click and choose the check option, and the result appears on the page. A Pangram trial page says the extension can also label posts on X, LinkedIn, Substack, Medium and Reddit. A Firefox version is listed on Mozilla's add-ons site.
For schools, Pangram connects to Canvas, Moodle, Brightspace and Google Classroom. Reviewers who checked the plan pages in August 2026 report that these LMS connections come with Institutional licences rather than the Individual plan. If your school uses one of these platforms, ask whether it already licenses Pangram before you buy your own plan.
What does Pangram do with the text you scan?
Pangram says it never uses submitted text to train its models and does not sell or share user data. Its FAQ says submissions are encrypted and stored on its servers so you can reopen your scan history. Its knowledge hub makes the same no-training statement. The Chrome extension stores scanned data by default for its feed report, and its trial page says you can opt out in settings.
These are the company's own statements. If you are checking an unpublished thesis or manuscript, read the privacy policy before uploading. Plagino's privacy page says it does not use uploaded documents to train AI models, and that deleting a scan removes the stored file and report.
Where Pangram is the better choice
Pangram does several things Plagino does not. For some readers those matter more than a single accuracy figure.
Independent evidence on false positives
Pangram has outside studies from the University of Chicago and the University of Maryland, plus more 2026 tests, all finding few false flags on human writing. Plagino's benchmark was run by Plagino, so Pangram's track record is easier to check from the outside.
Languages, integrations and volume
Pangram lists more than 20 languages, an API, browser extensions, a Google Docs add-on and LMS connections. Its Individual plan covers 300,000 words a month. Plagino's benchmark covered English only, and Plagino has no API or LMS integration.
Where Plagino is the better choice
Plagino cannot say it is more accurate than Pangram, because it has not tested Pangram. What it can say is that its numbers come with the method. In the March 2026 benchmark of 1,000 documents, Plagino caught 97.4% of AI-written documents and wrongly flagged 0.8% of human-written ones. Turnitin, the reference, caught 98.2% and flagged 1.4%.
Plagino is built for academic drafts. You upload a .docx file and get a PDF report with sentence-level AI-writing highlights and matched academic sources. You can keep that report and discuss it with a tutor. The free plan gives 3 scans a month with no card, and Pro costs $19 a month for 20 scans a day. Plagino checks writing and does not rewrite it.
How should you read a Pangram score?
Read a Pangram score as a reason to look closer, not as proof. Even a very low false positive rate produces some wrong flags when a large number of documents are checked. A single flag also says nothing about what happened while the essay was written.
A worked example: 10,000 human essays
Imagine a university scans 10,000 essays that were all written by students without AI. At Pangram's advertised 1-in-10,000 rate, about 1 essay would be wrongly flagged. At the Pangram 4 rate of 0.0041%, it would be fewer than 1 on average. At the 2% rate Pangram showed in the Maryland study, it would be about 200. The rate you can expect depends on how close your essays are to the test set behind the number.
A checklist before you act on a flag
Run through these five checks before treating a Pangram flag as meaningful.
- 1
Check the length. Text under 50 words is outside what Pangram 4 is built to score.
- 2
Check the format. Code, reference lists, tables and templated writing are out of scope.
- 3
Check the language. Pangram's lowest error rates are for English.
- 4
Scan the whole document. Scan the whole document, not one paragraph, and note which passages were flagged.
- 5
Look for other evidence. Look for other evidence, such as draft history, notes and a conversation with the writer.
If a detector has flagged your own writing, our guide on what to do if you are falsely accused of using AI covers the next steps.
Test conditions and limitations
Plagino's benchmark used 1,000 English documents: 500 AI-written by GPT-4o, Claude and Gemini, some lightly edited, and 500 human-written. It ran in March 2026 at each tool's default settings and scored at document level. Each rate rests on 500 documents, so it separates large gaps but not small ones. Pangram was not in that test.
The Pangram findings on this page come from Pangram's own pages and from published studies, most of which tested versions before Pangram 4. Some 2026 results are taken from summaries rather than the original papers. AI models and detectors change quickly, so check the date on any accuracy figure. Other head-to-head pages: Plagino vs GPTZero, Plagino vs Originality.ai, Plagino vs Copyleaks and Plagino vs ZeroGPT. The best AI detectors guide compares the wider field.
Frequently asked questions
Is Pangram AI free?
Yes, Pangram has a free plan, but it has a daily limit and no plagiarism check. Pangram's FAQ describes the limit as twenty free checks a day, and some reviewers describe it as 2,000 words a day. Paid plans start at $20 a month.
Is Pangram better than GPTZero?
Pangram had a lower false positive rate than GPTZero in the University of Chicago study, which tested both. The two companies publish benchmarks that each favour their own tool, on different data. Plagino's March 2026 benchmark included GPTZero but not Pangram, so compare both on writing like yours.
Is Pangram better than Turnitin?
They are bought differently, so the answer depends on who you are. Turnitin is licensed by institutions, while anyone can sign up for Pangram. Plagino has benchmarked Turnitin but not Pangram, so it cannot rank the two on accuracy.
Can Pangram detect humanized AI text?
Partly. Pangram says Pangram 4 found AI involvement in 98.83% of output from 13 humanizer tools. Independent tests of earlier versions found larger gaps, including a Notre Dame preprint where most humanizer-rewritten abstracts were missed.
Can Pangram be wrong about human writing?
Yes. Its advertised false positive rate is very low, but it is not zero. The University of Maryland study measured 2% false flags on its human articles, and users have reported human essays scored as AI.
What is the minimum word count for Pangram?
Pangram 4 does not score text under 50 words, according to reviewers who read its model card. Longer passages give more stable results. Very short answers and single sentences are outside its intended use.
Does Pangram work in languages other than English?
Yes. Pangram lists more than 20 languages, including Spanish, French, Arabic, Chinese and Urdu. A review of its Pangram 4 data reports that it misses more AI text in some languages than in English.
Does Pangram check for plagiarism?
Yes, on paid plans. Its Similarity Checker runs alongside the AI check, so one scan returns both results. The free plan does not include plagiarism detection.
Does Pangram store or train on my essays?
Pangram says it does not train on submitted text and does not sell or share user data. It does store scans so you can see your history. Read its privacy policy if the document is confidential.
Who owns Pangram?
Pangram is made by Pangram Labs, a private company founded in 2023 by Max Spero and Bradley Emi. Its investors include Menlo Ventures, ScOp Venture Capital and Haystack. Plagino has no affiliation with Pangram Labs.
Do universities use Pangram?
Some do, through its LMS connections for Canvas, Moodle, Brightspace and Google Classroom. Use differs from school to school. Ask your course leader which detector your school uses and how results are reviewed.
Should a Pangram score be used as proof?
No. A score estimates how text was probably produced; it does not show what a student did. Pangram's own pages acknowledge a non-zero error rate, so treat a flag as the start of a conversation.
