Pangram AI detector: how it works and how accurate it is

What is Pangram, and who makes it?

Pangram is a commercial AI detector from Pangram Labs. The company says it was founded in Brooklyn, NY, in 2023 by AI researchers from Tesla and Google. You paste or upload text, and Pangram splits it into segments and labels each one Human, AI-Assisted or AI-Generated. It says it works in more than 20 languages and also checks images and video.

The current text model is Pangram 4.1, which the model card says was released on October 19, 2026. It follows Pangram 4, described in a technical report posted to arXiv in July 2026. Version numbers matter, because every published study below tested an older model.

How does Pangram detect AI writing?

Pangram says it uses a trained classifier rather than a perplexity score. In Pangram's description of its method, a language model turns the text into embeddings, and a classifier head turns those into a label for each segment. Pangram says it trained a first model on about 1 million human and AI documents. It then used that model to search large datasets for human texts it got wrong, added those hard examples to the training set and retrained, over several cycles. For each human example it also generates a matching AI “mirror” in the same style and tone.

Perplexity-based detectors work differently. Perplexity measures how predictable each word is to a language model, and burstiness is how much that predictability varies across a document. In a 2025 post, Pangram's CTO argues that these measures flag famous human texts that appear often in training data, such as the Declaration of Independence, and have an elevated false-positive rate on nonnative English writing. Epoch AI likewise describes Pangram as a trained neural classifier that does not use perplexity. For the wider picture, see how AI detectors work.

How accurate does Pangram say it is?

Pangram says it is over 99% accurate, with a false-positive rate of 1 in 10,000. These are vendor claims. Pangram's October 2026 accuracy page gives these figures for Pangram 4.1 on its own benchmarks:

• False positives: 0.009% of 781,981 human documents were flagged as AI-generated or AI-assisted.
• False negatives: 0.365% of 520,000 AI documents from 26 models were missed.
• Short text: on 50 to 99 words, the false-positive rate rises to 0.02% (about 1 in 5,000) and the false-negative rate to 1.29%.
• Poetry: the highest false-positive rate of any domain Pangram reports, at 0.078%.

Pangram says it deliberately tunes the model to prefer false negatives, because a false positive “can cause real harm”. It also leans human on mixed text: in its own test of substantial AI edits, 41% of documents still came back as human.

Figures also vary by version. The Pangram 4 report gives a false-positive rate of 0.0041%, roughly 1 in 24,000. The 4.1 model card says Pangram then moved to a cleaner, harder test set and calibrated to roughly 1 in 10,000.

What have independent studies found?

Independent studies mostly put Pangram at or near the top, with clear exceptions. All of them predate Pangram 4.1, and Pangram itself said in November 2026 that its largest third-party evaluation by sample size was over a year old and used Pangram 3. The paper itself lists only an API version from May 2025.

These are different tests, not one score. The Brussels team also ran Pangram on 1,163 master's theses whose true authorship was unknown; it flagged 45.5%, typically at low to moderate levels of AI-associated text.

On independence: the Maryland paper thanks Pangram and GPTZero for API credits, and the Notre Dame paper thanks Pangram for API use. Pangram's July 2026 Pangram 4 report lists the University of Maryland alongside Pangram Labs as an author affiliation. The Chicago authors declare no financial or personal conflicts of interest, and Pangram says its research grants carry no publishing terms once a topic is approved.

Does Pangram flag human writing as AI?

Rarely in published tests, but it can. Pangram's own report says “both false positives and false negatives, although rare, do occur”. Epoch found no false positives in 495 human passages, and the Brussels study found all four tools correctly identified the fully human papers. In the Maryland study, Pangram's base model flagged 2% of the human articles.

Notre Dame found a harder case. Pangram 3.2 flagged none of the abstracts published in 2013 to 2015, but 14.9% of those from 2023 to 2025. The authors stress these are flag rates, not confirmed false positives, because some authors may have used AI. Pangram's report adds that the same text can get a different result in a different context, and that the model does not account for people who have “absorbed elements of LLM style into their writing”.

If you have been flagged, our guide to AI detector false positives explains what evidence to gather.

Is Pangram better than GPTZero or Turnitin?

In most independent comparisons, Pangram did as well as or better than the tools it was tested against. The Chicago study put GPTZero and Originality.ai in a secondary tier, with GPTZero largely losing its ability to detect AI text after humanizer rewriting. The Brussels study found Pangram consistently performed better than GPTZero, Copyleaks and Turnitin, which significantly underestimated AI content. In Epoch's test, Pangram and GPTZero were close: neither flagged human text, and they missed 10.1% and 10.8% of style imitations.

Vendor benchmarks favour their own tools. GPTZero's October 2025 benchmark puts GPTZero at 99.6% accuracy and Pangram at 97.5%, while Pangram's 2024 report claims over 38 times lower error rates than leading commercial tools on its own benchmark. For students, the result that counts comes from the tool your school uses. See does Turnitin detect AI? and which AI detector schools use.

Is Pangram free, and what does it cost?

There is a limited free option. As of November 10, 2026, Pangram's homepage says “We offer twenty free checks a day after signing up for our free plan”, so you need an account. The paid plans on its pricing page on the same date were:

• Individual: $20 a month, with a 7-day free trial. Up to 300,000 words a month and 100 image scans.
• Professional: $65 a month. Five times the words and image scans, plus $200 of API credits a month.
• Team: $40 a month for two seats, at $20 per seat.
• API: text detection at $0.05 per 100 words.

The page also shows savings for annual billing. Prices change, so check it before you pay.

Where is Pangram used?

Pangram offers integrations for classrooms, publishing and the browser. Its integrations page lists Canvas, Blackboard, Brightspace, Google Classroom, Moodle and Schoology, plus a browser extension, and its homepage adds Google Docs. In July 2026, Substack launched a feature that lets readers scan posts to see how much is estimated to be written by hand or with AI assistance, using Pangram. Publishers can run it on drafts and turn scanning off post by post.

What are Pangram's documented limits?

Pangram is designed for “natural language prose of 50 words or more”. It lists these as out of scope, where results are less reliable:

• Code, markup and Markdown
• AI and machine translations, such as Google Translate
• Tables of contents, title pages and reference lists
• Figure text, tabular data and mathematical notation
• Highly formatted text, such as poetry and recipes

It recommends raw text or .docx over PDF, and its 2026 report says predictions are “to some degree a black box”. Independent tests found more weak spots: AI text imitating a specific author, and AI text that rhymes or repeats. Mantzarlis also found that slightly altering the order of words in short passages could reverse a verdict.

Results on “humanized” text are mixed. Pangram says it can detect AI-generated text even after it has been humanized, and the Chicago study found it robust to the humanizer it tried. The Notre Dame study found that more than 96% of humanized AI rewrites went unflagged by Pangram 3.2 and GPTZero. Partly for that reason, its authors say detector scores should not be standalone evidence of misconduct.

How should you read a Pangram result?

As one piece of evidence. The Brussels researchers say detectors “should not be used as sole evidence in high-stakes decision-making”. Mantzarlis told the Reuters Institute that Pangram “at scale is probably a pretty solid tool”, but that he is “extremely worried about it being used in individual cases”. A few practical checks:

1. Check the length and format. Short or highly formatted text, such as poetry, is where Pangram reports more errors.
2. Read the segment labels. Pangram labels each segment, so check which parts were flagged and whether the label is AI-Assisted or AI-Generated.
3. Compare with the writer's process. The Reuters Institute piece suggests comparing earlier work, talking about process and asking for document history.
4. Follow the policy that applies. A course, journal or publisher policy decides what AI use is allowed, not a detector.

If your course or publisher allows some AI help, disclose it the way their policy asks and keep your drafts. Our guide to signs of AI writing covers what readers notice, and the responsible AI use guide covers disclosure.

Common questions

Sources

Checked on November 10, 2026. Pangram and GPTZero pages describe each vendor's own product. Studies are listed by their authors.

• Pangram: AI detector homepage and FAQ
• Pangram: How Pangram detects AI-generated content
• Pangram: Is Pangram accurate? (October 2026)
• Pangram: Pangram 4.1 model card
• Pangram: Why perplexity and burstiness fail to detect AI (March 2025)
• Pangram: A call for new Pangram evaluations (November 2026)
• Pangram: Pricing
• Pangram: Integrations
• Pangram Labs (Emi and Spero): Technical report on the Pangram AI-generated text classifier, arXiv (2024)
• Pangram Labs: Pangram 4 technical report, arXiv (July 2026)
• Jabarian and Imas (University of Chicago): Artificial Writing and Automated Detection, NBER Working Paper 34223 (2025)
• Becker Friedman Institute: Research brief: Artificial Writing and Automated Detection (2025)
• Russell, Karpinska and Iyyer (University of Maryland and Microsoft): People who frequently use ChatGPT for writing tasks are accurate and robust detectors of AI-generated text, ACL 2025
• Van Vlasselaer, Van Droogenbroeck and Spruyt (VUB): Who wrote this? Evaluating the reliability of AI detection tools in higher education, International Journal for Educational Integrity (2026)
• Epoch AI: AI detectors rarely flag human writing, but sometimes miss AI text imitating real authors (July 2026)
• Karr, Khvatskii, Hua and Chawla (University of Notre Dame): Why AI Detection Fails for Academic Integrity, arXiv (2026)
• Indicator (Alexios Mantzarlis): What I learned running an adversarial test on an AI text detector (June 2026)
• Reuters Institute: A human wrote this – believe me (July 2026)
• GPTZero: GPTZero vs Pangram (October 2025)
• Substack: How writers are reacting to Substack's AI transparency tools (July 2026)