Sign in or create an account

What Is the Pangram AI Detector?

By Ethen Walker · Editorial policy · Updated

A page of text divided into passages, each tagged with a small marker for human, assisted or generated.

Pangram is an AI-text detector made by Pangram Labs, a company founded in Brooklyn in 2023. It splits a document into passages and labels each one human-written, AI-assisted or AI-generated. Its own documentation and one independent study report very low error rates on long English text, and its own documentation also lists the kinds of writing where it is less dependable.

Disclosure: AI-Detection.io runs on the Pangram Labs API. We are a customer of the product we are describing, so read this page as an explanation from a user of the tool, not an independent review. We have not run our own accuracy testing, and we do not claim Pangram is the most accurate detector.

Key takeaways

  • Pangram classifies passages, not whole documents, and it returns labels with a High, Medium or Low confidence value, not reasons.
  • Most accuracy figures come from Pangram itself. One independent working paper, from University of Chicago researchers, found Pangram performed best of the four detectors it compared.
  • Pangram says it works best on long-form writing in complete sentences, and that bullet lists, tables of contents, reference sections, templated text and dense math are more prone to false positives.
  • A Pangram label does not show who wrote something. Treat it as a pointer to passages worth reading closely.

What is Pangram Labs, and what is its AI detector?

Pangram Labs sells a classifier that estimates whether text was written by a person or a large language model. Its homepage describes the company as founded in Brooklyn, NY, in 2023 by AI researchers from Tesla and Google. The founders, Bradley Emi and Max Spero, are the authors of the company’s technical report on the classifier. For how detectors of this kind work in general, see how AI detection works.

Pangram offers a web tool and an API, plus a separate image detector and similarity checker that this page does not cover. The text detector is a neural network trained on paired examples of human and machine writing. The technical report describes a training method it calls hard negative mining with synthetic mirrors, which is meant to cut false positives on text that is hard to tell apart.

What does Pangram actually label?

Pangram does not return one yes-or-no verdict. It reads the text in chunks and assigns each a label, then reports how much of the document falls into each class.

On the homepage, the labels are human, AI and AI-assisted. In the Pangram 3.2 model card, the segment labels are Human-Written, Lightly AI-Assisted, Moderately AI-Assisted and AI-Generated. Each segment also gets a score from 0.0 (no AI assistance) to 1.0 (fully AI-generated) and a confidence level of High, Medium or Low.

Two details from the model card matter when you read a result:

  • The resolution is coarse. The model card says resolution is about 50 words, so it cannot separate human and AI text at the sentence or word level. A highlighted sentence is part of a flagged passage, not a sentence-level finding.
  • Assistance is a spectrum. In its announcement of AI assistance detection (11 December 2025), Pangram describes light assistance as surface-level changes such as spelling, grammar, rephrasing and translation, and moderate assistance as changes where AI may have rewritten significant portions. It adds that “co-authorship is a spectrum” and says it cannot infer where an idea came from. The line between the two levels is not exact. For the wider picture, see AI-assisted vs AI-generated writing.

What does Pangram’s own research report?

Pangram publishes two kinds of material: a technical report on arXiv and model cards for each release. These are the vendor’s claims, so each figure below is attributed and comes with its conditions.

Technical report (February 2024, revised July 2024). Pangram reports that its classifier has over 38 times lower error rates than the leading commercial detectors it compared against, on a benchmark covering 10 text domains and 8 language models. It also says the classifier is not biased against nonnative English speakers and generalizes to domains and models it was not trained on. That is a 2024 comparison against 2024 products, so it does not describe today’s competitors.

Model card for Pangram 3.2 (27 February 2026). Pangram reports false positive rates on its human-text datasets from 0.00% to 0.54%. The highest were 0.54% for poetry and 0.35% for news; its academic writing set was 0.02%. For AI text, it reports false negative rates from 0.00% to 1.98%, with the highest on a random sample of chatbot conversations. It also reports results on outside datasets, including 0% false positives on 91 TOEFL essays in the sample used by Liang and colleagues and 0.07% false positives on a random sample of the RAID benchmark. The minimum text length is 50 words, lowered from 75 in the previous version.

Marketing figures. The homepage advertises 99.98% accuracy, while the model card reports 99.7% on its AI-generated samples and the FAQ says “over 99%”. The page does not explain the difference. We do not rely on any of these headline numbers, and we suggest you read the per-dataset tables instead.

The model card page also lists newer versions, up to Pangram 4.1 when we checked on 2026-10-27, with figures we did not review. We do not know which model version serves any given API request on a given day, so we cannot say that a figure for one version applies to the results you see.

What do independent evaluations show?

The strongest independent evidence we found is a working paper by Brian Jabarian and Alex Imas, “Artificial Writing and Automated Detection”, released in September 2025 as an NBER working paper and summarized in Chicago Booth Review.

According to those sources:

  • The researchers compared three commercial detectors (Pangram, GPTZero and Originality.ai) and one open-source model (RoBERTa).
  • They used about 2,000 human-written passages from six kinds of writing: blogs, consumer reviews, news articles, novels, restaurant reviews and résumés. Four language models generated the AI versions. Passages ranged from about 1,000 words down to under 50.
  • Pangram had essentially no false positives at most thresholds and false negatives of 2 to 4%. The authors describe it as the only detector that stayed within a strict policy cap of 0.5% false positives while still catching AI text.
  • GPTZero’s false positives stayed below 1% with low false negatives. Originality.ai’s false negatives ran from 10% to 40% depending on the model. RoBERTa was often close to guessing.
  • All three commercial tools lost accuracy on passages under 50 words.
  • The authors declare no financial or personal conflicts of interest. Their funding included the Booth Center for Applied Artificial Intelligence and the Becker Friedman Institute.

The paper also warns that performance will change as detectors, language models and rewriting tools compete, and it recommends regular published audits. Pangram’s homepage cites this study, and some third-party write-ups present it as “independent verification”. That is a fair description of one study, but it is one study, covering the kinds of writing it sampled. It does not show that Pangram performs the same way on every genre, language or future model.

We read the NBER abstract page and the Chicago Booth Review summary. We did not read the full paper text, so we do not quote more of its numbers than those two pages give.

How does the Pangram AI detector compare with GPTZero and Copyleaks?

On the evidence we found, there is one independent comparison that includes Pangram: the NBER working paper above, which tested Pangram, GPTZero and Originality.ai on the same passages. It found Pangram and GPTZero close on false positives and Pangram stronger on false negatives, and Originality.ai weaker at catching AI text. Copyleaks was not in that study, so we cannot compare it with Pangram on independent data. Vendor figures are measured on each vendor’s own tests and are not directly comparable.

For each tool’s own claims and the outside evidence, see our guides on GPTZero, Copyleaks and Originality.ai, and the side-by-side overview in which AI detector teachers use.

Where does Pangram say it is less reliable?

For a shorter answer to whether the results can be trusted, see is the Pangram AI detector reliable.

Pangram is unusually direct here. The 3.2 model card says the model “is intended to be used on long-form writing samples in complete sentences” and that bullet lists, instructions and technical manuals, tables of contents, reference sections, templated or automated writing and dense mathematical equations are more susceptible to false positives than other domains. It recommends removing headers, footers and other extra formatting, and says raw text or .docx files are better inputs than PDFs because of possible parser errors.

It also warns that short human passages sitting between AI text may be labeled AI-assisted. That is the same mixed-text problem every detector has, and it is why the middle label is the least settled one. The model card states that false accusations can cause serious harm and that the model has a non-zero error rate.

Pangram’s pages do not document how its detector handles every case. Its claim about nonnative speakers comes from its own report and the outside datasets above, not from the independent paper we read. The general reasons detectors err are covered in can AI detectors be wrong, and the evidence across tools is in how accurate AI detectors are.

What can Pangram not do?

Whatever the error rate, some things are outside what any classifier can do.

  • It cannot show authorship. It reports resemblance to patterns learned in training, not who typed the words or how.
  • It cannot explain a label. It returns the label, a score and a confidence value, not the reason a passage was flagged.
  • It does not compare against the web. Pangram says it analyzes only the submitted content, so it is not a plagiarism check. See AI detector vs plagiarism checker.
  • It cannot settle the in-between cases. Text that a person drafted and a tool polished sits on a spectrum, and the boundary between light and moderate assistance is soft.
  • Its confidence value is not a probability. High confidence means the model is more sure about a label, not that there is a given chance the label is right.

How does this map to what you see on AI-Detection.io?

Our checker sends your text to the Pangram Labs API and presents what comes back. The relationship between the two is simple:

Pangram returns Our result shows
Passage (window) labels and confidence Highlights, labelled AI-like, Mixed signal or Human-like
AI-Generated passages “AI-like signal” highlight
AI-Assisted passages “Mixed signal” highlight
Share of text in each class “Flagged as AI-like”: the share of AI-like passages, including mixed ones
Human Written passages “Human-like” highlight

We turn the shares into a verdict (Mostly human-like, Mixed signals, Likely AI-assisted, or Not enough text), and we require at least 50 words. The number is a share of passages, not a probability that the text was written by AI. What an AI detection percentage means explains the difference.

Three points about the relationship:

  1. We publish no accuracy figure. We have not tested Pangram ourselves, and the accuracy page says why we will not borrow a number from elsewhere.
  2. Your text is sent to Pangram. Our server does not store checked text, but Pangram receives it for analysis. We make no claim about what Pangram does with it afterward, so read its terms if that matters. Our privacy page describes the data flow, and checking your own text requires a free account.
  3. The model version can change. Vendor figures describe a specific release, and the service behind our checker can be updated without a change on our side.

How should you read a Pangram result?

  1. Check the length. Under a couple of hundred words, treat the result as rough, and under 50 words, as no result.
  2. Check the genre. Lists, outlines, reference sections, templates and equations are where Pangram itself expects more false positives.
  3. Read the highlights in context, not only the headline share.
  4. Weigh a Medium or Low confidence label lightly.
  5. Look for other evidence before acting: drafts, version history, sources, and a conversation with the writer.

For teachers who must act on a result, what to do when an AI detector flags student writing sets out a fair process, and which AI detector do teachers use compares Pangram with other tools. Our methodology page lists what we analyze and what a result does not mean.

If you already have the text, run it through our AI-generated text detector and review the flagged passages rather than relying on the overall score.

Frequently asked questions

Is Pangram the most accurate AI detector?

We cannot say that. One independent working paper found Pangram performed best of the four detectors it compared, on the writing types and lengths it sampled. Other studies use different texts, tools and versions, and detectors change often. Accuracy depends on the kind of text, so no single ranking holds everywhere.

Is the Pangram AI detector free?

Pangram’s homepage says a free tier allows twenty text checks per day after signup. Pricing and limits can change, so check its site. On AI-Detection.io, checking your own text requires a free account, and a single check accepts up to 5,000 words.

Does Pangram detect AI-assisted writing?

Yes, it has a label for it. Pangram describes light assistance as surface-level edits such as spelling and grammar and moderate assistance as significant rewriting. It says co-authorship is a spectrum, so treat an AI-assisted label as an approximate region, not a precise finding.

Does Pangram work in languages other than English?

Pangram’s model card lists 22 supported languages, including Spanish, French, Arabic, Chinese and Japanese, and says others may work. Our methodology page takes a cautious line: detection is most reliable for English, and we do not claim equal performance across languages.

How is Pangram different from a plagiarism checker?

A plagiarism checker compares text with a database of existing sources. Pangram analyzes only the text you submit and estimates whether it resembles machine writing. A text can be original and still flagged, or copied and not flagged.

Sources