Short answer
Checked in October 2026, Pangram claims over 99% accuracy and a 1 in 10,000 false positive rate; 2025 independent studies ranked it at or near the top. It is free for twenty checks a day after sign up. No study proves it is the single most accurate detector, and one study still recorded false positives.
Pangram has become one of the names AI answer engines cite when people ask which AI detector is best. Three questions come up again and again: is Pangram accurate, is it the most accurate AI detector, and is it free. This review answers them using Pangram's own pages and technical papers, which we label as the vendor's claims, and the published studies that name Pangram. We did not run our own tests.
Pangram at a glance
Everything in this table comes from Pangram's own website, checked in October 2026.
| Question | What Pangram says | Source |
|---|---|---|
| What it is | A detector for AI generated text that also offers plagiarism checks and AI image detection; founded in Brooklyn in 2023 by researchers who previously worked at Tesla and Google | Pangram homepage |
| Free? | Yes, partly: "twenty free checks a day after signing up for our free plan". Plagiarism detection and extra credits need a subscription; the Individual plan is listed at $20 a month | Pangram FAQ; Pangram pricing |
| Limit | A minimum of 50 words per scan. The Individual plan covers up to 300,000 words a month; we found no stated per scan cap for the free plan | Pangram blog on minimum length; Pangram pricing |
| What it reports | Whether text looks human, AI generated or AI assisted, plus "how much of the document" it believes is AI generated or AI assisted, with highlighted segments | Pangram homepage and FAQ |
| Languages | "More than 20 languages", including English, Arabic, Chinese, French, German, Hindi, Japanese, Portuguese, Spanish and Urdu | Pangram FAQ |
Pangram also says it accepts PDF, DOCX and RTF uploads and works in a Chrome extension, Google Docs and learning platforms such as Canvas and Google Classroom, according to its homepage.
How Pangram works
Pangram describes its method in a technical report written by its founders, Bradley Emi and Max Spero, first posted in February 2024 and updated since. Everything in this section is the vendor's own account.
- A trained classifier, not a perplexity score. The report describes "a slightly modified transformer-style architecture" trained on a mix of human writing and AI writing. Pangram's homepage adds that it does not rely on perplexity, the measure of how predictable each word is, because simple human writing can look predictable too.
- Synthetic mirrors. For each human example, Pangram prompts language models to write a matching AI version on the same topic, a method the report calls "mirror prompting".
- Hard negative mining. An early model scans a large pool of human writing, and any human text it wrongly flags is added back into training with a new AI mirror. Pangram says this "active learning" loop is how it pushes false positives down.
- Three labels. The homepage says a classifier head sorts text into human, AI or AI assisted, and that the first training stage used about one million publicly licensed human documents plus AI text from GPT-5 and other frontier models.
- Humaniser training. In a separate paper, DAMAGE, Pangram's researchers describe training the detector on text run through humaniser and paraphrasing tools.
The technical report also contains a line worth quoting to anyone using the tool for discipline: "We strongly discourage the use of our classifier as a sole arbiter of academic integrity". Our guide to how AI detection really works explains the wider family of methods Pangram is reacting against.
Vendor claims vs independent evidence
The table puts each headline claim next to what we could verify outside Pangram's own pages. Where Pangram's own staff wrote the study, we say so.
| Claim | Vendor's claim | Independent evidence |
|---|---|---|
| Overall accuracy | "Over 99% accurate" (FAQ); 99% on its own 1,976 document benchmark (technical report) | On 300 English articles, Pangram caught 98.0% of AI articles in its base mode and 99.3% in its humaniser mode (Russell, Karpinska and Iyyer, 2025) |
| False positives | "1 in 10,000", measured on public datasets of "tens of millions" of documents (FAQ) | The same 2025 study recorded false positive rates of 2.0% (base) and 2.7% (humaniser mode) on its human articles. A September 2025 NBER working paper reports near zero false positives (Jabarian and Imas) |
| Benchmark robustness | Robust to unseen models and domains (technical report) | In the RAID shared task at COLING 2025, Pangram caught 99.3% of AI text without attacks and 97.7% with attacks at a 5% false positive rate; its attack score tied for first (Dugan and colleagues, 2025) |
| Humanised text | "Pangram can detect AI-generated text even after it has been humanized" (homepage) | Pangram's humaniser mode reached 99.3% in the 2025 Russell study but faltered slightly on humanised o1-Pro articles. Pangram's own DAMAGE paper reports 98.26% (see conflict note below) |
| Non-native English writers | 0% false positives on the 91 TOEFL essays from Liang and colleagues, held out of training (technical report) | We found no independent replication of this result |
| Newest models | 99.7% of samples from 26 models detected, including tests dated July 2026 (homepage) | None found; all independent studies above predate these models |
The RAID shared task result
RAID is a public benchmark built by academic researchers. Pangram does not have an entry in RAID's main leaderboard submissions folder, but it does have one in the shared task directory on GitHub. Its results file shows 97.7% of AI text caught across all attacks at a 5% false positive rate and 94.6% at a stricter 1% rate. The weakest attacks for Pangram were paraphrasing and homoglyphs, where look-alike characters replace normal letters.
Source: RAID repository, Pangram shared task results file, 2024 entry
The shared task was deliberately easier than the open leaderboard: every domain and model in the test set was also in the training data, and the organisers' paper says Pangram "mined the RAID train set" for its hardest examples before submitting. The entry's metadata file is dated 21 February 2024, so it reflects an older Pangram model. RAID also sets the threshold for you, which is not how the live product decides.
Studies by other researchers
Russell, Karpinska and Iyyer, 2025. This team compared five expert human readers with five detectors on 300 non-fiction articles, including ones generated by GPT-4o, Claude and o1-Pro and some passed through humanisers. They found that only Pangram matched the experts' majority vote; GPTZero detected 85.3% of AI articles. The authors thank Pangram and GPTZero for providing API credits, which is worth knowing even though it is common practice (arXiv 2501.15654).
Jabarian and Imas, 2025. This NBER working paper from the University of Chicago tested Pangram, Originality.ai, GPTZero and an open source RoBERTa model. Its abstract says Pangram achieved "near-zero" false negative and false positive rates across models, short passages and humaniser tools, and was "the only tool to satisfy a strict cap (FPR ≤ 0.005) without sacrificing accuracy" (NBER w34223). It is a working paper, and one author, Alex Imas, is quoted praising Pangram on Pangram's homepage.
Conflict of interest: the DAMAGE study
The DAMAGE paper on humanised text, published at the GenAI Content Detection workshop in January 2025, was written by Elyas Masrour, Bradley Emi and Max Spero of Pangram Labs (ACL Anthology). It audits 19 humaniser and paraphrasing tools and reports that the authors' detector caught 98.26% of humanised academic text at a 5% false positive rate, against 60.04% for GPTZero. Because the authors built the detector they are measuring, treat these figures as vendor evidence, not independent confirmation.
Is Pangram the most accurate AI detector?
On current evidence, it is among the strongest, but "most accurate" cannot be proven. The three outside studies above put Pangram at or near the top of the tools they tested, yet each tested a small set of rivals, on English text, before the newest 2026 models. Other vendors also claim first place on RAID; our Winston AI review covers one of them, and the best AI detectors of 2026 roundup compares the field.
Main limitations
- False positives are rare, not impossible. Pangram's 1 in 10,000 figure is its own measurement. The 2025 Russell study saw a 2.0% false positive rate in base mode on its human articles, a small sample but a real one.
- Short text is weaker. Pangram will not give a verdict under 50 words, and its blog explains that short texts lack the context the model needs.
- The research on detectors in general is cautious. Weber-Wulff and colleagues tested 14 tools in 2023 and concluded they were "neither accurate nor reliable", with obfuscation making things worse (arXiv 2306.15666). Pangram was not among the tools tested.
- Non-native English writers face extra risk. Liang and colleagues found seven detectors labelled 61.22% of 91 TOEFL essays as AI written, and 89 of the 91 (97.80%) were flagged by at least one detector (arXiv 2304.02819). Pangram says it scores 0% on that essay set, but only Pangram has reported that. See our article on AI detection and non-native English writers.
- Much of the evidence comes from Pangram. The technical report, the DAMAGE paper and the model accuracy table are all written by the company.
- Benchmarks age quickly. The RAID shared task entry dates from 2024, and the outside studies were run in 2025. New models and new humanisers appear every few months.
If your own writing is flagged
- Ask which tool and which mode were used. Pangram reports human, AI and AI assisted labels, and a share of the document. Find out exactly what was shown and which segments were highlighted.
- Gather your process evidence. Export version history from Google Docs or Word, and collect notes, outlines and sources. A timeline of edits is stronger than any score.
- Quote the vendor's own caution. Pangram's technical report says it should not be the "sole arbiter of academic integrity" and that all detectors have "a nonzero false positive rate".
- Point to the research. Even Pangram recorded a 2.0% false positive rate in the 2025 Russell study, and the Liang study shows how non-native writing can be misread by detectors.
- Offer to talk through the work. Explaining your argument, reading and choices in person is often the most convincing answer.
If you write your own text and only want smoother rhythm, a light pass with a free humaniser is a polishing step, not a shield; our piece on whether humanised text can still be detected explains why tools like Pangram are trained against humanisers.
Your situation and what to do
| Your situation | What to do |
|---|---|
| You want a quick check on a draft | Sign up for the free plan, which Pangram says gives twenty checks a day, and use at least 50 words |
| You are a teacher with a flagged essay | Treat the result as a reason to talk with the student, as Pangram's own report advises, not as proof |
| The writer is a non-native English speaker | Be extra careful; Pangram's 0% TOEFL result has not been replicated by others |
| You need plagiarism checks too | Budget for a paid plan; Pangram says plagiarism detection needs a subscription |
| Your school uses Turnitin | Do not treat a Pangram result as a prediction of Turnitin's score; see our Turnitin vs GPTZero comparison |
| You want free alternatives | Compare our free AI detectors roundup and the GPTZero guide |
| You are choosing a detector for an institution | Read the independent studies above, then run a pilot on your own past, human written work |
Method note: We checked Pangram's own pages and published studies in October 2026; we did not run our own tests. Benchmark figures come from the public RAID repository and the papers linked above.
Frequently asked questions
Is Pangram accurate?
The evidence says it is one of the more accurate detectors. Pangram claims over 99% accuracy. Independent 2025 studies found it caught 98.0% of AI articles in base mode and its score with attacks tied for first in the RAID shared task, though one study still recorded a 2.0% false positive rate on human articles. Treat any single score as a signal, not proof.
Is Pangram the most accurate AI detector?
No study can settle that. Three outside studies from 2025 placed Pangram at or near the top of the tools they tested, but each compared only a few rivals on English text. Pangram has no entry in RAID's main leaderboard, so there is no like for like comparison with the vendors listed there.
Is Pangram free?
Partly. Pangram's FAQ says the free plan offers twenty checks a day after you sign up, including AI assistance detection. Plagiarism detection and extra credits need a subscription; the Individual plan is listed at $20 a month for up to 300,000 words, according to its pricing page checked in October 2026.
Can Pangram detect humanised AI text?
Pangram says it can, and its own DAMAGE paper reports 98.26% detection of humanised academic text, but its authors work for Pangram. In an outside 2025 study, Pangram's humaniser mode caught 99.3% of AI articles overall, while faltering slightly on humanised o1-Pro articles.
Does Pangram work in languages other than English?
Pangram says it detects AI writing in more than 20 languages, including Arabic, Chinese, French, German, Hindi, Japanese, Spanish and Urdu. We found no independent study testing Pangram outside English, so treat non-English results with extra caution.
