Pangram AI detector
Vendor Pangram Labs Official site pangram.com How this card is sourced
Updated · AI Detector Index editors
Pangram, launched in 2023, detects AI text in 24 named languages and AI images and video according to the vendor's home page read on 8 November 2026 (pangram.com). Pangram publishes a list of 26 models benchmarked, the longest in the index, and had no study record behind the card until 9 November 2026, when the Jabarian and Imas 2025 and Van Vlasselaer 2026 records were added. How this card is sourced explains the rules; the full detector index holds the other twenty-one cards.
What Pangram detects
| Field | Value | Source |
|---|---|---|
| Content types detected | Text, image, video | pangram.com |
| Generators covered | OpenAI, Anthropic, Google, Meta, DeepSeek, xAI, Claude Opus 5, GPT 5.6, Gemini 3, Grok 4.3, Claude Sonnet 5, Claude Fable 5, GPT-5.4, GPT-OSS 120B, DeepSeek V4, Llama 3.3, Mistral Medium 3.5, Kimi K2.6, Qwen 3.7 Max, Gemma 4, Nemotron 3 Ultra, GLM 5.2, 26 models benchmarked | pangram.com |
| Languages | English, Arabic, Chinese, Czech, Dutch, French, German, Greek, Hindi, Hungarian, Italian, Japanese, Korean, Persian, Polish, Portuguese, Romanian, Russian, Spanish, Swedish, Turkish, Ukrainian, Urdu, Vietnamese | pangram.com |
| Input limits | Supports PDF, DOCX, RTF files and up to 100 files at a time. Minimum and maximum words per text check not published. Image: JPG, PNG, WebP, min 512x512px, up to 30 MB each. Video: MP4, MOV, WebM, min 720x720px, up to 500MB. API: 5 QPS realtime; bulk 1,000 billable units. | pangram.com |
| Free tier | 20 free credits daily (twenty free checks a day after signing up for the free plan, including AI assistance detection) | pangram.com |
| API | Yes, docs.pangram.com/quickstart | docs.pangram.com/quickstart |
| Integrations | Chrome extension, Firefox extension, Google Docs, Gmail, Google Classroom, Canvas, Moodle, Brightspace (D2L), MCP server | pangram.com |
| Data retention | In our dashboard and browser extension, we store submitted text in order to show users results and their history of queries. This content is stored for as long as your account is active. It will be available in the history tab of the dashboard and can be deleted from your history to delete the content from our servers. Pangram does not train generalized AI models like ChatGPT, and our AI detection technology is based off of a large, proprietary dataset that doesn't include user submitted content. We never share or sell your data to any third parties. Zero data retention options for enterprise API. | pangram.com/data-privacy |
| Launch year | 2023 | pangram.com/about-us |
| Ownership | Pangram Labs (independent; co-founders Max Spero, CEO, and Bradley Emi, CTO; founded in Brooklyn, NY, in 2023). No parent company stated. | pangram.com/about-us |
| Independent studies | 4 records | Studies index |
Pangram covers three modalities: text, image and video. The home page names OpenAI, Anthropic, Google, Meta, DeepSeek and xAI as providers and lists 26 models benchmarked, among them Claude Opus 5, GPT 5.6, Gemini 3, Grok 4.3, Claude Fable 5, GPT-5.4, DeepSeek V4, Llama 3.3, Mistral Medium 3.5, Kimi K2.6, Qwen 3.7 Max and Gemma 4 (pangram.com, read 8 November 2026).
The FAQ names 24 languages: English, Arabic, Chinese, Czech, Dutch, French, German, Greek, Hindi, Hungarian, Italian, Japanese, Korean, Persian, Polish, Portuguese, Romanian, Russian, Spanish, Swedish, Turkish, Ukrainian, Urdu and Vietnamese, with the list ending "and more"; the page headline says 20+ (pangram.com, read 8 November 2026). The index records the named count of 24.
On humanized text, Pangram Labs says: "Pangram can detect AI-generated text even after it has been 'humanized,' or processed by tools that attempt to evade AI detection", and that an AI Assistance view highlights segments edited by AI rather than fully generated. Source: pangram.com, read 8 November 2026.
Pangram Labs, founded in Brooklyn in 2023 by Max Spero and Bradley Emi, owns the product (about page, read 8 November 2026). Integrations are Chrome and Firefox extensions, Google Docs, Gmail, Google Classroom, Canvas, Moodle, Brightspace (D2L) and an MCP server, plus an API with a quickstart (pangram.com, read 8 November 2026).
Limits and plans
| Plan | Price | Limits | Checked on |
|---|---|---|---|
| Individual | $20 / month | Scan up to 300,000 words per month, 100 image detection scans monthly, similarity detection, Pangram MCP, Google Docs and Gmail integration, $10 in monthly API credits. | |
| Professional | $65 / month | 5x more words, 5x more image scans, $200 of API credits monthly. | |
| Team | $40 / month, 2 seats | $20 / seat per month. | |
| API | Text (Pangram 4.1): $0.05 per 100 words; Image 1.0: $0.25 per image; Image Flash 1.0: $0.05 per image; Video: $0.05 per second | Bulk text API: 20% discount. |
Text uploads accept PDF, DOCX and RTF, up to 100 files at a time. Images must be JPG, PNG or WebP, at least 512x512 pixels and up to 30 MB each; video must be MP4, MOV or WebM, at least 720x720 pixels and up to 500MB. The API runs at 5 QPS in realtime, with bulk jobs of 1,000 billable units (pangram.com, read 8 November 2026).
The free plan grants 20 free credits daily, described as twenty checks a day after sign-up, including AI assistance detection (pangram.com, read 8 November 2026).
Prices below come from the Pangram pricing page, read 8 November 2026. The page shows a dollar symbol without naming the currency.
On retention, Pangram Labs says submitted text in the dashboard and browser extension "is stored for as long as your account is active" and can be deleted from history, and that zero data retention is available for the enterprise API (data privacy page, read 8 November 2026).
What the vendor claims
Pangram states accuracy of over 99% and third-party verification by researchers at the University of Chicago and The University of Maryland. We're accurate on 99.6% of these AI-generated samples.
Pangram Labs states that Pangram is accurate "over 99%" of the time and "has been third-party verified by researchers at the University of Chicago and The University of Maryland", adding: "We're accurate on 99.6% of these AI-generated samples." The sentences are from the Pangram home page, read 8 November 2026.
On false positives, Pangram Labs says: "Pangram's false positive rate (the rate at which human documents are incorrectly flagged as AI) is currently 1 in 10,000. This number is calculated on an aggregate of public datasets containing tens of millions of written documents." (pangram.com, read 8 November 2026). The same page carries a University of Chicago Booth chart showing 0.5% on product reviews and 0.8% on restaurant reviews. The home page claim gives 1 in 10,000; the Booth chart gives 0.5% and 0.8% on two narrow domains; this card uses the vendor's aggregate figure and marks the source as the vendor.
Pangram Labs also says: "AI detection is not magic" (pangram.com, read 8 November 2026). A 1 in 10,000 rate measured on public datasets says nothing about the odds for one specific essay in front of a teacher.
What independent studies found
- AIDI-S-0014Jabarian and Imas 20252025
False positive rate 0.0000 at a 0.5 threshold on 1,992 human passages and at most 0.0075 per genre at Youden-optimised thresholds; false negative rate at most 0.0382 on 1,992 AI passages per model and 0.0000 to 0.0500 after StealthGPT rewrites (Tables 3, 6, 4 and B.4).
- AIDI-S-0015Van Vlasselaer 20262026
Strict accuracy 65% and inclusive accuracy 97.5% on 40 fully AI-generated papers from ChatGPT Deep Research mode, 37 of 40 hybrid and 37 of 40 humanised papers within their known AI range, and all 40 human papers classified correctly; the authors then ran it alone on 1,163 theses and it flagged 45.5% (Tables 4 to 6).
- AIDI-S-0025DAMAGE 20252025
The vendor's own research models: DAMAGE held a 100% true positive rate on AI student essays and 98.26% after humanizing at a fixed 5% false positive rate (97.47% at the default threshold, 3.40% false positive rate); the LLM Baseline without humanized training data fell to 95.74% and 73.07% (Tables 3 and 7). The paper does not state whether DAMAGE is the shipping product.
- AIDI-S-0026Epoch AI 20262026
Pangram 3.3.2 flagged 0 of 495 pre-2022 human passages (0.00%), missed 0 of 297 AI passages from a plain prompt and 30 of 297 (10.10%) written to imitate a named author's style, at its default verdict with Mixed counted as a miss (Epoch AI results table, June to July 2026).
- AIDI-S-0032GPTZero and Pangram vendor benchmarks2026, vendor benchmark
Vendor model card (Pangram's own, Pangram 4.1, 19 October 2026): 70 false positives in 781,981 pre-ChatGPT English FineWeb texts (0.0090%, 95% CI 0.0071% to 0.0113%) and 1,899 misses in 520,000 AI generations from 26 models (0.3652%), Mixed counted as an error; the Pangram 4 card of 29 July 2026 printed 41 of 1,000,000 (0.0041%) on the uncleaned set. Measured by Pangram Labs on Pangram-chosen sets.
No study in the index had tested Pangram before 9 November 2026; two records were added that day, and the detectors with a study record are listed in the independently tested ranking. The Booth chart on Pangram's home page is a vendor-presented third-party result, not a record in this index.
Two records were added on 9 November 2026. Jabarian and Imas 2025 ran Pangram's API on 1,992 human passages written before 2020 and 7,968 AI passages from four models and reports a false positive rate of 0.0000 at a 0.5 threshold, at most 0.0075 per genre at Youden-optimised thresholds, false negative rates of at most 0.0382 per model, and 0.0000 to 0.0500 after StealthGPT rewrites (NBER w34223, read 9 November 2026). Van Vlasselaer 2026 reports strict accuracy of 65% and inclusive accuracy of 97.5% on 40 fully AI-generated papers from ChatGPT Deep Research mode, 37 of 40 hybrid and 37 of 40 humanised papers within their known AI range, and all 40 human papers classified correctly; the authors then ran Pangram alone on 1,163 master's theses without ground truth and it flagged 45.5% (DOI 10.1007/s40979-026-00226-w, read 9 November 2026).
Two records joined the list on 9 November 2026 and took the count to 4: DAMAGE 2025, the vendor's own paper on its research models, which prints 98.26% true positive rate on humanized essays at a fixed 5% false positive rate for a model the paper does not identify as the shipping product, so no field takes a value from it; and Epoch AI 2026, which recorded 0 of 495 pre-2022 human passages flagged, 0 of 297 plain-prompt AI passages missed and 30 of 297 style-imitation passages missed, so the false negative field moved from 1.31% (Jabarian and Imas 2025) to 0% while the false positive field keeps the Jabarian and Imas 2025 zero on its larger base of 1,992. The Pangram 4.1 model card's 70 of 781,981 (0.0090%) is recorded as a dated vendor measurement in the vendor benchmarks record and feeds no rank field.
Where Pangram appears in the index
Rankings
- AIDI-R-01AI detectors ranked by independent testing Two records added in the fifth wave on 9 November 2026, Pegoraro 2023 and Fleckenstein 2024, changed nine counts and none: Pegoraro 2023 tested ZeroGPT, GPTZero, Copyleaks, Originality.ai, Writer, Content at Scale (BrandWell), DetectGPT, the GPT-2 Output Detector and GLTR, lifting GPTZero to 12, the GPT-2 Output Detector to 10, Copyleaks to 7 and Originality.ai and ZeroGPT to 6; Fleckenstein 2024 tested 289 teachers and no software, so it counts for no card ([Fleckenstein 2024 record](/studies/fleckenstein-2024-teachers/)). Illuminarty, one of four cards added in the same wave, joined with its one image record from Ha and Passananti 2024, so 18 of 72 cards now qualify.
- AIDI-R-02AI detector false positive rates, ranked
- AIDI-R-04AI detectors ranked by languages supported
- AIDI-R-07AI image, voice, video and code detectors ranked by independent studies
- AIDI-R-08AI detectors ranked by published robustness to paraphrased and edited text
- AIDI-R-09AI detector false negative rates, ranked
- AIDI-R-10AI detector accuracy claims vs published results, ranked by the gap
- AIDI-R-11AI detectors ranked by named integration surfaces
- AIDI-R-12AI detectors ranked by the upload formats the vendor states
- AIDI-R-14AI detector APIs ranked by published price per 1,000 words
- AIDI-R-15AI detectors ranked by the LMS platforms the vendor names
- AIDI-R-16AI detectors ranked by the free checks per day the vendor states
- AIDI-R-17AI detectors ranked by the generator models the vendor names
- AIDI-R-18AI detectors ranked by study records on non-English and non-native text
- AIDI-R-20AI detectors ranked by launch year, with every change of hands and withdrawal the cards record
Fits these use cases, questions and comparisons
- AIDI-U-01AI detectors for business and managers
- AIDI-U-02AI detectors for school districts
- AIDI-U-03AI detectors for copywriters and agencies
- AIDI-U-08AI detectors for researchers checking their own manuscripts
- AIDI-U-09AI detectors for law firms and courts: an index fit check
- AIDI-U-10AI detectors for grant and proposal reviewers
- AIDI-U-12AI detectors for localisation and translation teams
- AIDI-Q-01Is there a 100% accurate AI detector? What published tests show
- AIDI-Q-02Is there a legit AI detector? How to tell from the record
- AIDI-C-01Pangram vs GPTZero: the 14 fields and the study results side by side
- AIDI-C-02Pangram vs Turnitin: the 14 fields and the study results side by side
- AIDI-C-04GPTZero vs Pangram vs Originality.ai: the three tools one 2025 study ran on the same 9,960 passages
- AIDI-C-08Pangram vs Originality.ai: the 14 fields and the study results side by side
- AIDI-C-13Pangram vs Copyleaks AI detector: two text detectors with one shared study, on the same 14 fields
- AIDI-UP-03Studies digest, Q4 2026: the nine evaluations added since launch, with one headline figure each
Pangram qualified for two rankings on the morning of 9 November 2026, false positive rate on the vendor figure of 0.01%, flagged as vendor-sourced until the Jabarian and Imas 2025 record replaced it with 0.0000 of 1,992 human passages later that day, and languages on a count of 24 (detectors.json, read 8 November 2026). With a daily rather than monthly free allowance the card is absent from the free tiers ranking.
Since 9 November 2026 Pangram also ranks eighth in independent testing on 2 records, fifth in false negative rate on 0.0131 of 1,992 GPT-4.1 passages, fourth in robustness to paraphrase on 37 of 40 humanised papers in Van Vlasselaer 2026, and sixth in the claim vs evidence gap on "over 99%" claimed against 97.5% inclusive accuracy in the same study.
Pangram fits business through the API and the Team plan, school districts through Canvas, Moodle, Brightspace and Google Classroom, agencies through the API and 100-file uploads, and recruiters through the 20 free checks a day on text (pangram.com, read 8 November 2026). The other image-capable detectors to compare are Winston AI and GPTZero, both with study records. The home page carries the full table.
The card also sits in the API pricing ranking added 9 November 2026 at $0.50 per 1,000 words, from $0.05 per 100 words.
Pangram's AI Assistance highlight places it among the sixteen cards on the paraphrased text question, added 9 November 2026; the same day it joined four new rankings, fourth on LMS coverage with four named platforms, third on free checks per day with 20 free checks a day, fifth on generators covered with 16 named models, and tenth on non-English evidence with the Van Vlasselaer 2026 record.
Pangram vs Copyleaks, added 10 November 2026, reads this card beside the other API-first text detector with a measured zero on human text, on the one record that ran both, Van Vlasselaer 2026 (97.5% inclusive accuracy here against 75% of the same 40 papers scored under 20% there), with both API prices converted to the same unit.
Questions about Pangram AI detector
Is Pangram free?
Partly. The free plan gives 20 free credits daily after sign-up, which the vendor describes as twenty checks a day, per the home page read 8 November 2026. Paid Individual access is $20 / month after a 7-day free trial, per the pricing page read the same day.
What is Pangram's false positive rate?
1 in 10,000 by the vendor's own statement, calculated on an aggregate of public datasets with tens of millions of documents, per the home page read 8 November 2026. The Booth chart on the same page shows 0.5% and 0.8% on two review domains. Jabarian and Imas 2025, indexed 9 November 2026, measured 0.0000 of 1,992 human passages at a 0.5 threshold, and Van Vlasselaer 2026 measured 0 of 40 human papers.
How many languages does Pangram detect?
24 are named in the FAQ, with the list ending "and more", while the headline says 20+, per the home page read 8 November 2026. The index ranks Pangram on the named count of 24.
Sources and revisions
Sources
- pangram.comRead 8 November 2026
- pangram.com/pricingRead 8 November 2026
- pangram.com/about-usRead 8 November 2026
- pangram.com/solutions/apiRead 8 November 2026
- pangram.com/data-privacyRead 8 November 2026
- docs.pangram.com/quickstartRead 8 November 2026
Revision log
- Card created
- Jabarian and imas 2025 and van vlasselaer 2026 records added, 2 study records
- Damage 2025 and epoch ai 2026 records added in the sixth wave, 4 study records
- Paraphrased text question and four seventh-wave rankings linked2026-11-10: pangram vs copyleaks comparison linked (tenth wave)
). Illuminarty, one of four cards added in the same wave, joined with its one image record from Ha and Passananti 2024, so 18 of 72 cards now qualify.: register plate](/img/independently-tested-480.webp)
