GPTZero vs Pangram vs Originality.ai: the three tools one 2025 study ran on the same 9,960 passages
Pangram recorded a false positive rate of 0.0000, Originality.ai 0.0011 and GPTZero 0.0071 on the same 1,992 human passages at a 0.5 threshold in Jabarian and Imas 2025, the only study in the index that ran all three detectors on one corpus (NBER w34223, Table 3, read 9 November 2026). GPTZero carries 11 study records, Originality.ai 5 and Pangram 2, by the index data file at version 2026.11.13, and the three share that one record, with GPTZero and Originality.ai sharing four more. On every other field the three cards are read side by side below, with the vendor's wording and the page it came from.
Updated Fields checked 8 and 9 November 2026
Short answer
On the one shared corpus, Pangram posted the lowest false positive rate (0.0000 of 1,992 at threshold 0.5), Originality.ai the next (0.0011) and GPTZero the highest (0.0071), and after a StealthGPT rewrite Pangram missed 0.0000 to 0.0500 of AI passages, Originality.ai 0.0227 to 0.2073 and GPTZero 0.2552 to 0.7730 (NBER w34223, Tables 3 and B.4, read 9 November 2026). GPTZero has the longest outside record, 11 studies from 9 false accusations in 18 texts (Weber-Wulff 2023) to 0.00% at threshold 0.5 (RAID 2024); Originality.ai states 30+ languages and Pangram names 24; Pangram's Individual plan is $20 per month, Originality.ai's Pro is $12.95 per month billed yearly, and GPTZero's premium plans start at $12.99 per month (pangram.com/pricing, originality.ai/pricing, gptzero.me, checked 8 November 2026). No detector score should be the only evidence in a misconduct or hiring decision.
How do GPTZero, Pangram and Originality.ai compare on the 14 fields?
| Field | GPTZero | Pangram | Originality.ai |
|---|---|---|---|
| 1. Name, vendor, official domain | GPTZero, GPTZero LLC, gptzero.me | Pangram, Pangram Labs, pangram.com | Originality.ai, Originality AI Inc., originality.ai |
| 2. Modalities detected | text, image (AI Vision, specs not published) | text, image, video | text, code |
| 3. Generators covered | Claude, ChatGPT, GPT-3, GPT-4, GPT-5, GPT-6, Gemini, Llama, Deepseek, Grok, AI services based on those models | OpenAI, Anthropic, Google, Meta, DeepSeek, xAI and 26 models benchmarked by name, from Claude Opus 5 to GLM 5.2 | GPT-6 Astra, GPT-6.1 Sol, Claude Fable 5, Claude Opus 5.5, Gemini 3, Gemini 4 Argon, Kimi K3, DeepSeek V4, Grok 4.7, Llama 4 Maverick and Scout, "and more" |
| 4. Languages supported | 5: English, German, Portuguese, French, Spanish | 24 named, "and more" | 30+, none named |
| 5. Input limits | 10,000 characters per scan without an account; 150,000 characters or a batch of 250 files with a free account | PDF, DOCX, RTF, up to 100 files at a time; words per check not published; image 30 MB, video 500 MB; API 5 QPS | 100-word minimum; 1 credit per 100 words; paste, upload or URL; maximum per scan not published |
| 6. Free tier | 10,000 words per month (vendor blog, 25 October 2026); 10,000 characters per scan without login | 20 free credits daily (twenty checks a day on the free plan) | 10K words (100 credits) one time, no subscription required |
| 7. Paid plans | Premium plans starting at $12.99/month (vendor blog); Professional Team Plan 500,000 words per member per month, min 2 seats, price not rendered | Individual $20 / month; Professional $65 / month; Team $40 / month, 2 seats; API $0.05 per 100 words | Pro $12.95/month billed yearly or $14.95 monthly, 200K words a month; Enterprise $136.58/month billed yearly or $179 monthly, 1.5M words |
| 8. API | Yes, docs.gptzero.me | Yes, docs.pangram.com | Yes, docs.originality.ai (Enterprise plan) |
| 9. Integrations | Chrome Extension, Google Docs, Google Classroom, Canvas, Zapier, API | Chrome and Firefox extensions, Google Docs, Gmail, Google Classroom, Canvas, Moodle, Brightspace (D2L), MCP server | Chrome and Firefox extensions, Google Classroom (via Chrome), Moodle Plugin 2.0, API, MCP, Bulk Scan |
| 10. Vendor accuracy and false positive claims | States an accuracy rate of 99.9% when spotting AI-generated text vs. human writing, validated on RAID, detecting 95.7% of AI texts at 1% false positives; 1 in 10,402 false positives in the GPTZero 4o post of 18 October 2026 | Accuracy of over 99%; 99.6% on AI-generated samples; false positive rate 1 in 10,000 on public datasets of tens of millions of documents | 99%+ accuracy for 5.0 and 5.0 Max; human specificity 99.954% and 99.958% on PERSUADE 2.0 |
| 11. Independent studies | 11 records: Weber-Wulff 2023, Liang 2023, Walters 2023, Elkhatat 2023, Perkins 2024, RAID 2024, Chaka 2023, Krishna 2023, Orenstrakh 2023, Jabarian and Imas 2025, Van Vlasselaer 2026 | 2 records: Jabarian and Imas 2025, Van Vlasselaer 2026 | 5 records: Liang 2023, Walters 2023, RAID 2024, Orenstrakh 2023, Jabarian and Imas 2025 |
| 12. Data retention and privacy statement | API inputs not stored; dashboard inputs stored and used in aggregate; FERPA, GDPR and CCPA compliance stated (policy of 20 May 2026) | Submitted text stored while the account is active, deletable from the history tab; no training on user submissions; zero data retention for the enterprise API | Scan history may train the models unless the user opts out (policy of 17 June 2026); retained only as long as reasonably necessary |
| 13. Launch year and ownership | 2023; GPTZero LLC; announced 23 June 2026 that it plans to join Superhuman | 2023; Pangram Labs, Brooklyn; no parent stated | 2022; Originality AI Inc., Collingwood, Ontario; parent not published |
| 14. Rankings the card appears in | Independently tested, false positive rate, free tiers, languages, input limits, robustness, false negative rate, claim vs evidence gap, studies by content type | Independently tested, false positive rate, languages, robustness, false negative rate, claim vs evidence gap, studies by content type | Independently tested, false positive rate, languages, robustness, false negative rate, claim vs evidence gap, studies by content type |
Which studies tested all three, or two of the three?
One record ran all three on the same texts, and it is the reason the three-way exists. Jabarian and Imas 2025, an NBER working paper, submitted 1,992 human passages written before 2020 (Amazon reviews, blog posts, news, novels, restaurant reviews and resumes) and 7,968 AI passages from GPT-4.1, Claude Opus 4, Claude Sonnet 4 and Gemini 2.0 Flash to the three commercial APIs and a RoBERTa baseline, and printed every rate at fixed thresholds (NBER w34223, read 9 November 2026; Jabarian and Imas 2025 record). GPTZero and Originality.ai also share four earlier records: Liang 2023, Walters 2023, RAID 2024 and Orenstrakh 2023. GPTZero and Pangram share Van Vlasselaer 2026. Pangram and Originality.ai share Jabarian and Imas 2025 alone.
| Source | GPTZero | Pangram | Originality.ai |
|---|---|---|---|
| Jabarian and Imas 2025, false positive rate at threshold 0.5, 1,992 human passages (Table 3) | 0.0071 | 0.0000 | 0.0011 |
| Jabarian and Imas 2025, false positive rate at the loosest threshold, 0.1 (Table 3) | 0.0071 | 0.0010 | 0.0027 |
| Jabarian and Imas 2025, false positive rate at Youden thresholds, highest genre (Table 6) | 0.0238 | 0.0075 | 0.0218 |
| Jabarian and Imas 2025, false negative rate at threshold 0.9, highest model (Table 4) | 0.0296 | 0.0382 | 0.4242 (Claude Opus 4) |
| Jabarian and Imas 2025, false negative rate at Youden thresholds, highest cell (Table 7) | 0.0700 | 0.0200 | 0.0448 |
| Jabarian and Imas 2025, false negative rate after one StealthGPT rewrite (Table B.4) | 0.2552 to 0.7730 | 0.0000 to 0.0500 | 0.0227 to 0.2073 |
| Jabarian and Imas 2025, false negative rate at a 0.0001 false positive cap (Table C.2) | 1.0000 on all four models | 0.0095 to 0.0211 | 0.0749 to 0.5129 |
| Jabarian and Imas 2025, cost per correctly flagged AI passage, six-genre average (Table B.6) | $0.0575 | $0.0228 | $0.0416 |
| Liang 2023, 91 TOEFL essays | One of seven detectors averaging 61.22% false positives | Not tested | One of the same seven |
| Walters 2023, 126 documents | 81% correct (Pro), 2% of 42 human papers flagged | Not tested | 98% correct (model 1.4), 0 of 42 flagged, 2 uncertain |
| RAID 2024, shared benchmark | 0.00% false positives at threshold 0.5; 98.8% accuracy on open-source chat models at 5% FPR | Not tested | 0.25% false positives at threshold 0.5; 85.0% with no attack, 9.3% under the homoglyph attack |
| Orenstrakh 2023, 164 submissions | 54.39% on 114 human, 52 false positives; 20% of 10 after QuillBot | Not tested | 93.86% on 114 human, 7 false positives; 40% of 10 after QuillBot |
| Van Vlasselaer 2026, 160 papers | 0 of 40 hybrid papers within range; 70% of fully AI papers under 20%; 0 of 40 human papers flagged | 37 of 40 hybrid and 37 of 40 humanised papers within range; 0 of 40 human flagged; 45.5% of 1,163 theses flagged | Not tested |
Results as each study reports them: NBER w34223, arXiv 2304.02819, De Gruyter 10.1515/opis-2022-0158, arXiv 2405.07940, arXiv 2307.07411, DOI 10.1007/s40979-026-00226-w, read 8 and 9 November 2026. "Not tested" means the paper did not run that tool.
One of four comparisons in the compare hub; the two-way pages Pangram vs GPTZero and Pangram vs Turnitin read two of these cards against each other and against Turnitin. The cards are GPTZero, Pangram and Originality.ai, and the home board answers which AI detectors exist and who has tested them?
Which card fits which reader?
- Choose Pangram ifthe false positive rate on a large outside sample decides it: 0.0000 of 1,992 human passages at threshold 0.5 and at most 0.0075 per genre at Youden thresholds in Jabarian and Imas 2025 (NBER w34223, read 9 November 2026).
- Choose Pangram ifyour texts are in languages beyond English and you want them named: 24 listed, against GPTZero's 5 and Originality.ai's unnamed 30+ (pangram.com, read 8 November 2026).
- Choose GPTZero ifyou want the longest outside record to read before you decide: 11 study records from 2023 to 2026, including 0.00% false positives at threshold 0.5 in RAID 2024 and 9 of 18 in Weber-Wulff 2023 (independently tested ranking).
- Choose GPTZero ifyou check long documents free: 150,000 characters per text with a free account and 10,000 words a month (gptzero.me/faq, read 8 November 2026).
- Choose Originality.ai ifyou also check code or scan whole sites: text and code modalities, URL and full-site scans, Bulk Scan on Pro (originality.ai/pricing, read 8 November 2026).
- Choose Originality.ai ifa published 100% on human papers matters: 0 of 42 human papers flagged in Walters 2023 with model 1.4, and 0.0011 of 1,992 in Jabarian and Imas 2025 (De Gruyter PDF, read 8 November 2026).
- Choose Pangram ifhumanised or rewritten AI text is the worry: 0.0000 to 0.0500 missed after a StealthGPT rewrite against 0.0227 to 0.2073 for Originality.ai and 0.2552 to 0.7730 for GPTZero in the same paper, and 37 of 40 humanised papers caught in Van Vlasselaer 2026 (DOI 10.1007/s40979-026-00226-w, read 9 November 2026).
How does the evidence compare, tool by tool?
GPTZero
- Register ID
- AIDI-D-0001
- Study records
- 16 in the index: Weber-Wulff 2023, Liang 2023, Walters 2023, Elkhatat 2023, Perkins 2024, RAID 2024, Chaka 2023, Krishna 2023, Orenstrakh 2023, Jabarian and Imas 2025, Van Vlasselaer 2026, Pegoraro 2023, Kar 2024, Alshammari and Rao 2025, DAMAGE 2025, Epoch AI 2026
- Shared with Pangram and Originality.ai
- 2: Jabarian and Imas 2025, Epoch AI 2026
- Vendor accuracy claim
- An accuracy rate of 99.9% when spotting AI-generated text vs. human writing, validated on RAID, detecting 95.7% of AI texts while only incorrectly predicting 1% of human texts as AI; mixed documents with a 96.5% accuracy rate. (FAQ)
- Free tier
- Free plan covering 10,000 words per month (vendor blog, Oct 25, 2026); 10,000 characters per scan without login
- Languages
- 5 stated
Pangram
- Register ID
- AIDI-D-0006
- Study records
- 4 in the index: Jabarian and Imas 2025, Van Vlasselaer 2026, DAMAGE 2025, Epoch AI 2026
- Shared with GPTZero and Originality.ai
- 2: Jabarian and Imas 2025, Epoch AI 2026
- Vendor accuracy claim
- 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.
- Free tier
- 20 free credits daily (twenty free checks a day after signing up for the free plan, including AI assistance detection)
- Languages
- 24 stated
Originality.ai
- Register ID
- AIDI-D-0004
- Study records
- 7 in the index: Liang 2023, Walters 2023, RAID 2024, Orenstrakh 2023, Jabarian and Imas 2025, Pegoraro 2023, Epoch AI 2026
- Shared with GPTZero and Pangram
- 2: Jabarian and Imas 2025, Epoch AI 2026
- Vendor accuracy claim
- Originality.ai 5.0 and 5.0 Max have 99%+ accuracy (accuracy blog, October 25, 2026). 5.0 launch post: AI recall UChicago Booth AI 97.69% (5.0) / 98.00% (5.0 Max), Epoch AI 99.33%; 5.0 Max identified 90.1% of mixed documents.
- Free tier
- Free, no subscription required, $0: 10K words (100 credits one time)
- Languages
- 30 stated
How this comparison is built
- Cards compared
- GPTZero, Pangram and Originality.ai
- Fields
- The 14 fields of every card, in card order
- Shared study records
- 2: Jabarian and Imas 2025, Epoch AI 2026
- Fields checked
- 8 and 9 November 2026
- Records in the studies index
- 35 records
All three columns are drawn from the same 14 fields in the index data file, read on 8 November 2026 from the vendor page recorded on each card, and from the studies index, where each result is copied as the paper reports the figure, with the base stated. The shared-studies table puts the one three-way record first and lists every two-way record after it, marking the tool a paper did not run as "Not tested" rather than leaving a blank that could read as a zero. A vendor's accuracy or false positive statement is quoted and attributed and never placed in the same cell as a study figure. The rules, the recheck schedule (next pass 8 December 2026) and the correction route are on the methodology page; the independently tested ranking sorts every card on the study count used in the first paragraph.
Two of the three pairs inside this page have their own two-card comparison: Pangram vs GPTZero and, since 9 November 2026, Pangram vs Originality.ai, which reads the Jabarian and Imas 2025 rows for those two tools alone beside RAID 2024 and Van Vlasselaer 2026.
Where each figure comes from
Values as each vendor states them on gptzero.me, gptzero.me/faq, gptzero.me/pricing, pangram.com, pangram.com/pricing, pangram.com/data-privacy, originality.ai, originality.ai/pricing and originality.ai/privacy-policy, all read 8 November 2026. Each criterion row holds the same field for all three cards; a cell carries the vendor's own wording.
What do the three cost over 12 months?
Prices in USD as shown with the $ symbol on gptzero.me/pricing and the GPTZero blog, pangram.com/pricing and originality.ai/pricing, checked 8 November 2026; Originality.ai names USD, Pangram and GPTZero show the symbol only. GPTZero's pricing cards require JavaScript and did not render on that date; the card notes a browser recheck before any purchase decision. The plans include different word allowances, so the figures are not like for like.
Questions about GPTZero, Pangram and Originality.ai
Which study tested GPTZero, Pangram and Originality.ai together?
Jabarian and Imas 2025, an NBER working paper, ran all three APIs on 1,992 human passages written before 2020 and 7,968 AI passages from four 2025 models (NBER w34223, read 9 November 2026). It is the only record in the index that covers all three.
Which of the three had the lowest false positive rate?
Pangram, at 0.0000 of 1,992 human passages at threshold 0.5, then Originality.ai at 0.0011 and GPTZero at 0.0071 (NBER w34223, Table 3, read 9 November 2026). At Youden-optimised thresholds the highest genre figures were 0.0075, 0.0218 and 0.0238 in the same order.
Which missed the least AI text after a rewrite?
Pangram, at 0.0000 to 0.0500 after one StealthGPT rewrite, then Originality.ai at 0.0227 to 0.2073 and GPTZero at 0.2552 to 0.7730 (NBER w34223, Table B.4, read 9 November 2026). Before the rewrite GPTZero's highest miss rate at threshold 0.9 was 0.0296, lower than Pangram's 0.0382.
How many studies tested each tool?
GPTZero 11, Originality.ai 5 and Pangram 2, by the index at version 2026.11.13 (independently tested ranking). GPTZero and Originality.ai share five records, GPTZero and Pangram two, Pangram and Originality.ai one.
Which costs less for an individual?
Originality.ai's Pro is $12.95 per month billed yearly ($155.40) or $14.95 monthly, GPTZero's premium plans start at $12.99 per month per the vendor blog, and Pangram's Individual plan is $20 per month, all checked 8 November 2026 (originality.ai/pricing, gptzero.me, pangram.com/pricing). Word allowances differ: 200K, unstated and 300,000 a month respectively.
Which supports the most languages?
Originality.ai states 30+ without naming them, Pangram names 24 and GPTZero names 5, as read 8 November 2026 (originality.ai, pangram.com, gptzero.me); the languages ranking places Originality.ai above Pangram on the stated count and notes that no study measures accuracy per language.
Which stores the text you paste?
GPTZero stores dashboard inputs and not API inputs; Pangram stores text while the account is active and offers zero retention on the enterprise API; Originality.ai may use scan history to train its models unless the user opts out, all per the policies read 8 November 2026 (data retention ranking; do AI detectors store your text?).
Do all three offer an API?
Yes: GPTZero at docs.gptzero.me, Pangram at docs.pangram.com with metered text at $0.05 per 100 words, and Originality.ai at docs.originality.ai on the Enterprise plan, all read 8 November 2026 (API view).
Sources and revisions
Sources
- GPTZerogptzero.me
- GPTZero FAQgptzero.me/faq
- GPTZero pricinggptzero.me/pricing
- GPTZero blog, 25 October 2026gptzero.me/news/best-ai-detectors
- GPTZero privacy policygptzero.me/privacy-policy.html
- Pangrampangram.com
- Pangram pricingpangram.com/pricing
- Pangram data privacypangram.com/data-privacy
- Originality.aioriginality.ai
- Originality.ai pricingoriginality.ai/pricing
- Originality.ai privacy policyoriginality.ai/privacy-policy
- Jabarian and Imas 2025, NBER w34223nber.org/papers/w34223
- Liang et al. 2023arxiv.org/pdf/2304.02819v3
- Walters 2023degruyterbrill.com/document/doi/10.1515/o...
- Dugan et al. 2024 (RAID)arxiv.org/pdf/2405.07940
- Orenstrakh et al. 2023arxiv.org/abs/2307.07411
- Van Vlasselaer et al. 2026doi.org/10.1007/s40979-026-00226-w
- Weber-Wulff et al. 2023arxiv.org/pdf/2306.15666
