GPTZero AI detector

Vendor GPTZero LLC Official site gptzero.me How this card is sourced

AIDI-D-0001

Updated ·

GPTZero, launched in 2023, detects AI text in 5 languages according to its vendor page read on 8 November 2026 (gptzero.me). Eleven of the 35 study records in the index cover GPTZero, seven on the morning of 9 November 2026 and four more added later that day. How this card is sourced explains the rules; the full detector index holds the other twenty-one cards.

What GPTZero detects

The GPTZero register entry: field, value, source and read date
FieldValueSource
Content types detectedText, imagegptzero.me
Generators coveredClaude, ChatGPT, GPT-3, GPT-4, GPT-5, GPT-6, Gemini, Llama, Deepseek, Grok, AI services based on those modelsgptzero.me/faq
LanguagesEnglish, German, Portuguese, French, Spanishgptzero.me
Input limitsWidget: 10,000 characters per scan without an account. With a free account: texts up to 150,000 characters, or a batch of up to 250 files. File types and size not published.gptzero.me/faq
Free tierFree plan covering 10,000 words per month (vendor blog, Oct 25, 2026); 10,000 characters per scan without logingptzero.me/news/best-ai-detectors
APIYes, docs.gptzero.medocs.gptzero.me
IntegrationsChrome Extension (Gmail, Google Docs, Classroom, social media), Google Docs, Google Classroom, Canvas, Zapier, APIgptzero.me
Data retentionPrivacy policy (Last updated May 20, 2026) lists submitted content for use of providing our services as collected data. FAQ: We do not store or collect the documents passed into any calls to our API. However, we do store inputs from calls made from our dashboard. This data is only used in aggregate by GPTZero to further improve the service for our users. Compliant with FERPA, GDPR, and CCPA; personal data is deleted on request.gptzero.me/privacy-policy.html
Launch year2023gptzero.me
OwnershipGPTZero LLC; announced June 23, 2026 that GPTZero plans to join Superhuman. Footer Company menu lists Superhuman and Grammarly.gptzero.me/news/preserving-whats-human
Independent studies16 recordsStudies index

GPTZero scans text and, through a feature the vendor calls AI Vision, images. The FAQ lists Claude, ChatGPT, GPT-3, GPT-4, GPT-5, GPT-6, Gemini, Llama, Deepseek and Grok (gptzero.me/faq, read 8 November 2026). The 5 languages are English, German, Portuguese, French and Spanish (gptzero.me, read 8 November 2026).

On edited AI text, the GPTZero 4o launch post states that the model "deliberately tackles the harder cases for AI detection: AI-generated text that has been paraphrased or mixed with human writing" (launch post, read 8 November 2026). The FAQ still says the classifier "is not trained to identify AI-generated text after it has been heavily modified after generation". Source: gptzero.me/faq, read 8 November 2026. This card carries both sentences.

GPTZero LLC announced on June 23, 2026 that GPTZero plans to join Superhuman (announcement, read 8 November 2026); Grammarly is the sister card. Integrations are a Chrome extension, Google Docs, Google Classroom, Canvas, Zapier and an API documented at docs.gptzero.me (gptzero.me, read 8 November 2026).

Limits and plans

GPTZero plans with price, limits and the date each was checked
PlanPriceLimitsChecked on
Premium (individual plans)premium plans starting at $12.99/monthIndividual plan cards on gptzero.me/pricing did not render (JavaScript) on 2026-11-08; the only readable figure is the vendor blog's starting price. Re-check in a browser before shipping.
Professional Team Plan500,000 words per member, per month; min 2 seats. Price not rendered.

Without an account the widget takes 10,000 characters per scan. With a free account GPTZero accepts texts up to 150,000 characters or a batch of up to 250 files (gptzero.me/faq, read 8 November 2026). The free plan covers 10,000 words per month, as stated in a vendor blog post dated October 25, 2026 (vendor blog, read 8 November 2026).

Prices below come from the GPTZero pricing page, read 8 November 2026; the individual plan cards there load through scripts, so the only readable individual price was the starting figure in the vendor blog.

On retention, the FAQ says: "We do not store or collect the documents passed into any calls to our API. However, we do store inputs from calls made from our dashboard." The privacy policy of May 20, 2026 lists submitted content as collected data (read 8 November 2026).

What the vendor claims

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)
gptzero.me/faq, read

GPTZero LLC states 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." The sentence is from the GPTZero FAQ, read 8 November 2026. The home page headline gives 99.9%; the FAQ gives the RAID operating point of 95.7% at 1%; this card uses the FAQ sentence. The RAID leaderboard entry GPTZero submitted on 7 October 2025, read 9 November 2026, scores 97.24% of 652,800 generations at a 5% false positive rate and 92.93% at 1%, with an AUROC of 0.9925 (github.com/liamdugan/raid, read 9 November 2026); the FAQ's 95.7% at 1% is the vendor's own figure, undated on the page, and the two are set side by side on the leaderboard question.

On false positives, GPTZero LLC says: "Only incorrectly predicting 1% of human texts as AI" and "We repeatedly train our model to reduce false positive rate for ESL writing to 1%" (gptzero.me/faq, read 8 November 2026). The 4o launch post of October 18, 2026 gives "a false positive rate of only 1 in 10,402 on diverse student writing corpra" (gptzero.me/news/introducing-gptzero-4o/, read 8 November 2026). The FAQ gives 1%; the launch post gives 1 in 10,402; this card uses the FAQ figure.

A GPTZero percentage shows how the model weighed the text; the percentage is not a record of who typed the words. GPTZero LLC itself says: "These results should not be used to punish students" (gptzero.me/faq, read 8 November 2026).

What independent studies found

  1. AIDI-S-0001Weber-Wulff 20232023

    GPT Zero: 29 of 54 documents correct (54% binary accuracy, rank 8 of 14); 68% logarithmic accuracy; 9 false accusations in 18 human-written documents (50%, the highest of any tool; 3 on human texts, 6 on machine-translated human texts); only 3 of 36 AI documents undetected (8.3%, the lowest of any tool). Tested March 7 to 28, 2023.

  2. AIDI-S-0002Liang 20232023

    GPTZero was one of seven detectors tested (accessed March 15, 2023). Across the seven, the paper reports an average false positive rate of 61.22% on 91 human-written TOEFL essays and near-perfect accuracy on 88 US 8th-grade essays; 18 of 91 TOEFL essays were flagged as AI by all seven detectors. Per-detector rates appear only in a bar chart (Fig. 1a) and are not stated as numbers.

  3. AIDI-S-0003Walters 20232023

    GPTZero (Pro): 81% of 126 documents correct, 4% incorrect, 15% uncertain; 98% correct on 42 GPT-3.5 papers, 57% on 42 GPT-4 papers; 88% of 42 human papers correct, 2% incorrect (false positive rate 2%). Tested June 25 to July 12, 2023.

  4. AIDI-S-0004Elkhatat 20232023

    GPTZero: GPT-3.5 sensitivity 93% (14 of 15 paragraphs flagged) and specificity 80% (4 of 5 human controls correct, 1 false positive, Human 2 labelled Likely AI-Generated); GPT-4: 4 of 15 flagged, 11 false negatives.

  5. AIDI-S-0005Perkins 20242024

    GPTZero: lowest baseline mean accuracy of the seven tools, 26.3% (Table 7); 26.4% on non-manipulated AI text falling to 16.7% after adversarial techniques (drop 9.7%, Table 8); 1 false accusation in 10 human samples (FAS 10%); 58 of 104 AI samples undetected (UDC 55%).

  6. AIDI-S-0006RAID 20242024

    GPTZero: false positive rate 0.03% at naive threshold 0.25 and 0.00% at 0.5, 0.75 and 0.95 on human-written documents (Table 4); accuracy at a fixed 5% FPR 98.8% on open-source chat models (greedy) but 9.4% on open-source non-chat models with sampling (Table 5); unusually robust to adversarial attacks, losing only 0.3% under the homoglyph attack (Table 6).

  7. AIDI-S-0009Chaka 20232023

    Chaka 2023: GPTZero classified 8 of 21 English chatbot responses as AI-generated, misclassified 10 as human-written and was indecisive on 3 (rank 3 of 5); all 30 ChatGPT responses machine-translated into German, French, Spanish, Southern Sotho and isiZulu were classified as human-written. Free version, tested 30 March to 2 April 2023.

  8. AIDI-S-0011Krishna 20232023

    GPTZero detected 13.9% of GPT-2 XL, 8.7% of OPT-13B and 7.1% of GPT-3.5 text at a fixed 1% false positive rate on 3,000 Wikipedia prompts of 300 tokens, falling to 1.2%, 1.0% and 1.8% after one DIPPER paraphrase (Table 1).

  9. AIDI-S-0012Orenstrakh 20232023

    Ranked last of eight at 54.39% on 114 human and 45.00% on 20 ChatGPT English coursework submissions, with 52 false positives in 114; accuracy on 10 QuillBot-paraphrased ChatGPT texts fell from 70% to 20% (Tables 6, 8 and 9).

  10. AIDI-S-0014Jabarian and Imas 20252025

    False positive rate 0.0071 at every fixed threshold on 1,992 human passages and at most 0.0238 per genre at Youden-optimised thresholds; false negative rate 0.0020 to 0.0296 before and 0.2552 to 0.7730 after StealthGPT rewrites of 1,992 AI passages per model (Tables 3, 6, 4 and B.4).

  11. AIDI-S-0015Van Vlasselaer 20262026

    Scored 70% of 40 fully AI-generated papers under 20% and the rest between 20% and 40%, classified 0 of 40 hybrid and 1 of 40 humanised papers within their known AI range, and classified all 40 human papers correctly with a mean overestimate of 1.95 points (Tables 4 and 5).

  12. AIDI-S-0018Pegoraro 20232023

    True positive rate 27.3% on ChatGPT responses and true negative rate 93.5% on human responses in a 20-method sweep on a ChatGPT and human Q&A benchmark, March 2023 (Table I).

  13. AIDI-S-0023Kar 20242024

    Read one 500-word ChatGPT 3.5 article as 97% AI and the same text after a QuillBot paraphrase as 29 AI; 38 after Grammarly and 97 after ChatGPT paraphrases (Table 1, single text).

  14. AIDI-S-0024Alshammari and Rao 20252025

    Average AI score 100% on 49 DeepSeek-V3 answers and average human score 98.2% on 49 human answers; 94.1% on DeepThink text, falling to 52% after a humanized paraphrase (Tables III and V, March 2025).

  15. AIDI-S-0025DAMAGE 20252025

    True positive rate 99.73% on AI-written student essays and 60.04% after humanizing at a fixed 5% false positive rate; 95.60% and 34.53% at the default threshold, with a 1.47% default-threshold false positive rate on human essays (Tables 3 and 7; essay count not printed).

  16. AIDI-S-0026Epoch AI 20262026

    GPTZero (model 2026-05-11-base) flagged 0 of 495 pre-2022 human passages (0.00%), missed 2 of 297 AI passages from a plain prompt (0.67%) and 32 of 297 (10.77%) 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).

  17. AIDI-S-0032GPTZero and Pangram vendor benchmarks2026, vendor benchmark

    Vendor benchmark (GPTZero's own, model 4.3b, page of 5 February 2026 updated 20 July 2026): false positive rate 0.08%, recall 99.60%, 99.76% (accuracy) averaged over four domains of 1,000 human and 1,000 AI texts each; 0.21% and 91.80% on 1,000 bypassed AI texts. Measured by GPTZero on GPTZero-built sets.

Weber-Wulff 2023 scored GPT Zero at 29 of 54 documents correct (54%, rank 8 of 14), with 9 false accusations in 18 human-written documents, the highest of any tool, and 3 of 36 AI documents undetected, the lowest (arXiv 2306.15666, read 8 November 2026).

Liang 2023 ran seven detectors, GPTZero among them, on 91 human-written TOEFL essays; the average false positive rate across the seven was 61.22%, with per-detector rates only in a bar chart (arXiv 2304.02819, read 8 November 2026).

Walters 2023 found GPTZero (Pro) 81% correct on 126 documents, 57% on 42 GPT-4 papers, and 2% of 42 human papers classified as AI (De Gruyter PDF, read 8 November 2026).

Elkhatat 2023 reports sensitivity of 93% on GPT-3.5 (14 of 15 paragraphs), specificity of 80% (1 false positive in 5 human controls), and only 4 of 15 GPT-4 paragraphs flagged (Springer PDF, read 8 November 2026).

Perkins 2024 gives GPTZero the lowest baseline mean accuracy of seven tools, 26.3%, with 1 false accusation in 10 human samples and 58 of 104 AI samples undetected (arXiv 2403.19148, read 8 November 2026).

The RAID benchmark (2024) reports a false positive rate of 0.03% at threshold 0.25 and 0.00% at 0.5 and above, accuracy at a fixed 5% FPR of 98.8% on open-source chat models but 9.4% on non-chat models with sampling, and a loss of only 0.3% under the homoglyph attack (arXiv 2405.07940, read 8 November 2026).

Chaka 2023 ran the free version of GPTZero on 21 English chatbot responses between 30 March and 2 April 2023 and reports 8 classified as AI-generated, 10 misclassified as human-written and 3 indecisive (rank 3 of 5 tools); all 30 ChatGPT responses machine-translated into German, French, Spanish, Southern Sotho and isiZulu were classified as human-written (DOI 10.37074/jalt.2023.6.2.12, read 9 November 2026).

Four records were added on 9 November 2026. Krishna 2023 reports GPTZero detecting 13.9% of GPT-2 XL, 8.7% of OPT-13B and 7.1% of GPT-3.5 text at a fixed 1% false positive rate on 3,000 Wikipedia prompts, falling to 1.2%, 1.0% and 1.8% after one DIPPER paraphrase, and the authors "recommend against using" GPTZero and RankGen (arXiv 2303.13408, read 9 November 2026). Orenstrakh 2023 ranked GPTZero last of eight at 54.39% on 114 human and 45.00% on 20 ChatGPT coursework submissions, with 52 false positives in 114 and a fall from 70% to 20% on 10 QuillBot-paraphrased texts (arXiv 2307.07411, read 9 November 2026). Jabarian and Imas 2025 reports a false positive rate of 0.0071 at every fixed threshold on 1,992 human passages, false negative rates of 0.0020 to 0.0296 before and 0.2552 to 0.7730 after StealthGPT rewrites (NBER w34223, read 9 November 2026). Van Vlasselaer 2026 reports 70% of 40 fully AI-generated papers scored under 20%, 0 of 40 hybrid and 1 of 40 humanised papers classified within their known AI range, and all 40 human papers classified correctly (DOI 10.1007/s40979-026-00226-w, read 9 November 2026).

Pegoraro 2023, added 9 November 2026 as the twelfth record, ran 20 detection methods on a ChatGPT and human question-and-answer benchmark in March 2023 and records GPTZero at a 27.3% true positive rate on ChatGPT responses and a 93.5% true negative rate on human responses (arXiv 2304.01487, Table I, read 9 November 2026); no ranking field takes a value from the printed rates.

Four records joined the list on 9 November 2026 and took the count to 16: Kar 2024 read one ChatGPT 3.5 article at 97% AI and 29 after a QuillBot paraphrase; Alshammari and Rao 2025 recorded a mean AI score of 100% on 49 DeepSeek-V3 answers falling to 52% after a humanized paraphrase; DAMAGE 2025, written by Pangram Labs, measured a 60.04% true positive rate on humanized student essays at a fixed 5% false positive rate, which replaced the 20% Orenstrakh 2023 figure in the robustness field; and Epoch AI 2026 recorded 0 of 495 pre-2022 human passages flagged and 2 of 297 plain-prompt AI passages missed, a rate that does not replace the 0.25% already in the false negative field. GPTZero's own benchmarking page (0.08% false positive rate for model 4.3b, 5 February 2026, updated 20 July 2026) is recorded as a dated vendor measurement in the vendor benchmarks record and feeds no rank field.

Where GPTZero appears in the index

Rankings

Fits these use cases, questions and comparisons

GPTZero qualifies for all five launch rankings and four of the five added on 9 November 2026: independently tested on 11 studies, false positive rate on the RAID 2024 figure of 0.00%, free tiers on 10,000 words per month, and languages on a count of 5 (detectors.json, read 8 November 2026).

Since 9 November 2026 GPTZero also ranks fourth in false negative rate on 0.0025 of 1,992 GPT-4.1 passages in Jabarian and Imas 2025, seventh in robustness to paraphrase on 2 of 10 QuillBot-paraphrased texts in Orenstrakh 2023, fifth in the claim vs evidence gap on 99.9% claimed against 98.8% in RAID 2024, and second in studies by content type with no record on its image detector.

GPTZero fits business through the API and the team plan, school districts through Canvas and Google Classroom, agencies through the API and the 250-file batch, and recruiters through the free monthly allowance on text (gptzero.me, read 8 November 2026). Comparable LMS-connected cards are Turnitin and Copyleaks; the home page carries every AI detector we track. The only other cards with four or more study records are Turnitin and the GPT-2 Output Detector, OpenAI's free 2019 baseline, which four of GPTZero's six studies also ran.

The card also sits in the two rankings added on 9 November 2026 where its fields qualify: API pricing lists GPTZero in the no-published-price table because the API price renders by script, and minimum input lists it among the cards that state no minimum.

Since 9 November 2026 GPTZero also leads two pages added that day, the false positive studies question with 11 of its 16 records printing a human-text figure and the non-English evidence ranking with five records, appears on the GPT-4 question with seven records, and ranks eighth on LMS coverage with two named platforms and thirteenth on generators covered with ten named models.

Questions about GPTZero AI detector

How many languages does GPTZero detect?

Five. GPTZero lists English, German, Portuguese, French and Spanish on its home page, read 8 November 2026, and that count places the card in the languages ranking.

Is GPTZero free?

Yes, within limits. The free plan covers 10,000 words per month per the vendor blog of October 25, 2026, read 8 November 2026, and the widget takes 10,000 characters per scan with no login. Paid plans start at $12.99/month per the same blog.

What false positive rate do studies give GPTZero?

Between 0.00% and 50%, depending on the study. RAID 2024 reports 0.00% at threshold 0.5; Walters 2023 reports 2% of 42; Perkins 2024 reports 1 of 10; Elkhatat 2023 reports 1 of 5; Weber-Wulff 2023 reports 9 of 18. The card ranks on the RAID figure.

Sources and revisions

Sources

  1. gptzero.meRead 8 November 2026
  2. gptzero.me/faqRead 8 November 2026
  3. gptzero.me/pricingRead 8 November 2026
  4. gptzero.me/news/best-ai-detectorsRead 8 November 2026
  5. gptzero.me/news/introducing-gptzero-4oRead 8 November 2026
  6. gptzero.me/news/preserving-whats-humanRead 8 November 2026
  7. gptzero.me/privacy-policy.htmlRead 8 November 2026
  8. docs.gptzero.meRead 8 November 2026
  9. arxiv.org/pdf/2306.15666Read 8 November 2026
  10. arxiv.org/pdf/2304.02819v3Read 8 November 2026
  11. degruyterbrill.com/document/doi/10.1515/o...Read 8 November 2026
  12. link.springer.com/content/pdf/10.1007/s40...Read 8 November 2026
  13. arxiv.org/pdf/2403.19148Read 8 November 2026
  14. arxiv.org/pdf/2405.07940Read 8 November 2026
  15. doi.org/10.37074/jalt.2023.6.2.12Read 9 November 2026

Revision log

  1. Card created
  2. Chaka 2023 record added, 7 study records
  3. Krishna 2023, orenstrakh 2023, jabarian and imas 2025 and van vlasselaer 2026 records added, 11 study records
  4. Pegoraro 2023 record added to the study list in the fifth wave2026-11-09: kar 2024, alshammari and rao 2025, damage 2025 and epoch ai 2026 records added in the sixth wave, 16 study records
  5. Seventh-wave question pages and three rankings linked2026-11-09: raid leaderboard entry of 7 october 2025 added beside the faq claim

Corrections log

Related reading