Is there a 100% accurate AI detector? What published tests show

Walters 2023 is the only record with a perfect score. Copyleaks and Turnitin classified all 126 documents correctly, and Originality.ai was correct on 124 of 126 (DOI 10.1515/opis-2022-0158). The same paper reports the floor under the ceiling: across all 16 detectors the average was 95% correct on the 42 GPT-3.5 papers and 61% on the 42 GPT-4 papers, every AI paper was submitted unaltered, and the author names Turnitin's possible training on the human essays as the first limitation. The Walters 2023 record carries every row of Table 4.

Weber-Wulff 2023 tested 14 detectors on 54 documents three months earlier, and no tool passed 80% on averaged accuracy; Turnitin led the binary count at 41 of 54 (DOI 10.1007/s40979-023-00146-z). The class of document decided the score: average binary accuracy in Weber-Wulff 2023 was 94% on human text, 63% and 70% on the two unedited ChatGPT classes, 30% on hand-edited AI text and 15% on Quillbot-paraphrased text. Perkins 2024 measured the same drop a year later: seven detectors averaged 39.5% on unedited AI samples and lost a further 17.4 points once the generating tool applied six adversarial edits, with Turnitin falling from 50% to 7.9% (DOI 10.1186/s41239-024-00487-w). The Weber-Wulff 2023 and Perkins 2024 records carry the full tables.

The claim vs evidence gap ranking, added 9 November 2026, sets each vendor's headline figure beside the best study result for the same tool, from 3.5 points under to 2.67 points over. RAID 2024 shows why one number cannot describe a detector: at a fixed 5% false positive rate GPTZero scored 98.8% on open-source chat models with greedy decoding and 9.4% on open-source non-chat models with sampling (DOI 10.18653/v1/2024.acl-long.674). Human text is the other half of the question: Liang 2023 reports an average false positive rate of 61.22% across seven detectors on 91 essays by non-native English writers (DOI 10.1016/j.patter.2023.100779), and eight of the ten detectors with a record posted 0 false positives in at least one study, per the false positive rate ranking. All nine records sit in the studies index.

The closest any record comes to a clean sheet on human text is Epoch AI 2026, added 9 November 2026: Pangram 3.3.2 and GPTZero model 2026-05-11-base flagged 0 of 495 pre-2022 passages by 99 professional authors, and Pangram also missed 0 of 297 AI passages written from a plain prompt; the same three detectors then missed 30, 32 and 53 of 297 passages written to imitate a named author's style, and Originality.ai flagged 19 of the 495 human passages (epoch.ai, results table, read 9 November 2026; not peer reviewed). Two zeros on 495 and one on 297 are rows in a table, not a claim of 100%: the authors print 90% confidence intervals up to 0.54% and 0.90% on those cells, and the imitation condition shows the same tools missing one AI passage in ten. Two question pages added 9 November 2026 read the same records from other angles: the repeatability question, on whether a score repeats, and the paraphrased text question, on the eleven records that scored rewritten AI text.

Fourteen of the 22 vendors in the index publish an accuracy figure, and eight of those state 99% or higher, by the count of the vendor claim field across all 22 cards read 8 and 9 November 2026. GPTZero states an accuracy rate of 99.9%, validated on RAID, detecting 95.7% of AI texts while flagging 1% of human texts (gptzero.me/faq, read 8 November 2026); RAID's own Table 5 puts GPTZero at 98.8% in one setting and 9.4% in another (DOI 10.18653/v1/2024.acl-long.674), and Weber-Wulff 2023 put GPTZero at 29 of 54. Winston AI states an accuracy rate of 99.89% at a 1% false positive rate (gowinston.ai, read 8 November 2026), against 36 of 54 for Go Winston in Weber-Wulff 2023 (DOI 10.1007/s40979-023-00146-z). The other claims are quoted on each card with source and date: Copyleaks over 99%, Originality.ai 99%+, Pangram over 99%, Grammarly 99%, QuillBot a 99% detection rate, Phrasly 99.8%, ZeroGPT 98.4%, Proofademic 97.6% to 99.9%, JustDone 80%, Codequiry 90% of AI-written code caught, the GPT-2 Output Detector detection rates of about 95% on GPT-2 text, and GLTR a rise in the human detection rate from 54% to 72% with its annotations. Turnitin, AIDetectPlus, Sapling, Walter Writes, Hive, WasItAI and Reality Defender publish no accuracy number; Scribbr publishes the figures it found in other tools.

Every card carries a vendor caveat, and four vendors say the opposite in so many words. Scribbr: "we can't guarantee 100% accuracy", with the highest accuracy Scribbr found at 84% in a premium tool and 68% in any free tool (scribbr.com, read 8 November 2026). Phrasly: "no method of detection is 100% accurate" (phrasly.ai, read 8 November 2026). JustDone: "no honest detector claims 100%" (justdone.com, read 8 November 2026). Walter Writes: "No AI detector can guarantee 100% accuracy" (Walter Writes card). Turnitin publishes no accuracy figure, only a false positive rate under 1% for documents with over 20% AI writing, validated on over 700,000 pre-ChatGPT papers (Turnitin FAQ, read 8 November 2026). The vendor picks the sample, threshold and date; a study picks its own, which is why the false positive rate ranking keeps vendor rates in a separate table and the ranking by independent testing counts records rather than scores.

Four checks turn a headline into a usable figure, and a fifth asks which generator the claim covers: the question page on ChatGPT coverage marks which of the 72 cards name ChatGPT, GPT-4 and GPT-5 and sets the claims beside the four studies that tested ChatGPT text. What was counted: GPTZero's 99.9% is overall accuracy, Turnitin's under 1% is a false positive rate, and RAID's 98.8% is a true positive rate at a fixed 5% false positive rate, so the three cannot be compared. The base: 126 documents in Walters 2023, 54 in Weber-Wulff 2023, 6,287,820 texts in RAID 2024. The text: unedited ChatGPT output scored 63% to 70% in Weber-Wulff 2023 and paraphrased output 15% (DOI 10.1007/s40979-023-00146-z). The date: Walters 2023 ran from 25 June to 12 July 2023, and every vendor has shipped new models since. The methodology page explains why the index never averages two studies, and the home board shows which AI detectors exist and who has tested them.

All read 8 November 2026.

• Walters 2023: https://doi.org/10.1515/opis-2022-0158
• Weber-Wulff et al. 2023: https://doi.org/10.1007/s40979-023-00146-z
• Perkins et al. 2024: https://doi.org/10.1186/s41239-024-00487-w
• Dugan et al. 2024 (RAID): https://doi.org/10.18653/v1/2024.acl-long.674
• Liang et al. 2023: https://doi.org/10.1016/j.patter.2023.100779
• Elkhatat et al. 2023: https://doi.org/10.1007/s40979-023-00140-5
• OpenAI, 2023 classifier announcement and withdrawal note: https://openai.com/index/new-ai-classifier-for-indicating-ai-written-text/
• GPTZero FAQ: https://gptzero.me/faq
• Winston AI home page: https://gowinston.ai/
• Scribbr AI detector: https://www.scribbr.com/ai-detector/
• Phrasly AI detector: https://phrasly.ai/ai-detector
• JustDone AI detector: https://justdone.com/ai-detector
• Turnitin AI writing detection FAQs: https://guides.turnitin.com/hc/en-us/articles/28477544839821
• Chaka 2023: https://doi.org/10.37074/jalt.2023.6.2.12 (read 9 November 2026)
• Epoch AI 2026, AI detectors rarely flag human writing: https://epoch.ai/data-insights/ai-detectors-false-negatives (read 9 November 2026)

Revision log: 2026-11: page created; 2026-11-09: 22-card recount; 2026-11-09: ChatGPT coverage question page linked; 2026-11-09: Pegoraro 2023 and Fleckenstein 2024 added to the records paragraph (fifth wave); 2026-11-09: Epoch AI 2026 row and paragraph added at index version 2026.11.7, three records and twelve cards gained records in the sixth wave; 2026-11-09: repeatability and paraphrased text questions linked (seventh wave)