Which AI detectors were tested on GPT-4 text? The 13 cards scored on GPT-4 family output in 7 study records, with every result and its base

Seven records, and the GPT-4 share runs from 5 samples to 1,992 passages. Walters 2023 generated 42 literature reviews with ChatGPT-4 in the first week of April 2023, matched to 42 from ChatGPT-3.5 and 42 student essays, and tested 16 detectors between 25 June and 12 July 2023 (DOI 10.1515/opis-2022-0158, read 9 November 2026). Elkhatat 2023 wrote 15 GPT-4 paragraphs and 15 GPT-3.5 paragraphs against 5 human controls (DOI 10.1007/s40979-023-00140-5, read 9 November 2026). Perkins 2024 had GPT-4 write 5 of its 15 original samples of about 500 words, beside Claude 2 and Bard, then applied six adversarial techniques (DOI 10.1186/s41239-024-00487-w, read 9 November 2026). RAID 2024 lists GPT-4 among 11 generators behind 509,014 non-adversarial generations and prints its detector results pooled (DOI 10.18653/v1/2024.acl-long.674, read 9 November 2026). Jabarian and Imas 2025 generated 1,992 passages each from GPT-4.1, Claude Opus 4, Claude Sonnet 4 and Gemini 2.0 Flash (NBER w34223, read 9 November 2026). Van Vlasselaer 2026 used GPT-4o for the literature review and conclusion of 40 hybrid papers and a GPT-4o prompt to humanise them, while its 40 fully AI papers came from ChatGPT Deep Research mode (DOI 10.1007/s40979-026-00226-w, read 9 November 2026). DAMAGE 2025 names the GPT-4 family among the models that wrote one mirror essay per human essay and uses 2,000 GPT-4o mirrors in its fine-tuned humanizer test (ACL Anthology 2025.genaidetect-1.9, read 9 November 2026).

In the two records that print GPT-4 results alone, Copyleaks is at the top both times: 100% of 42 GPT-4 documents in Walters 2023, shared with Turnitin and Originality.ai, and 14 of 15 GPT-4 paragraphs in Elkhatat 2023, where GPTZero fell from 14 of 15 GPT-3.5 paragraphs to 4 of 15 and Writer flagged 1 of 15 (DOI 10.1007/s40979-023-00140-5, read 9 November 2026). Walters 2023 shows the same version gap across its field: 61% average on GPT-4 documents against a higher or equal figure on GPT-3.5 for every card with a printed split, with Sapling at 33% and GPTZero at 57% on GPT-4 (DOI 10.1515/opis-2022-0158, read 9 November 2026). Perkins 2024 is the low point: GPT-4 output was identified 23.9% of the time across seven tools, against 76.9% for Bard and 17.7% for Claude 2 (DOI 10.1186/s41239-024-00487-w, read 9 November 2026). On the 2025 and 2026 text the order changes: GPTZero missed 0.0025 and Pangram 0.0131 of 1,992 GPT-4.1 passages at threshold 0.5 in Jabarian and Imas 2025 (NBER w34223, read 9 November 2026), while Pangram placed 37 of 40 GPT-4o hybrid papers in range in Van Vlasselaer 2026, where GPTZero placed 0 of 40 (DOI 10.1007/s40979-026-00226-w, read 9 November 2026). The false negative rate ranking orders the cards on the miss rates these records produce.

Twenty-nine cards. Eight have study records whose generators are something else, from ChatGPT 3.5 in Kar 2024 to DeepSeek-V3 in Alshammari and Rao 2025 and images in Ha and Passananti 2024, and the table below names them. The other 21 cards have no study record at all: Grammarly, Phrasly, JustDone, Proofademic, Walter Writes, AIDetectPlus, Codequiry, WasItAI, Reality Defender, Grammarly Authorship, Turnitin Clarity, SynthID, Smodin, Surfer, Quetext, TruthScan, Deepware, Sensity, Codeleaks, Pindrop Pulse and Resemble Detect, by the study lists on their cards read 9 November 2026. Of the 13 cards that were tested, 11 name GPT-4, a later GPT model or ChatGPT in the generator field of their card, while GLTR and the GPT-2 Output Detector name only GPT-2 and were scored on GPT-4 era text anyway; the ChatGPT coverage question sets the vendor lists beside the tests.

Cards with study records but no GPT-4 text test

Records and generators as each record's sample table names them, read 9 November 2026.

A GPT-4 result dates the detector as much as the generator: the 2023 figures came from versions that Turnitin, GPTZero and Originality.ai have each replaced by their own dated release notes, and no GPT-4 record scores the versions of November 2026.

All read 9 November 2026 in the primary source; the source of every row is linked in the tables.

• Walters 2023: https://doi.org/10.1515/opis-2022-0158
• Elkhatat 2023: https://doi.org/10.1007/s40979-023-00140-5
• Perkins 2024: https://doi.org/10.1186/s41239-024-00487-w
• RAID 2024 (Dugan et al.): https://doi.org/10.18653/v1/2024.acl-long.674
• Jabarian and Imas 2025, NBER w34223: https://www.nber.org/papers/w34223
• Van Vlasselaer 2026: https://doi.org/10.1007/s40979-026-00226-w
• DAMAGE 2025: https://aclanthology.org/2025.genaidetect-1.9/
• NIST AI 700-1 (2025): https://doi.org/10.6028/NIST.AI.700-1
• Epoch AI 2026 data insight: https://epoch.ai/data-insights/ai-detectors-false-negatives
• Kar 2024: https://doi.org/10.1177/02537176241247934
• Alshammari and Rao 2025: https://arxiv.org/abs/2507.17944
• Weber-Wulff 2023: https://doi.org/10.1007/s40979-023-00146-z
• Sadasivan 2023: https://arxiv.org/abs/2303.11156
• Krishna 2023: https://arxiv.org/abs/2303.13408
• Pegoraro 2023: https://arxiv.org/abs/2304.01487
• Ha and Passananti 2024: https://doi.org/10.1145/3658644.3670306
• RAID image 2025: https://arxiv.org/abs/2506.03988

Revision log: 2026-11-09: page created from the sample and results tables of 7 study records and the study lists of 42 cards at index version 2026.11.7.