Which AI detectors have a public leaderboard entry? The six cards on the RAID leaderboard, and the four benchmarks where no indexed detector appears

Six cards, holding 7 of the 62 entries in the submissions directory on 9 November 2026. Four were submitted by the vendors themselves, each file naming a company contact: QuillBot on 30 January 2025, GPTZero on 7 October 2025, Grammarly on 9 February 2026 and Winston AI on 24 September 2026 (github.com/liamdugan/raid, read 9 November 2026). Two are open tools the RAID authors entered as baselines: GLTR, dated to its 2019 paper, and the GPT-2 Output Detector in both RoBERTa sizes. Winston AI's entry posts the highest figure of the six, 99.98% of 652,800 generations detected at a 5% false positive rate across all domains, generators and adversarial attacks, with 99 misses; GLTR and the two RoBERTa models detect about half. Two commercial detectors that the 2024 paper scored, Originality.ai and ZeroGPT, have no entry in the submissions directory, so their RAID figures remain the paper's: 85.0% with no attack for Originality and a 1.42% false positive rate at threshold 0.5 for ZeroGPT (DOI 10.18653/v1/2024.acl-long.674, read 8 November 2026).

The leaderboard figure and the paper figure for the same vendor are not the same number. The paper's Table 5 placed GPTZero at 98.8% on open chat models under greedy decoding and 9.4% on open non-chat models under sampling, in 2024, on the version then live (RAID record); the leaderboard entry of 7 October 2025 is one aggregate over every cell, 97.24%, for a model more than a year newer. Readers quoting either need the date, which is why the repeatability question lists RAID among the records that change under a tool.

Four, all with a record in the index, and all for reasons the organisers state. NIST GenAI scored 348 submissions from 11 discriminator teams on news summaries, identifies each team by a five-character label and names no product; its RoBERTa-base-openai-detector baseline is the GPT-2 Output Detector's model, but no result is printed under that name (DOI 10.6028/NIST.AI.700-1, Table 2 and Section 4.2.5, read 9 November 2026). ASVspoof 5 scored 53 teams on 680,774 utterances and states that "neither the team name nor the names of team members can be revealed"; none of the six voice cards is a participant, and Pindrop appears as a co-author's affiliation and workshop sponsor rather than an entrant (arXiv 2601.03944, Section IV footnote 8, read 9 November 2026). SAFE 2025 names its top five teams, ISPL, Viper-Purdue, JAIST-HIS, Anon_Peking and Hangzhou Dianzi University, and no indexed card is among its 12 (arXiv 2510.03387, Table 5, read 9 November 2026). DFDC 2020 published its whole Kaggle leaderboard of 2,114 teams, led by Selim Seferbekov at a log loss of 0.4279, and the entry ranked eighth, the biometrics vendor ID R&D, has no card (Meta AI results post, read 9 November 2026).

Public benchmarks in the index with no entry from an indexed card

Entrant counts and figures as each record prints them from its primary paper, read 9 November 2026.

Because the three media benchmarks score research teams, not products. ASVspoof 5 and SAFE 2025 ran on the organisers' terms, with anonymised or institution-named teams, and DFDC ran as a Kaggle competition that closed on 31 March 2020, with no indexed vendor among its named prize winners (DFDC 2020 record). The nearest thing to a leaderboard claim among the media cards is a vendor's own footnote: Pindrop states "99% Detection accuracy against known engines" computed "on the ASVspoof5 dataset of 32 engines", a run on the challenge's data that neither ASVspoof paper prints (pindrop.com, read 9 November 2026), and Resemble Detect cites a "Podonos Audio DFD Bench" with pooled accuracy figures for itself and four competitors (resemble.ai, read 9 November 2026). Both are vendor pages, so they sit on the cards as claims and not here as entries. The modality records ranking shows the same gap from the study side: Hive has 2 records, AI or Not 1 and Sightengine 1, all on images, and no voice or video card has a record of any kind.

A leaderboard entry is a score on one test set, submitted by whoever chose to submit, on the date the file carries, and 36 of the 72 cards, read 9 November 2026 in the data file, have none at all.

All read 9 November 2026 unless an earlier date is shown.

• RAID leaderboard: https://raid-bench.xyz/leaderboard (page shell only; data read from the repository below)
• RAID leaderboard submissions, metadata and results files: https://github.com/liamdugan/raid/tree/main/leaderboard/submissions
• Dugan et al. 2024, RAID: https://doi.org/10.18653/v1/2024.acl-long.674 (read 8 November 2026)
• NIST AI 700-1: https://doi.org/10.6028/NIST.AI.700-1
• ASVspoof 5 challenge summary: https://arxiv.org/abs/2408.08739
• ASVspoof 5 evaluation paper: https://arxiv.org/abs/2601.03944
• SAFE 2025: https://arxiv.org/abs/2510.03387
• DFDC 2020 dataset paper: https://arxiv.org/abs/2006.07397
• Meta AI, Deepfake Detection Challenge results: https://ai.meta.com/blog/deepfake-detection-challenge-results-an-open-initiative-to-advance-ai/
• Pindrop Pulse for Contact Centers: https://www.pindrop.com/pulse-for-contact-centers
• Resemble AI Detect: https://www.resemble.ai/products/detect

Revision log: 2026-11-09: page created from the RAID leaderboard files and the NIST, ASVspoof 5, SAFE 2025 and DFDC 2020 records at index version 2026.11.8.