Verified 2026-10-08
The original RAID research was published at ACL 2024. Its findings describe the detectors evaluated at that time. The maintained RAID leaderboard includes later detector submissions and updated results. This page presents a dated, readable snapshot of those results, with links to the official leaderboard and underlying data.
Quick answer: As of October 8, 2026, Winston AI ranks first on the maintained RAID leaderboard under the selected settings, with a 99.998% TPR at a fixed 1% false-positive rate. Domain, generator model, decoding strategy and repetition penalty are set to all, while adversarial attack is set to none. The ranking changes when those settings change.
Why this leaderboard matters: Many articles still reference results from the original 2024 RAID publication. However, the maintained RAID leaderboard includes newer detector submissions. This snapshot makes those updated results accessible and links directly to the official evidence. Rankings depend on the evaluation settings selected.
Important limitation: This ranking uses RAID’s non-adversarial aggregate. It measures performance on samples with adversarial attack set to none and should not be interpreted as a universal ranking for paraphrased, modified or attacked AI text.
Current RAID results, not the original 2024 rankings. This table reflects a dated snapshot of the maintained RAID leaderboard. The original ACL paper remains the reference for the benchmark methodology.
Setting | Selected value |
|---|---|
Snapshot date | 2026-10-08 |
Primary metric | TPR @ 1% FPR |
Domain | All |
Generator model | All |
Decoding strategy | All |
Repetition penalty | All |
Adversarial attack | None |
Included submissions | 47 |
Latest RAID AI detector rankings
The 44 submissions with a compatible aggregate are sorted by TPR at 1% FPR for non-adversarial text. 3 additional merged submissions are retained as unranked because their official result files do not provide the selected aggregate. Detector names are reproduced from the official result files. A version is shown only when the official detector name includes one.
Rank | Detector | TPR @ 1% FPR | Evidence |
|---|---|---|---|
1 | Winston AI | 99.998% | |
2 | Grammarly | 99.686% | |
3 | Veredict Labs | 99.572% | |
4 | GiantMelon | 99.563% | |
5 | SpeedAI | 99.550% | |
6 | Trinka AI | 99.454% | |
7 | Zhuque AI Detection Assistant | 99.426% | |
8 | MELD | 99.397% | |
9 | QuillBot | 99.235% | |
10 | Slop Or Not | 99.088% | |
11 | TMR (Target Mining RoBERTa) - AI Text Detector | 98.561% | |
12 | Linda-RAID | 98.463% | |
13 | ModernBERT AI Detection | 98.178% | |
14 | slopx-gray-v1 | 97.463% | |
15 | Tabularis ModernBERT R1.1 (NFKC) | 97.202% | |
16 | DeBERTa-ConPara | 96.982% | |
17 | Tabularis AI Detector | 96.643% | |
18 | It's AI | 96.355% | |
19 | GPTZero | 95.721% | |
20 | ADAL | 95.283% | |
21 | GPTypo | 94.002% | |
22 | Argus v5 | 93.425% | |
23 | Fast AI Detector | 91.371% | |
24 | Desklib AI Text Detector v1.01 | 89.107% | |
25 | WPF V141 Photon Fractal | 87.847% | |
26 | Snidewider (panel-HGB) | 87.813% | |
27 | Desklib | 86.272% | |
28 | e5-small-lora | 82.807% | |
29 | BERT-tiny-4M | 80.847% | |
30 | DivEye | 79.572% | |
31 | AIDetector.review | 78.233% | |
32 | MultiScale-Hybrid-mpnet | 78.160% | |
33 | Kareem Elsamadicy (Independent Researcher) | 77.623% | |
34 | Binoculars | 69.537% | |
35 | THE DETECT LAW v7 | 68.915% | |
36 | warm_silence_129 | 65.064% | |
37 | RADAR | 48.086% | |
38 | decting-v0.10 | 47.053% | |
39 | GLTR | 45.572% | |
40 | SuperAnnotate AI Detector | 44.535% | |
41 | RoBERTa-base (GPT2) | 39.827% | |
42 | RoBERTa-large (GPT2) | 38.827% | |
43 | RoBERTa (ChatGPT) | 32.840% | |
44 | LLMDet | 20.164% | |
— | FastDetectGPT | Not available | |
— | Gaussian Extreme | Not available | |
— | Luminar | Not available |
At the selected threshold, Winston AI recorded 54,399 true positives and 1 false negative across 54,400 eligible non-adversarial AI samples.
Official evidence
Readers can verify this snapshot against the official RAID leaderboard, the official GitHub repository, and the ACL 2024 paper. Every ranking row links directly to its merged official results.json file.
Source revision: 3794044125a6. Snapshot extracted on 2026-10-08.
Why the 2024 RAID paper and today's leaderboard show different results
The ACL 2024 paper is a historical evaluation. It reports the systems and results available for the published study, so its tables should not be presented as a continuously updated ranking.
The official RAID repository supports later detector submissions. Submitters provide predictions for the test set, automated evaluation generates a results.json file, and maintainers review the submission before it is merged. A merged result can therefore appear on the maintained leaderboard even though it was not part of the original paper.
The maintained leaderboard also exposes several evaluation views, including TPR at 1% FPR, TPR at 5% FPR and AUROC. Rankings can change across metrics and filters because each view asks a different performance question.
How to interpret RAID scores
TPR at 1% FPR
The true-positive rate measures the share of eligible AI-generated samples detected as AI. Fixing the false-positive rate at 1% constrains the threshold so that approximately 1% of the relevant human samples may be flagged. A higher TPR is better only within the same threshold, dataset and filter configuration.
TPR at 5% FPR
Allowing a 5% false-positive rate usually makes it possible to catch more AI-generated text, but it accepts a looser human-text error constraint. A detector can therefore score differently at 1% and 5% FPR.
AUROC
AUROC summarizes how well detector scores separate AI and human samples across many possible thresholds. It is useful for threshold-independent comparison, but it does not describe performance at one operational false-positive limit.
Adversarial attacks and filters
RAID includes domains, generator models, decoding settings, repetition penalties and adversarial transformations. Changing any of these filters changes the evaluated subset. This page does not combine scores from incompatible configurations.
Methodology and provenance
This snapshot was generated from the 47 merged leaderboard/submissions/*/results.json files present in the official repository at source revision 3794044125a6. For each submission, the page reads score_agg.no_adversarial with domain, model, decoding and repetition penalty set to all, attack set to none, and accuracy key 0.01. The 44 compatible rows are sorted by the unrounded accuracy value, descending, with detector name used only as a deterministic tie-break. A merged result without that aggregate remains visible but unranked.
The displayed score is the official aggregate converted to a percentage and rounded to three decimal places. True-positive and false-negative counts are retained in the generated snapshot files. Missing values are not estimated. Unmerged pull requests are excluded. Detector names and release dates are taken from each official result file.
Independent verification of the Winston AI aggregate
The verification used RAID’s evaluation code and Winston AI result file from pinned source revision 3794044125a6. RAID calibrates a separate threshold for each domain from human-written, non-adversarial samples. Winston’s official result reports a 1% false-positive rate in each of the eight domains. Independently summing the 272 most granular non-adversarial result cells produces 54,399 true positives and 1 false negative across 54,400 samples. Dividing 54,399 by 54,400 reproduces the published aggregate of 0.9999816176470588.
Frequently asked questions
What is the latest RAID AI detector leaderboard?
This page is a 2026-10-08 snapshot of the maintained official RAID repository. It includes 47 merged detector submissions, of which 44 provide the complete aggregate required for this configuration.
Is the 2024 RAID paper outdated?
No. The ACL 2024 paper remains the methodological and historical reference. Its detector rankings describe the systems evaluated for that publication, while the maintained leaderboard can include later approved submissions.
Which AI detector ranks first on RAID?
As of October 8, 2026, Winston AI ranks first on the maintained RAID leaderboard at 99.998% TPR @ 1% FPR with all domains, generator models, decoding strategies and repetition-penalty settings included and adversarial attack set to none. This is a configuration-specific result, not a claim about every RAID view.
How often is the RAID leaderboard updated?
The maintained leaderboard changes when approved submissions or evaluation outputs are merged into the official repository. This page therefore records a source revision and verification date instead of implying that a snapshot is permanently current.
What does TPR at 1% FPR mean?
It is the share of eligible AI-generated samples correctly detected when the decision threshold is calibrated to a 1% false-positive rate on the relevant human samples.
Why do RAID rankings change when filters change?
Different domains, generator models, decoding methods, repetition settings, attacks and metrics create different evaluation subsets or thresholds. Detectors do not perform identically in every condition.
Where can I verify the official RAID results?
Use the official leaderboard for interactive filtering and the official submission directory for the merged result files. Each table row above links to its own results.json evidence.
Update history
2026-10-08 source snapshot
Pinned RAID revision: 3794044125a6.
Configuration: TPR @ 1% FPR; all domains, generator models, decoding strategies and repetition-penalty settings; adversarial attack set to none.
Coverage: 47 merged result files, 44 ranked submissions and 3 unranked submissions whose official files do not provide the selected aggregate.
Leader: Winston AI at 99.998%, independently reproduced from 54,399 true positives and 1 false negative.