Which AI detector is most accurate? Our benchmark results

Why benchmark results matter more than accuracy claims

“99% accurate” is easy to print and hard to verify. Accuracy depends on the mix of samples, the threshold, the generators included and whether the test text was edited. A detector that labels everything as AI is 100% accurate on a test set of AI text. Useful numbers say which data was used, how the threshold was chosen and how many human samples were wrongly flagged at that threshold.

We report two kinds of result. Public benchmarks are research datasets that other detectors are also measured on, so the numbers can be compared. Our own labelled sets are built to resemble what people actually check, such as 2026 chatbot output and human writing by non-native English speakers. Each section says which is which, and every model listed here is the one running on the site.

Text: our Ettin-150M detector

The deep scan uses Ettin-150M, an open encoder model that we fine-tuned to tell human writing from machine text. It was trained on the training splits of two public research corpora, RAID and MAGE, after removing invisible characters and mapping look-alike letters back to Latin.

The RAID figure above uses RAID’s own measure and attack set, but on a held-out portion of its public training data, not on the official hidden test set. We also submitted an earlier version of the model, trained on RAID alone, to the official RAID leaderboard. The RAID maintainers’ evaluation on the hidden test set gave it:

That RAID-only version is a research model: it scores well on RAID but wrongly flags far more human writing outside RAID’s domains, so the site runs the later version described above. RAID is one benchmark; the labelled-set results below are a better guide to everyday use.

The weak spot is clear: paraphrased text. When AI output is run through a rewriting model, the deep scan catches only about a third of it at a 5% false-positive rate. Our humanizer page explains what still gives such text away.

Text check with deep scan — AUC 0.959 on 2,467 labelled text samples (cross-validated)

2,467 labelled samples (1,258 AI, 1,209 human), run 2026-10-11. At the “Likely AI” line it caught 78% of AI samples and wrongly flagged 5% of human ones.

Voice: our WavLM voice-clone detector

The optional neural voice check is a WavLM-base-plus model that we fine-tuned on ASVspoof 2019 LA, WaveFake and LJSpeech. During training, clips were re-encoded as MP3, Opus and AAC so that the model learns the voice and not the file format. It is scored over four windows of each clip.

In-the-Wild and 2021 DF were each measured on random 6,000-clip samples. For reference, the XLSR-SLS model reports 7.46% on In-the-Wild and 3.93% on 2021 DF in its paper, using a backbone roughly three times larger. The In-the-Wild result matters most here, because that benchmark was assembled from the internet rather than a lab and none of it was in our training data. One caution: these are clean benchmark clips, and a voice note recorded through a phone speaker and re-recorded is a harder case.

On the site, the neural score is combined with the other audio measurements. Those combination weights were fitted on 300 ASVspoof 2019 and 300 In-the-Wild clips. The “Likely AI” line was set so that no more than about 5% of the real In-the-Wild voices cross it, because recordings from the internet are closer to what people check than studio-clean benchmark audio. The cross-validated results of the complete audio check:

Audio check with the neural voice check — AUC 0.978 on 600 labelled audio clips (cross-validated)

600 labelled samples (300 AI, 300 human), run 2026-10-12. At the “Likely AI” line it caught 89% of AI samples and wrongly flagged 3% of human ones.

Images: Community Forensics classifier plus our fine-tune

The image check averages two vision transformers: the Community Forensics detector (CVPR 2025) and our own fine-tune of it on recent generators, using only commercially licensed images. Each image is also scored on native-resolution crops. On the DRAGON benchmark of diffusion-model images, measured against CC0 human photos from the megalith collection, the original Community Forensics model alone reaches an AUC of 0.996. We don’t quote an ensemble figure for DRAGON, because our fine-tune was trained on other images from it. DRAGON images are also clean generator output, so the labelled set below, with modern generators and re-encoded photos, is the more realistic test.

Image check — AUC 0.98 on 635 labelled images (cross-validated)

635 labelled samples (360 AI, 275 human), run 2026-10-12. At the “Likely AI” line it caught 84% of AI samples and wrongly flagged 5% of human ones.

How these numbers are produced

The site runs 8-bit versions of each model to keep downloads small. Before one ships, we compare it with the full-precision original on the same samples. One audio export that quantised the convolutional layers raised the error rate from 5.7% to 29%, so it was rejected. Thresholds on our own sets are chosen so that at most about 5% of human samples are labelled “Likely AI”, and the reported scores are cross-validated so the threshold isn’t tuned on the same samples it is judged on. Some research benchmarks forbid commercial use, so we use them only to guide development and do not publish scores on them. The methodology page describes the scoring model in more detail.

Changelog

• 2026-10-12 — Image check becomes a two-model ensemble (labelled-set AUC 0.94 → 0.98). Audio combination weights fitted. This page published.
• 2026-10-11 — Neural voice check (WavLM) added as an option.
• 2026-10-11 — Deep text scan switches to our Ettin-150M model (labelled-set AUC 0.912 → 0.959).
• 2026-10-12 — Official RAID test-set evaluation: 99.73% TPR at 5% FPR with attacks.
• 2026-10-10 — RAID leaderboard submission opened.