Forensic AI detection for text, image, audio, and video. Six peer-reviewed signals. Visualized residuals. Citations on every finding. Not a black-box score.
Six signals + watermark scan + C2PA, in your browser.
Best open detector across 291 generators (May 2026).
Checks run locally. Nothing leaves your device by default.
Adobe · OpenAI · Sony · Leica · BBC and more.
A pathologist doesn’t hand you a diagnosis — they walk you through the slides. We took the same posture. Below is a real case file: the six forensic signals, the residuals they produced, what each reading means, and the paper it traces to.
Recompressing at the original quantization table yields an almost-uniform residual — characteristic of a single fresh encoding from a generative pipeline. Real photos preserve edge-localized JPEG scars this image lacks across 96.4% of the frame.
[01] Krawetz, Hacker Factor (2007)
The 2-D Fourier magnitude shows six off-DC bright lobes arranged in a near-symmetric pattern — a known fingerprint of diffusion samplers leaking U-Net upsampling stride into the frequency domain.
[02] Corvi et al., CVPR (2023)
Green-channel histogram peaks 14% higher than red and blue means — an artifact of how generative models hallucinate skin and foliage. A Bayer-array sensor would distribute that luminance evenly.
[03] McCloskey & Albright, ICIP (2019)
The residual carries a repeating horizontal pattern at a 13-pixel period. A real CMOS sensor produces spatially independent shot noise; this regularity is consistent with a VAE decoder learning a low-rank residual prior.
[04] Lukáš et al., IEEE TIFS (2006)
A learned detector trained on 2.6M frontier samples concentrates on iris detail and jaw-skin transitions — the two regions where current diffusion models still struggle with specular-reflection consistency.
[05] Selvaraju et al., ICCV (2017)
No cryptographically-signed manifest was found in the container. This is not itself evidence of synthesis — most camera output is unsigned — but it removes the strongest counter-signal we could have offered.
[06] C2PA Spec 2.1 (2024)
We re-run the ensemble after common evasion passes. A real forensic verdict has to be stable under recompression, resize, and screenshot capture — single-classifier detectors collapse from AUC 0.93 → 0.62 here.
Even a perfect detector tells you what — not who. When a file ships a signed C2PA manifest, the signature chain tells you the camera or model, every recorded edit, and a verifiable timestamp. Adobe, OpenAI, Sony, Leica, and the BBC all sign at the point of capture.
Each modality runs the same examination structure — multiple signals, visualized residuals, plain-English findings, citations.
Per-token predictability heatmap, sentence-rhythm burstiness, AI-tell phrase lookup, n-gram repetition, and a Kirchenbauer green-list watermark scan that flips at z > 4.
Six panels: ELA · FFT log-magnitude · RGB decomposition · PRNU noise residual · Grad-CAM model attention · C2PA manifest reader. All client-side.
STFT spectrogram, spectral flatness uniformity, harmonic stability, cepstral echo-hiding watermark. Codec-aware for VoIP.
12-frame sampling, noise-σ series, motion residual, and per-key-frame ELA + noise residual.
When a file ships a signed manifest, we read it before heuristics. Provenance answers who, not just what.
Originality, GPTZero, Hive — every major detector outputs one confidence score. The May 2026 GenImage benchmark put the best open single-classifier at 75% mean accuracy across 291 generators — with Flux Dev pulling it to 21%. Six independent signals are harder to evade and easier to audit.
The combined verdict is a posterior — but every contributing measurement is separately citeable. If you don’t trust the ensemble, read the residual yourself.
| Method | Description | Modalities | Reference |
|---|---|---|---|
Error Level Analysis ELA · recompression residual | Re-compresses the specimen at a known JPEG quality and reports per-block delta. Splices, edits, and fresh diffusion output produce signature patterns. | image · video | Krawetz 2007 ↗ |
Fourier Spectrum FFT · log-magnitude · DC-centered | 2-D DFT magnitude. Diffusion and GAN samplers leak periodic structure invisible to the eye, loud in frequency space. | image · video | Corvi 2023 ↗ |
Channel Decomposition RGB / YCbCr histogram divergence | Splits into per-channel grayscale, compares against a Bayer-array baseline. AI generators leak channel statistics no real sensor produces. | image · video | McCloskey 2019 ↗ |
Noise Residual PRNU proxy · 5×5 high-pass | Sensor pattern-noise fingerprint. Real cameras stamp every pixel with a unique PRNU; AI either has none or shows a learned, repeating one. | image · audio · video | Lukáš 2006 ↗ |
Model Attention Grad-CAM · ConvNeXt-V2 saliency | A learned detector trained on frontier samples returns its attention regions. Surfaces where the verdict came from inside the model. | image · video | Selvaraju 2017 ↗ |
C2PA Provenance Content Credentials 2.1 manifest | Reads the manifest, verifies the signature chain. When present, this beats every forensic signal. | all | C2PA 2024 ↗ |
Burstiness · Perplexity text only · entropy variance | Sentence-length variance + per-token surprise heatmap. Human writing is bursty; LLM output is fluently flat. | text | Gehrmann 2019 ↗ |
Watermark Scans Kirchenbauer · DCT · cepstral | Three watermark detectors — green-list n-gram bias for text, mid-band DCT for image, cepstral echo-hiding for audio. | text · image · audio | Kirchenbauer 2023 ↗ |
Every method traces to a published paper or open specification. No mystery models, no secret thresholds.
Foundational ELA writeup; underpins the q=0.90 re-encoding pass.
Off-DC lobes as the canonical diffusion fingerprint.
Channel-divergence baseline for RGB decomposition.
The original PRNU sensor-fingerprint paper.
Saliency overlay we render on the learned-detector pass.
Provenance side of every verdict; trust roots are open.
AUC 0.93 → 0.62 collapse cited in robustness panel.
Calibration corpus; ground truth for the 75.0% accuracy ceiling.
Green-list n-gram bias powering the text watermark scan.
Cepstral echo-hiding test used in audio.
DCT mid-band scheme used in image.
EER 3.3% → 10–18% generalization drop, referenced in audio thresholds.
A working map of deepfake video detection in May 2026: the new threat surface (Sora 2, Veo 3, Runway Gen-4, Kling 2, HeyGen, Hedra), the benchmarks that matter (SAFE, RobustSora, DeepfakeEval-2024), and the three-layer detection stack that has the best chance of surviving the year.
13 min read
What actually works for detecting AI-generated speech and music in 2026: ASVspoof 5 winners, Speech DF Arena rankings, the AudioSeal and SynthID watermark stack, why telephony codecs break detectors, and why Suno songs need a different model from ElevenLabs voices.
12 min read
What actually works for detecting AI-generated images in 2026: a survey of classical forensics (ELA, FFT, PRNU), modern learned detectors (CO-SPY, AIDE, FatFormer, DRCT), provenance via C2PA, and the benchmark numbers that should change how you read any vendor claim.
14 min read
The full forensic instrument — every signal, every visualization, the C2PA reader — is free to try. Pro unlocks unlimited verification, watermark-free shareable reports, JSON export, and API access.
Anonymous preview at 3/day, signed-in free tier at 10/day. No card.
Unlimited verification, watermark-free reports, JSON export, REST API.
One-time payment, never billed again. Locks in early-supporter pricing.
Running a moderation pipeline or newsroom? Enterprise options →
No signup required. No upload by default. Just open the instrument and look at the residuals.