How does AI image detection work?
We are not a magic mirror. Findings are graded by evidence strength: signed strong evidence (C2PA / watermarks) always wins; the classifier is only a probabilistic reference. Verdicts come in three states (A: provenance found / B: classifier lean / C: inconclusive) — never "100% AI" and never "verified real". A negative result does not prove an image is real: screenshots and re-encoding strip every active signal.
Want to just run a check? Try the free AI image detector →
C2PA digital signature Free
Cryptographically signed provenance (verify + trust list). A hit yields a high-confidence verdict naming the tool. Stripped by screenshots and platform re-encoding.
Invisible watermark (DWT-DCT) Free
Default watermark of the open-source diffusers pipeline. Note: A1111 / Forge / ComfyUI never embed it — a negative tells nothing.
EXIF metadata scan Free
Generator software fields (e.g. "Made with Google AI"). Fields can be forged or removed.
PNG text chunks Free
A1111 parameters, ComfyUI workflow JSON, NovelAI markers. Can be turned off when saving.
IPTC declarations Free
Standardized DigitalSourceType / trainedAlgorithmicMedia statements (newer Midjourney exports, OpenAI…).
Forensic signals Free
FFT spectrum, noise residual, compression traces. Explanatory only: produces notes like "heavily recompressed — reliability degraded".
AI classifier (2nd opinion) Within quota
Sightengine genai probability score. 0.4–0.6 is an abstain zone; reliability drops on unseen generators and recompressed images (industry tests show up to 13% false positives).
Known limitations (the honest part)
- VCT² benchmark (2025, 17 detectors / 166,000 images): zero-shot average accuracy ≈ 58%.
- Commercial detectors falsely flag real news photos ~13% of the time (NewsGuard audit, 2026-05).
- On "real photo with local AI inpainting", mainstream classifiers drop from >91% to ~55% (INP-X, 2026).
- Diffusion regeneration attacks can strip watermarks and fool classifiers simultaneously; watermarks can also be transplanted to forge positives.
- Detection results must never be the sole basis for judicial, insurance, academic, or account-ban decisions.
“Is this image AI — and why?”
The “why” is the product. Every result page replays the six steps with their findings (e.g. “C2PA manifest signed by OpenAI”, “Stable Diffusion parameters chunk found”), marks short-circuited steps, and lists the known failure modes next to the verdict — so you can judge the evidence yourself instead of trusting a black-box score. See it live on the detector home page.