Verifying AI-generated images at scale

When people ask how Verify detects AI images, the honest answer is: a stack of overlapping signals. No single detector survives the next-gen generator. So we run an ensemble — and the engineering story is mostly about how we keep the ensemble fast.

The four signal classes

1. Generative artefacts. Frequency-domain residuals, latent-space inconsistencies, and learned embedding anomalies.
2. Capture-time signals. EXIF, hardware attestation (C2PA), and physical-camera fingerprints.
3. Content plausibility. Light direction, shadow geometry, reflection consistency.
4. Provenance. Reverse image search, perceptual hash matches against known-synthetic seed sets.

We weight these per request based on observed confidence — a clean smartphone capture with valid C2PA never goes through the expensive ensemble.

Why naive ensembles fail

Most public benchmarks evaluate a single detector on a fixed test set. Production is different: adversaries adapt. We retrain quarterly on red-team outputs and publish our confusion matrices internally — the moment a class slips below 90% TPR, it goes back into the queue.