AI-Generated Image Detection

Is this image real — or did a model make it?

Plica detects AI-generated and edited images by reading the layers a generator can't fully control: pixel microstructure, GAN fingerprints, and the metadata trail. One verdict, with the reasons behind it.

Try the API — free → PDF & KYC documents → See use cases
No cryptographic content credentials in a file? Then pixel and metadata findings are forensic review signals — strong grounds for suspicion, not mathematical proof. We say which is which.
Our angle

A detector alone breaks under scrutiny.

Many tools score the pixels and stop there — and proudly need "no metadata." But a probabilistic score with no provenance is exactly what falls apart when a result is challenged. Plica reads the pixels and the provenance, and tells you which signal is which.

Detector-only
"This image is 87% likely AI." A single number. No trail of why, no record of what was examined, nothing to defend when someone disputes it.
Plica
Pixel signals + metadata + provenance, merged into one verdict with named reason codes — and an honest line on what's proof and what's a review signal.
What we read

The layers a generator leaves behind.

A fraudster controls the layer you look at — the visible image. The layers beneath are harder to fake cleanly. That's where Plica looks.

GAN & pixel forensics
Structural GAN fingerprint clusters and intrinsic pixel-microstructure analysis — the statistical traces generative models leave in the raster.
GAN fingerprint · intrinsic forensics · moiré · quantization
Metadata & provenance
EXIF and container fingerprints: camera anchors present or stripped, sterile profiles, cloned metadata, third-party software traces.
EXIF validation · sterile profile · clone signals · software traces
Capture & pipeline signatures
Scanner-pipeline detection, monitor-screenshot signatures, and re-encode artifacts that reveal how an image actually reached you.
scan-pipeline · screenshot · re-encode · JFIF
Content credentials
When present, C2PA and SynthID are read as cryptographic signals. When absent, we say so — and fall back to forensic review, not false certainty.
C2PA · SynthID
Layer-conflict logic
When an authentic container holds an elevated-risk raster, we surface the conflict and route to review — instead of forcing a false clean verdict.
LAYER_CONFLICT_EMBEDDED · HITL
Merged verdict
All signals merge into one outcome with a single source of truth — not six numbers that contradict each other.
one verdict · reason codes
The verdict

Three outcomes, not a mystery number.

🟢 Authentic
No anomalies across the layers. Verdict and reason codes recorded.
🟡 Review / HITL
Ambiguous or conflicting signals. Routed to human review, not auto-rejected.
🔴 AI-generated / Manipulated
Strong synthetic signals — GAN clusters, screenshots, cloned EXIF. Flagged for rejection or escalation.
What we don't claim

Honest about the boundary.

Without cryptographic content credentials in a given file, pixel and metadata findings are review signals — not proof of authenticity or synthesis. We label which is which, every time.

When an image is degraded — heavily recompressed, passed through messengers — forensic signal weakens. We say the signal is weak rather than pass a guess off as a finding.

Plica handles images and PDF documents. It is an independent layer that plugs into your KYC stack and works alongside Sumsub, Onfido, Veriff — it does not replace them.

Check an image now.

Free trial API · images + PDF · no integration required to start

Try the API — free →