Two questions, asked separately.
Image authenticity is not one problem. A frame can be wholly synthetic (generated by a model, never photographed), or it can be a genuine photograph with a manipulated face dropped into it. Those failure modes leave different traces, so Raven checks for them separately.
The first pass asks whether the image came from a generator at all. If it did, Raven reports the most likely generator and scores the alternatives by per-tool confidence.
The second pass asks whether a real photograph carries deepfake or face-swap artefacts. A low score there alongside a high synthetic score tells you the whole frame is fabricated rather than a real scene with an altered subject.
A frame that reads as real.
The image below is convincing at a glance: a person in tactical gear on a street, framed like a press photograph. Raven flags it as likely AI-generated.
The breakdown matters as much as the verdict. A 92% concentration on one generator with near-zero scores across named tools, and a 2% face-manipulation reading, describes a fully synthetic frame rather than a doctored photograph.
A verdict with its working shown.
- Verdict: a plain-language judgment, such as likely AI-generated.
- Closest match: the generator the signals most resemble.
- Per-tool confidence: the remaining generators, each scored, so you can see how concentrated the result is.
- Face manipulation score: a separate deepfake and face-swap reading.
Presenting the distribution rather than a single label is deliberate. A verdict resting on one narrow margin should be treated differently from one where the signal is overwhelming, and the interface makes that difference visible.
Check the frame before you trust the lead.
A synthetic image will still produce a plausible-looking location estimate. Nothing about Find Region or Find Street refuses to run on a fabricated scene; they will faithfully analyze a street that never existed.
Running Verify Image first is the cheap insurance: it stops an investigation from committing hours, and potentially an operational response, to a frame that was generated rather than captured.
Common questions.
What does it detect?
Whether an image was generated by an AI model rather than captured, and separately whether a real photograph carries deepfake or face-swap edits.
Can it tell which AI tool made the image?
It names the most likely generator and scores the rest by per-tool confidence, so the answer reads as a distribution rather than a single label.
Is it production ready?
No. It is a beta preview and accuracy is not guaranteed.
Why check authenticity before geolocation?
Because a synthetic image still yields a plausible location estimate. Verifying first avoids investigating a scene that was never photographed.
