- Your deepfake detector waves through the oldest fraud in insurance: a real photo, lifted from a settled claim, resubmitted as fresh damage. It passes every authenticity test because it genuinely is real. We built a forensic gate that catches it anyway 🧵
- A lone authenticity detector asks one thing: was this image generated or manipulated? A recycled real photo answers no, honestly. Clean EXIF, real sensor noise, no artifacts. The detector clears it, and the same damage quietly gets paid twice.
- So we stopped leaning on one signal. At ingest we compute a 256-bit perceptual hash of the photo and run a nearest-neighbour search over a historical claims index. A recycled image lights up as a near-match to a claim that already settled.
- Perceptual-hash reuse is one stream. A second reads the FNOL narrative against the picture: the account says rear-end at night, the photo shows front damage in daylight. Two independent signals, one from provenance and one from language, landing on the same claim.
- In the demo, claim CLM-C-1044 is exactly this. Provenance passes. Our simulated pixel-forensics signal passes. It is a real camera photo. But the reuse hash matches a settled prior claim and the narrative contradicts the image, so it routes to SIU-REFERRAL.
- The part that matters to a regulator: the analysts only advise. A deterministic policy gate outside the LLM aggregates their verdicts and the raw signals into one routing call. Same inputs, same decision, every run. Never a model grading its own work.
- On a fixed, labeled 120-claim synthetic benchmark, 35 of the frauds carry no authenticity artifact at all (recycled-real plus metadata-spoofed). A single-signal baseline catches 0 of those 35. The cross-signal gate catches all 35. Same photos a detector alone clears.
- That set is constructed, so those figures describe it, not the open world. What holds regardless: on it, 0% of clean claims were wrongly referred, and every decision writes to a hash-chained record you can re-verify. An adversarial Challenger argues to clear each flag first.
- Genuine question for fraud and claims teams: when a resubmitted photo is genuinely real, provenance clean and pixel-forensics clean, what are you using to catch it today? Perceptual-hash reuse, narrative checks, something else? #InsurTech #FraudDetection #AIGovernance
- We built this as a demo, not a deployed pipeline: the connectors are stubbed and every claim is synthetic. If your team is rethinking evidence authentication before damage assessment, we'd like to hear how you draw the line. https://veriprajna.com/demos/insurance-claims-ai
Published on X · August 7, 2026
On social media
See this post on its original platform
In our archive