- A fraudster photographs an undamaged car. A diffusion model adds a smashed bumper — correct lighting, shadows, reflections. Your claims AI confirms the damage, scores severity, generates an estimate. It pays out. This isn't hypothetical. It's 2026. 🧵
- The problem is structural. Damage-assessment AI evaluates content — what does the damage look like? It was never built to ask authenticity — was this damage physically present when the camera captured the photo? Two different questions. Most stacks answer only the first.
- The scale stopped being marginal. US AI-enhanced fraud cases went from under 20,000 in 2022 to over 80,000 in 2025. In April 2025, UK motor carriers disclosed fraudsters using diffusion models to inject damage — inflating average payouts by ~£13,000 per incident.
- And consumers are willing. Verisk's March 2026 study: 36% say they'd alter a claim image. Among Gen Z it's 55%. Of those who've already tried, 44% called their results "very realistic."
- Insurers know they're exposed. Only 32% feel "very confident" they can detect a deepfake. 76% say manipulated submissions have grown more sophisticated. The detection gap is wide and widening.
- But the threat most carriers never audited is their own pipeline. If your app "enhances" uploaded photos — upscaling, denoising — a GenAI model can read a real dent as noise and smooth it. The adjuster now sees less damage than the policyholder actually submitted.
- Under US law, that's spoliation. The test isn't intent — it's whether synthetic pixels (pixels the camera sensor never captured) entered the evidence. Overwrite the original on a claim that later goes to litigation, and you're facing sanctions or adverse inference.
- Now stack the regulation. The NAIC Model Bulletin (24 states) demands explainable claim decisions. The EU AI Act classes insurance AI as high-risk — Aug 2026 deadline, penalties up to €35M or 7% of global turnover. A black-box score can't produce the explanation either requires.
- The market has strong players, but real gaps. Tractable: 95% damage accuracy, no deepfake detection, no chain-of-custody, SaaS-only — you don't own the model. VAARHAFT does fraud detection but no damage assessment. Detection and assessment sit in separate products.
- Our take: authenticity can't be a bolt-on after assessment. It belongs in the pipeline — forensic computer vision that authenticates, measures, and preserves every pixel, with a segmentation mask and chain-of-custody an adjuster (and a court) can actually inspect.
- If your photo pipeline runs any GenAI "enhancement" before an adjuster sees the image — have you audited it for spoliation exposure? Most carriers haven't. Where are you drawing the line on enhancement before adjuster review? #InsurTech
- We wrote up the full two-threat model — synthetic fraud and evidence spoliation — and what a forensic claims-CV stack actually needs to handle in 2026: https://veriprajna.com/solutions/insurance-claims-ai