I generated fake insurance claims photos with GPT Image 2 and nobody I showed could tell they were AI

I've been working in fraud detection for P&C insurance and ran a small experiment last week that honestly shook me.

I used GPT Image 2 to generate 5 images. Smashed car window, a receipt, an invoice, injury photos, property damage. Then I showed them to 5 people — a mix of regular people and professionals who work in insurance and legal — and asked them to pick out which were real and which were AI.

Nobody could tell with confidence. Not the insurance people, not the legal people, not the regular folks. Everyone's confidence was low.

These weren't even sophisticated attempts. Just straight prompts. No editing, no post-processing. The raw output was good enough to fool everyone I tested, regardless of their background.

And that got me thinking about what this means for claims processing. Fraud has always been a thing in insurance. People have been staging accidents and faking injuries forever. But the barrier just dropped to basically zero. You don't need to stage anything anymore. You just need a free AI tool and 30 seconds.

A few stats that hit different after running this test:
– 20-30% of claims already include some form of altered media
– P&C fraud costs around $45-50B annually in the US
– Gen Z is 4x more likely than boomers to consider "small edits" to claims photos

Curious what others are seeing. If you work in claims or fraud, are you noticing more suspicious submissions? How are you thinking about detection when the fakes are this good?

Also interested if anyone's looked into purpose-built detection vs general purpose AI models for this. The non-deterministic nature of large LLMs seems like a real problem for security use cases where you need consistent, repeatable results.