Image generators have crossed the threshold where casual inspection fails. But "impossible to spot casually" is not "impossible to spot" — a structured check still catches most fakes, and matters for finance teams more than most people realize: manipulated images now appear in invoice fraud (doctored invoices, fake remittance letters), fake ID documents in onboarding, and fabricated "proof" photos in supplier disputes.
The 8 checks, in order
- Provenance beats forensics. Ask where the image came from before analyzing pixels. Reverse-image search it; an image that only exists in one place, sent by the person who benefits from it, starts suspicious.
- Text and numbers. Generators still mangle small text — letterheads, invoice line items, serial numbers. Zoom into every character.
- Geometry and reflections. Check that shadows agree on a light source, reflections match the scene, and straight lines (table edges, door frames) stay straight through the background.
- Hands, ears, teeth, jewelry. Still the classic failure zones in generated people — asymmetric earrings, fused fingers, irregular teeth.
- Texture uniformity. AI images often have a subtle over-smoothness — skin without pores, fabric without weave — or the opposite, hallucinated detail that repeats in tiles.
- Metadata (EXIF). A real phone photo carries camera model, timestamps, and often GPS. Absent metadata proves nothing (platforms strip it), but present, coherent metadata is a positive signal.
- Content credentials (C2PA). A growing share of cameras and editors cryptographically sign capture and edit history. Verify credentials when present — this is the direction the industry is moving.
- Run a detector — and treat it as one vote. Automated AI image detectors analyze statistical fingerprints invisible to the eye. They're genuinely useful and genuinely fallible: compression, resizing, and novel generators degrade accuracy. Use the score to decide how much human scrutiny to apply, not as a verdict.
For finance teams: the process answer
The same principle that governs voice applies to images: detection triggers verification, it doesn't replace it. A suspicious invoice image shouldn't start a forensics debate — it should trigger the callback protocol: verify the request with the counterparty on a number already on file. Pixels can be argued about; a verified callback can't.