Visual fraud glossary. The terms, explained in a few lines.
The words that come up when talking about fake photos used as evidence: what kinds of images exist, how they are detected and what forms the fraud takes.
Vouch teamOctober 23, 20266-minute read
Types of fake images
Visual fraud
Using a fake or misleading image as evidence to obtain a payment, a refund or a benefit. More in what visual fraud is.
AI-generated image
An image created from scratch by a model from a text prompt. There is no original photo behind it.
AI-edited image
A real photo where an AI tool adds, removes or changes something, like a dent, a stain or an object.
Shallowfake
A real photo or video with a simple, targeted edit: a retouch, a crop or a change of context. More in what a shallowfake is.
Cheapfake
Synonym for shallowfake. The term became popular in 2019 to name manipulations made with cheap software or no software at all.1
Deepfake
An image, audio or video generated or altered with AI to create something that never existed, like a face or a scene.
Recycled photo
A real image taken from the internet, social media, a sales listing or another claim, and presented as proof of a new event.
Staged scene
A situation set up for the photo, like staged damage or a toy model photographed as if it were a real object.
False context
A real photo presented as if it were from another day, another place or another order.
How it is detected
Metadata
The data stored inside a photo file: date, device, location and editing software. Many social networks and messaging apps strip it,2 which is why some platforms require photos with date, time and location.3
EXIF
The most common metadata format in photos from cameras and phones.
C2PA
An open standard, created in 2021, that signs an image with information about how it was created and edited.4 More in what C2PA is.
Content Credentials
The public-facing name of the C2PA standard. They are checked with the Verify tool.5
Invisible watermark
A hidden signal in the pixels that identifies an image as AI-generated. Google’s is called SynthID and only recognizes content made with its tools.6
Reverse image search
Searching for a photo using the image itself, with tools like Google Lens or TinEye, to find out whether it existed before.
Error level analysis (ELA)
A technique that compares compression across areas of an image to flag possible edits. It fails with photos compressed several times.7
AI image detector
A model trained to estimate whether an image was AI-generated. It loses accuracy with compressed photos and with generators it doesn’t know.
Risk score
A score that sums up how suspicious a photo is, for example from 0 to 100, ideally with the reason.
Human review
The step where a specialist reviews doubtful cases with the evidence organized and decides.
Step by step, in the guide to telling whether a photo was generated or edited with AI.
Types of fraud
Refund abuse
Requesting a refund with a false or exaggerated claim. It is the most widespread type of fraud in online commerce.8
Policy abuse
Taking advantage of a company’s return, warranty or compensation rules to get a benefit you’re not entitled to.
First-party fraud
Fraud committed by the customer, under their real identity, rather than by a third party who steals an account.
“Item not received”
Claiming an order never arrived when it did. It is the most common form of refund abuse.
Wardrobing
Buying a product, using it and returning it as if it were new.9
Refund-as-a-service
Groups that file fraudulent refunds on behalf of others in exchange for a percentage.10
Fake proof of delivery
A generated or reused photo used to make it look like an order was delivered.11
Multi-accounting
One person opening several accounts to multiply refunds or promotions.
Collusion
Two parties, like a courier and a customer, agreeing to report a problem that never happened.
Operations
Self-inspection
When the customer takes and uploads photos of the item, for example their car after an accident, instead of an in-person inspection.
Instant refund
Returning the money before the product is received or checked, with the photo as the only evidence.
Chargeback
A payment reversal ordered by the bank or card network when a customer disputes a purchase.
How Vouch solves it
Vouch is a visual fraud detection platform. It analyzes every photo in under 30 seconds with three layers: detection of images generated or edited with AI, a search for the image online and in previous claims, and the customer’s context. It returns a 0–100 risk score with the reason, and gray-area cases go to a specialist. It works from Vouch Web, with no integration, or via API.
Frequently asked questions
What is the difference between a shallowfake and a deepfake?
A deepfake is generated or altered with AI to create something that never existed. A shallowfake starts from a real photo and applies a targeted edit, like adding damage or changing a number plate.
What is a photo’s metadata?
The data stored inside the file: date, device, location and the software that edited it. It helps verify a photo, but many messaging apps and social networks strip it, and it can be modified.
What is refund abuse?
Requesting a refund with a false or exaggerated claim: that the order didn’t arrive, arrived damaged or was wrong. It is the most widespread type of fraud in online commerce.
Sources
This glossary is based on public sources only; every figure can be verified.
- Data & Society. Deepfakes and Cheap Fakes: The Manipulation of Audio and Visual Evidence (Sep 2019)
- IPTC. IPTC study shows some social media networks remove rights information from photos (Mar 2013)
- Turo. Reporting damage | Hosts
- C2PA. Technology and media entities join forces to create standards group aimed at building trust in online content (Feb 2021)
- Content Authenticity Initiative. Using the Verify tool
- Google. SynthID Detector — a new portal to help identify AI-generated content (May 2025)
- 29a.ch. Image Error Level Analysis with HTML5 (Apr 2012)
- Merchant Risk Council. 2026 Global eCommerce Payments & Fraud Report (Mar 2026)
- Retail Rewired. Forter: Retailers report rising returns fraud, with 80% of fake claims fuelled by AI evidence (Jun 2026)
- Sourcing Journal (WWD). Amazon Lawsuit Names 27 Defendants Running Global Refund Fraud Scheme (Dec 2023)
- TechCrunch. DoorDash says it banned driver who seemingly faked a delivery using AI (Jan 2026)