Veryfi Blogpost: AI_Generated Receipts
AI-generated document detection is an integrated component of the Veryfi Fraud Suite that identifies documents created using AI tools. The system employs advanced detection methods to flag potentially fraudulent AI-generated documents and returns a "Generated Document" decision in the Fraud Type when suspicious patterns are detected.
Detection Methods
Veryfi uses two complementary detection approaches that work together to provide comprehensive AI-generated content identification:
1. Image Artifact Analysis
The primary detection method uses sophisticated computer vision models that perform pixel-level analysis to identify patterns typically associated with AI-generated images. This approach:
Analyzes image artifacts and pixel patterns characteristic of AI generation
Employs multiple specialized models for enhanced accuracy
Functions independently of file metadata
Provides confidence scores for detection results
2. Metadata Analysis
AI-generated images can also contain distinctive metadata signatures that can indicate artificial creation. Veryfi system also examines file metadata and EXIF data embedded within documents, including:
EXIF data, Document metadata, Technical metadata
☝🏼 Fraud type: Generated Document mechanics update
New logic: Generated Document =
meta.pages.ai_generated+EXIF data
No changes in the fraud_type: Generated Document that flags AI-Generated images - please continue relying on its response. However, when debugging, you might want to update your internal workflow referencing a new field that powers Generated Document Fraud type.
Old: Check
meta.pages.screenshot.type: ai_generated+scoreNew: Check
meta.pages.ai_generated+score
Also, some early adopters might have built integration based on meta.pages.screenshot.type: ai_generatedflag, in this case we you need to switch to the most reliable and robust flag - Fraud_type: Generated Document, that consumed the prediction from both vision model and EXIF checker.
How Results Are Delivered
Fraud Integration
When Veryfi Fraud Suite is enabled, AI-generated documents are flagged through:
Fraud Type: Returns "Generated Document"
Fraud Color Coding: Red color indicates detection threshold is exceeded
Dedicated Field
Results for the vision model also appear in the meta.pages.ai_generated field with possible values: true / false + score
📚 Read more about Artifact Analyses
Limitations & Considerations
Metadata Vulnerabilities
Metadata can be easily removed or altered
File conversion, screenshots, or format changes often eliminate metadata evidence
The vision model serves as a backup detection when metadata is unavailable
Physical Document Challenges
Screen Photography: Taking photos of screens displaying AI-generated content significantly reduces detection accuracy
Print-and-Scan: AI-generated documents that are printed and then re-digitized present detection challenges
PDF Limitations
AI-generated PDFs created programmatically (rather than as image files) may not be detected by the vision model or Exif data checker, as these are rendered documents rather than generated images. To improve coverage of PDF documents, we are currently working on Layout Analysis.
Best Practices
Make sure your implementation combines fraud_type response rather than single signal.
Adjust thresholds based on your risk tolerance and false positive acceptance, either via Business Rules or in your business validation logic, if you handle it on your side
Monitor detection patterns to identify potential fraud trends
Consider the limitations when evaluating edge cases
Report to us new undetected cases, and we will work together on their coverage
This detection system provides a robust foundation for identifying AI-generated fraudulent images while maintaining operational efficiency and minimizing false positives.
How can I help improve Veryfi's fraud detection accuracy?
You can establish a feedback loop by sending fraud decision corrections back to Veryfi using a PUT API call with the fraud_review object. This gives Veryfi team valuable insights into fraudulent and non-fraudulent submissions, helping us continuously improve our models.
How does this work in practice?
If your workflow includes sending warnings or flags to your end users (giving them the option to approve or reject a document), you can pass those decisions back to Veryfi.
For example:
User submits a document flagged as potentially fraudulent
You present a warning and ask them to confirm or dispute the flag
User makes their decision (approve/reject)
You send that feedback to Veryfi via the
fraud_reviewobject with just a decision or Fraud Type. API docs link for reference
Decision : [
fraud,not fraud,unknown]Types: [
other,handwritten characters,digital tampering,generated document,ai generated,LCD photo,screenshot,not a document,duplicate,high velocity,fraudulent pdf,invalid qr data]
This closed-loop feedback helps Veryfi models learn from real-world decisions and improve detection accuracy over time.
Please note that Fraud Suit (including AI-generated detection) is an add-on feature. Please reach out to [email protected] if you would like to access this feature.
How can I test Veryfi's AI-generated receipt detection?
Try uploading a suspicious receipt here and Vee, our mascot, will let you know if it’s fake or not. If you already have an account with us, please contact us at [email protected] and request it on your account through our Customer Support team.