AI Provenance and Training Mining Guide

A guide for AI agents auditing media files for generative AI origin, model attribution, and machine learning training consent.

---

1. IPTC Digital Source Types

C2PA leverages standard IPTC Digital Source Type URIs within actions and source assertions to differentiate synthetic content from human capture:

Digital Source Type URI	Meaning	Provenance Category
.../digitalsourcetype/digitalCapture	Captured directly by a physical camera, microphone, or sensor	Human / Camera
.../digitalsourcetype/trainedAlgorithmicMedia	Created entirely or primarily using generative AI algorithms	Synthetic / Generative AI
.../digitalsourcetype/compositeWithTrainedAlgorithmicMedia	Human media edited or composited with generative AI elements	Hybrid / AI Edited
.../digitalsourcetype/algorithmicMedia	Generated by non-AI algorithmic procedures (e.g. procedural shaders)	Algorithmic
.../digitalsourcetype/virtualRecording	Recorded inside a synthetic environment (e.g. 3D engine, video game)	Virtual Simulation

The full namespace prefix for these URIs is: https://cv.iptc.org/newscodes/digitalsourcetype/

---

2. Machine Learning Training and Data Mining Assertions

Publishers use the c2pa.training-mining assertion to signal automated scrapers and AI developers whether media content may be used for model training.

Assertion Structure

{
  "c2pa.training-mining": {
    "entries": {
      "c2pa.ai_generative_training": { "use": "notAllowed" },
      "cawg.data_mining": { "use": "notAllowed" },
      "c2pa.ai_inference": { "use": "allowed" },
      "c2pa.ai_training": { "use": "notAllowed" }
    }
  }
}

Standard Permission Keys

1. c2pa.ai_generative_training: Governs use for training generative models (e.g. image, video, voice synthesis).
2. c2pa.ai_training: Broad prohibition across all machine learning training pipelines.
3. c2pa.ai_inference: Governs using the content as input for real-time model inference (e.g. visual question answering).
4. cawg.data_mining: Governs automated text and data extraction under copyright mining exceptions.

Values are either "allowed", "notAllowed", or "constrained".

---

3. Generative Model Attribution

When generative models sign output content, they attach model parameters and creation metadata in the c2pa.generative assertion:

{
  "c2pa.generative": {
    "model": {
      "name": "Imagen",
      "version": "3.0"
    },
    "prompt": "A serene mountain lake at sunrise",
    "seed": 429184
  }
}

The inspect_c2pa.py script automatically parses these keys and extracts the model name, version, and prompt for quick review.