AI provenance metadata for generated images (EU AI Act deepfake disclosure)

Context

Stable Diffusion and SDXL generate images that get shared and distributed. Under EU AI Act Article 50 (August 2, 2026), AI-generated images must carry transparency metadata and deepfake disclosure becomes mandatory.

Currently, generated images carry no provenance — no indication they're AI-generated, no model info, no generation parameters.

Possible approach

Embed provenance in image metadata (EXIF/PNG tEXt) during generation:

# After generating
image = pipe("a sunset").images[0]

# Embed provenance
from PIL.PngImagePlugin import PngInfo
pnginfo = PngInfo()
pnginfo.add_text("ai_generated", "true")
pnginfo.add_text("model", "stabilityai/stable-diffusion-xl-base-1.0")
pnginfo.add_text("generated_at", "2026-03-31T10:00:00Z")
image.save("output.png", pnginfo=pnginfo)

This uses Pillow (already a dependency), zero additional packages needed.

Why

• EU AI Act deepfake provisions target AI-generated images specifically
• Photos carry EXIF (camera, GPS). AI images carry nothing.
• Stability AI's open models are widely deployed — setting a provenance default would influence the ecosystem

Reference

• AKF embeds provenance into images via EXIF/XMP
• C2PA handles media provenance but requires PKI infrastructure
• A lightweight EXIF-based approach has broader adoption potential