Summary
In this Computerphile episode, computer vision researcher Mike Pound explains how watermarks are embedded in images, building on an earlier discussion of watermarking AI-generated text. He starts by contrasting watermarking with steganography: where steganography aims to hide a message and maximize payload, watermarking aims to prove who or what process created an image.
The video walks through why simple approaches fail. Metadata is easily stripped, especially when an image is converted to JPEG and uploaded to the web, and hiding data in the least significant bits of pixel values is trivial to detect and overwrite. Pound then explains the modern approach: an encoder-decoder neural network that embeds a watermark and is trained against thousands of random augmentations such as crops, flips, color changes, resizing, deleted regions, and rotations, so the signature survives real-world editing.
He also touches on technical details from the TrustMark paper and later implementations, including how to keep watermarks detectable when an image changes size. It is a clear, approachable introduction for anyone curious about provenance and copyright protection for images, including AI-generated content.
📺 Source: Computerphile · Published October 22, 2026
🏷️ Format: Deep Dive
