A Guide to Watermarking Technology for AI-Generated Content

Watermarking is a technology that adds identifiable patterns to AI-generated content to convey provenance information.

• Visible watermarks: A method like Dall-E 2, which leaves an explicit mark in the corner of an image.
• Invisible watermarks: A method that embeds patterns invisible to the human eye but detectable by algorithms.
• Implementation methods: There are methods that embed watermarks directly during the model's generation process to increase durability, and post-processing methods that can be applied to outputs from closed models after generation.

In addition to watermarking, the following supplementary technologies are also used.

• Data Poisoning: Technologies like Glaze and Photoguard, which disrupt AI algorithm processing, or Nightshade and Fawkes, which break the assumptions of AI training models to make deepfake generation difficult.
• Signing: A method like Truepic, which follows the C2PA standard to combine and authenticate provenance information within metadata.

Regarding the scope of disclosure for these technologies (Open vs Closed), the core issue being discussed is the balance between promoting innovation and preventing misuse.