AI Watermarking

AI Watermarking embeds signals, often designed to be hidden and robust, into content produced or processed by Generative AI systems to help identify its association with a particular supported system. Detection is system-specific and imperfect: signals may be weakened or removed by edits, transformations, or platform processing, and their presence or absence does not reliably establish whether content is machine-generated in general. In Large Language Models, watermarking often applies a Statistical Constraint to Token selection, such as favoring a keyed Green-list of tokens, creating a pattern detectable by a compatible detector while aiming to minimize effects on output quality and Perplexity. In Computer Vision and image generation, techniques such as Steganography or Latent Space manipulation can embed visually imperceptible signals. Some implementations withstand particular transformations, but resistance depends on the method and transformation severity; cropping or compression can reduce or eliminate detectability. OpenAI, Google DeepMind, and Meta have researched or deployed watermarking and Content Provenance methods to support transparency, AI Safety, and identification of Deepfakes. These methods can also support regulatory compliance: the EU AI Act requires machine-readable marking and detectability of certain AI-generated or manipulated content, but does not mandate watermarking specifically; other technical approaches can include Metadata and Cryptographic Provenance, used individually or in combination as appropriate.