Security & Policy

Watermarking

Embedding a hidden, detectable signal in AI-generated text, images, or audio to mark it as machine-made.

AI watermarking subtly biases a model output in a pattern that is invisible to people but detectable by an algorithm, so content can later be identified as AI-generated. It is one proposed answer to deepfakes, misinformation, and academic dishonesty. The catch is robustness: watermarks can be weakened by editing or paraphrasing, and they only work if the generator cooperates, so they are a partial measure rather than a guarantee.

Reference: Wikipedia: Watermarking