AI watermarking embeds hidden markers or signatures into content produced by AI systems — images, text, audio, or video — so that the origin of that content can be verified later. Unlike a visible logo or caption, the watermark survives typical post-processing steps such as compression, cropping, or reformatting, and can be detected by a corresponding decoder without being obvious to a human viewer or reader. The technique serves two distinct goals: attribution (tracing output back to a specific model or organisation) and authenticity (confirming that content has not been manipulated after generation).

Think of it like a serial number stamped invisibly into a banknote. You cannot see it in normal use, but a UV scanner in a bank instantly reveals whether the note is genuine and where it was printed.

In practice, text watermarking often works by slightly biasing the model’s token sampling so that a statistically detectable pattern is baked into the word choices. Image watermarking uses frequency-domain perturbations — the pixel values are adjusted in ways invisible to the eye but readable by a detector. For example, Google’s SynthID system embeds a watermark at generation time into images produced by Imagen, then provides a separate API endpoint to query whether a given image carries that mark.

Watermarking intersects closely with Hallucination mitigation (it cannot stop a model from fabricating facts, but it can prove which model produced the output), with RLHF (reward signals can be conditioned on watermark-preserving behaviour), and with Diffusion and Multimodal systems where provenance tracking is increasingly required by regulators. The effectiveness of any scheme depends heavily on how robust the watermark is against adversarial removal.

You will encounter AI watermarking most often in policy discussions around deepfake legislation, in enterprise content-authenticity pipelines, and in model providers’ API documentation where provenance features are advertised to media and legal teams.