Digital Watermarking

Digital watermarking embeds a visible or imperceptible signal into an image or video so that authorised tools can later detect it. For AI-generated visuals, the watermark can provide machine-readable evidence of provenance, such as association with a particular generation system, while remaining part of the visual data itself.

How authentication works

A watermarking system combines an embedding process with a corresponding detector. Authentication depends on whether the detector can reliably recover or recognise the expected signal.

• During or after generation, an embedding algorithm modifies selected pixels or frequency components according to a predefined pattern, sometimes controlled by a secret key.
• The watermark may encode a content identifier or indicate that the visual originated from a participating AI system.
• A robust watermark is designed to remain detectable after common operations such as resizing, compression or limited cropping.
• Verification produces evidence or a confidence score, rather than an infallible declaration of authenticity.

What it can establish

• An imperceptible watermark enables automated labelling and screening of large volumes of synthetic visual content without visibly obstructing the image.
• A fragile watermark is designed to be disrupted by modification, helping reveal or localise tampering.
• A valid watermark can support a claim about the content's source or processing history, provided the embedding system and verification keys remain trustworthy.

Limits and complementary safeguards

Watermarking authenticates an embedded claim, not the truth of what an image depicts. It is therefore one component of content authentication, not a complete solution to deepfakes.

• Cropping, heavy editing, adversarial manipulation or format conversion may weaken or remove a watermark, while poor detector design may produce false positives or false negatives.
• Absence of a detectable watermark does not prove that content is human-made because many AI systems may not apply compatible watermarks.
• Compromise of embedding keys or tools can permit forgery, making key management and access control essential.
• Watermarks work best alongside cryptographic signatures, secure provenance metadata, forensic analysis and platform disclosure rules.