The signal is part of the content
An invisible watermark is designed to be detected by a matching system while remaining hard for a person to notice. In an image or video it may be embedded in pixels or frames; in audio it may be embedded in the signal. This differs from metadata, which is stored in fields or container structures around the media data.
Google DeepMind's SynthID description is a concrete example: it describes imperceptible watermarks in generated images, video, audio and text. For video, Google says the signal is placed in generated segments; its technical explanation describes watermarking frames. SynthID detection depends on Google's technology and coverage. It is not equivalent to opening an MP4 file and reading its metadata boxes.
Can you see or remove it with a metadata tool?
No. Free AI Wiper reads supported portable metadata and specific hidden Unicode characters. It does not decode images or video frames to detect vendor watermarks, and it does not edit pixels, audio samples or statistical patterns. Removing a C2PA record from a file should not be described as removing an embedded watermark.
A watermark can also be absent, weakened by editing, or unreadable to a tool that lacks the matching detector. Therefore neither a clean metadata report nor a failed watermark check proves human authorship. Read metadata versus watermarks for a comparison, or how video detection works for the full set of video signals.