Modern watermarks are statistical rather than visual: a pattern spread through the pixels or the audio that survives compression, cropping and re-encoding, and that a detector can read even when a human cannot see it.
They are one layer, not the answer. A determined actor can degrade a watermark; a provenance manifest can be stripped; a visible label can be cropped. Anyone serious runs all three and assumes each will fail sometimes.
QUESTIONS PEOPLE ASK
Can you detect AI-generated images?
Reliably when the generator embedded a watermark and you have the matching detector. Unreliably otherwise: general-purpose "AI detectors" produce enough false positives to be unsafe as evidence.