Generated images, audio and text have crossed the line where people cannot reliably identify them.
Two different technical responses exist, and they answer two different questions.
Watermarking asks whether this artefact was generated, using only the artefact itself.
Provenance asks where this file came from, using metadata attached to it.
For images, a watermark is a perturbation spread across the pixels, invisible but statistically detectable.
Good schemes survive resizing, compression and cropping, because those happen in ordinary use.
For text the mechanism is cleverer, since there are no pixels to perturb.
At each step the vocabulary is split into a favoured and unfavoured list by a secret key.
Sampling is nudged toward the favoured list whenever several tokens are about equally good.
Over a few hundred tokens the resulting bias is measurable, while the writing reads normally.
Short passages carry too little signal to judge, which is a real limit for detection in practice.
Provenance standards such as C2PA sign a manifest recording the origin and edit history of a file.
That manifest travels with the file and can be verified cryptographically by anyone.
Its weakness is that stripping the metadata is trivial, so absence proves nothing at all.
Both approaches raise the cost of casual deception and neither survives a motivated adversary.