Generative images: copyright, deepfakes, watermarking

Image generation raises four questions that are all real and partly unresolved:

Question	The core
Training data	models trained on artists' works, often without permission
Style imitation	«in the style of [a living artist]» — style is not protected, but the use is contested
Deepfakes	images of real people in invented situations
Non-consensual material	by far the most common harmful use

The last one is not an edge case: studies of synthetic material online have shown that the overwhelmingly dominant part is non-consensual intimate material of women. That is the actual principal harm, and it belongs first — not last.

The technical countermeasures and what they actually manage:

Countermeasure	Gives	Does not give
A visible mark	clarity for the attentive	cropped away in seconds
An invisible watermark (SynthID)	survives some editing	can be removed by anyone who wants to
C2PA / Content Credentials	cryptographic proof of origin	proves nothing about images without a signature
Detectors	a rough indication	a high error rate; getting worse as the generators improve
Protection tools (Glaze, Nightshade)	makes copying the style of individual works harder	broken by new training methods

The uncomfortable conclusion: technology can establish authenticity but not prove falsity. The absence of a signature means nothing, since the vast majority of genuine images lack one too.

So the emphasis moves to provenance and responsibility: who published it, for what purpose, and what do independent sources say? That is source criticism, not image analysis.

The rules are moving: the EU AI Act requires synthetic content to be labelled and deepfakes to be labelled particularly clearly. Several countries have introduced criminal provisions against non-consensual synthetic images.

A policy of your own — four lines that go a long way:

1. Do not generate images of identifiable real people without consent.
2. Do not imitate a named living artist's style commercially.
3. Label generated material where it could be mistaken for documentary.
4. Check the licence of the generator you use — the terms differ.