Labelling Generated Content: Why and When

Labelling generated images has moved from ethics into practice: platforms are deploying mechanisms automatically, and regulation in various countries is gradually taking shape. It helps to understand how it works and what to do yourself.

Why it is needed

The reason is not that generation is bad but that viewers make decisions based on what they see. An image that looks like a photograph of an event reads as evidence. If the event did not happen, that misleads — regardless of the author's intent.

Labelling restores the match between what an image is and what it appears to be.

Technical mechanisms

1. File metadata: the service records the image's origin in service fields. Simple, but easily lost on re-saving.
2. Cryptographic provenance signatures: a more robust option being developed by industry initiatives. It allows an image's history to be traced.
3. Invisible watermarks: a compression-resistant mark inside the image itself.
4. Automatic detection by the platform: analysing the image for signs of generation.
5. An explicit interface label: a "made with AI" note beside the post.

Platform requirements

They differ and change. The general tendency: large social platforms require generated content to be declared at upload and add a label automatically when they detect signs of it.

The practical consequence: a label will appear either way, and it is better to set it yourself than to receive it automatically — in the second case a platform may reduce reach for what looks like concealment.

A practical criterion

Formal requirements do not cover every situation, and for the rest a rule helps: label wherever a viewer might take the image as documentary.

• An overtly stylised illustration — no label needed, it is obvious.
• A photograph-like portrait in a personal profile — a borderline case.
• An image presented as a picture of an event — labelling mandatory.
• Advertising showing a product — requirements are usually stricter and a label is often formally required.
• News or informational material — mandatory.

How to label

Briefly and without apology. One line under the image or in the caption: "image created with AI". That is enough.

What does not work: fine print at the end of a long text, a note in the comments, a passing mention. If a label is needed, it belongs where the image is.

Metadata and its fate

Provenance information is lost through most operations: re-saving, screenshotting, uploading to a platform that recompresses, cropping in an editor.

So metadata alone cannot be relied on: if the disclosure matters, it belongs in the caption rather than only in the file.

What comes next

The direction is clear: requirements will tighten, mechanisms will become more robust, automatic detection more accurate. The practical strategy today is to build the habit of labelling wherever it is warranted by meaning, without waiting for a formal requirement.

That is cheaper than reworking accumulated material once the requirement arrives, and it removes the risk of reputational questions raised retrospectively.