Your AI-generated content needs a label. Here's what that actually means.

A pattern I keep running into when I look at how teams actually handle AI content labeling: it works for the first week. Someone adds the “AI-generated” caption by hand, carefully, every time. Then a launch happens, or that person goes on vacation, or the content starts coming from three tools instead of one, and the caption becomes something that happens most of the time instead of every time. Nobody decided to stop labeling things. It just stopped being the first thing anyone thought about at publish time.

That's the part of content labeling under Article 50 that doesn't get talked about enough: it's not really a legal question. Almost everyone already agrees AI-generated content should say so. It's an operations question. And most teams are solving it with a process, when what it actually needs is infrastructure.

For context on the obligation itself: I wrote a longer explainer on Article 50 a little while back. The short version for this post is that content labeling is one of its three obligations, alongside disclosure and audit logging, and it covers AI-generated text, images, audio, and video that gets published or distributed. It needs to be identifiable as AI-generated where the Act requires it. It doesn't say how, and that gap is exactly where the gallery-page problem lives.

“How” usually turns into: someone adds a caption when they remember, a CMS field that's optional and therefore empty half the time, or a note in the team wiki that says “always label AI content” that nobody has opened since the day it was written. None of these are bad-faith. They're just what happens when a compliance requirement gets treated as a checklist item instead of a property of the content itself.

The alternative is to make the label a byproduct of generating the content, not a separate step after it. Concretely, that means two different things depending on what you're labeling, and it's worth being specific here because the two cases really don't work the same way:

• For an image, the label has to be part of the file. A caption on the page around the image doesn't survive that image being downloaded, reposted, or screenshotted, and images get downloaded, reposted, and screenshotted constantly. The only way a label travels with an image is if it's baked into the pixels.
• For text or an article, baking anything into the content itself would mean mangling the actual words, which obviously isn't acceptable. What works instead is a badge positioned with the content, generated at the same time as the content and returned alongside it, so there's no separate step where someone has to remember to add it.

That image/text split maps onto a distinction worth understanding on its own: machine-readable marking versus a visible label aren't interchangeable, and Article 50(2) asks for the former even when a visible badge is also present.

This is, not coincidentally, exactly how Disclarion's SDK approaches it: a real watermark pixel-baked into images, a positioned HTML badge for text, both produced by the same call that logs the interaction in the first place. I'm not going to pretend that's an unbiased example, it's literally the product I work on, but I think the underlying point holds regardless of which tool anyone uses for it: if labeling depends on a human remembering to do it at publish time, it will eventually not happen. Not because anyone stops caring, but because that's what happens to every manual step over a long enough timeline.

Most of the unlabeled AI content out there almost certainly isn't trying to mislead anyone. It's just what a missed caption looks like at scale, and scale is the normal operating condition for anything generating content with AI. Worth checking your own output against, before an auditor, or a customer, does it for you.