Do I Need to Label AI Images on Social Media?

Not all of them. The duty turns on whether the image could be mistaken for something real, which means an obviously synthetic illustration and a photorealistic fake are treated very differently.

Marketing teams generate images by default now. The question of whether to label AI images on social media usually gets answered with a shrug, or with a blanket policy that labels everything. Both are wrong, in opposite directions, and the second one is worse because it trains people to ignore the label.

What the Act actually covers

Article 50 of the EU AI Act, in force since 2 August 2026, treats synthetic visual content as a deployer duty where it constitutes a deepfake: image, audio or video content that appreciably resembles existing persons, objects, places or events, and would falsely appear to a person to be authentic.

That definition is doing all the work. What matters is whether someone could reasonably take it for a record of something that happened. Whether AI made it is a separate question.

An obviously illustrated hero image is not pretending to be a photograph. A photorealistic image of your office that does not exist is.

When you do need to label AI images on social media

• Photorealistic images of people, whether or not they resemble anyone specific. A fabricated customer, a stock-style team photo that is not your team, a face that does not exist.
• Images of real places or events that have been generated or altered. A product in a location it was never in, a venue dressed differently, a crowd that was not that size.
• Altered photographs where the alteration changes what the image is evidence of. Retouching a blemish is not this. Removing a competitor’s sign, or adding people to a half-empty room, is.
• Anything presented as documentation. Before-and-after shots, case study imagery, site photos.

When you probably do not

• Obvious illustration and abstract art. Nobody is going to mistake a stylised vector header for a photograph.
• Icons, patterns and backgrounds. No claim about reality is being made.
• Clearly fantastical imagery. If it could not have happened, it does not falsely appear authentic.
• Internal material. The Act is about what reaches other people.

There is a judgement call in the middle of that list. The line between stylised and photorealistic is not always obvious, especially once a designer has been at it. Label it when you are unsure. It costs almost nothing.

Does the platform doing it for me count

Several platforms now attach their own AI labels, either from metadata or from detection. That is useful, and it is not a substitute for your own duty.

Platform labels are inconsistent, they get stripped when an image is re-uploaded or screenshotted, and they are applied by a third party using signals you do not control. Relying on one means your compliance depends on somebody else’s detection working. Put the disclosure in the post text as well.

The machine-readable part, and who it belongs to

There are two different marking duties in Article 50 and they belong to different parties.

Providers of generative systems have to mark output in a machine-readable format, so software can tell it was artificially generated. That is the model maker’s job, not yours, and systems already on the market before August 2026 had until 2 December 2026 to implement it.

Deployers who publish deepfakes have to disclose that the content is artificially generated or manipulated. That is a visible disclosure to a human reader, and it is yours.

Confusing the two is common. Your image generator adding invisible metadata does nothing for your disclosure duty.

How to label AI images on social media without ruining the post

Short, factual, and in the post rather than buried in a thread:

Keep it: in the caption, not only in alt text, and not only in a reply. The same principle applies to AI-generated text.

The practical problem

Most marketing teams cannot say, three months later, which images on their feed were generated and which were photographed. Half the companies in our study keep no written record of where AI is used at all.

That is survivable until a client asks, or a platform flags a post, or someone asks whether an image was photographed. At that point you need to know what you made, when, with what, and who approved it.

That is what Art50 records: what was AI-generated, who reviewed it, and a sealed proof anyone can check from a link. The free plan covers twenty records a month, no card.

Figures come from The Human Behind It, our October 2026 study of 53 companies. Art50 provides evidence, not legal advice. See the EU AI Act timeline and the full text of Article 50.