The most common question we have had since August goes roughly like this: "We made an image with AI. Do we have to mark that somehow now?"
The honest answer is: it depends, and it depends at exactly the point where most people take the wrong turn. Because there is not one rule. There are two layers that apply independently of each other. Mixing them up produces wrong answers in both directions — you either label everything out of caution, or nothing at all.
This post separates the two cleanly. By the end you will know which question to ask of which post.
The two layers
Layer 1: the law. The European AI Act sets out transparency obligations in Article 50. They have applied since 2 August 2026. In Germany, the AI Market Surveillance and Innovation Promotion Act has been in force since 29 July 2026, designating the Bundesnetzagentur as the central market surveillance authority and complaints body.
Layer 2: the platforms. TikTok, Meta and LinkedIn have introduced requirements of their own. They predate the law, follow a different logic, and do not line up with it.
Both apply side by side. A platform can require something the law does not. And the law can require something the platform gives you no switch for.
Layer 1: what Article 50 covers
The provision that matters most for social media is Article 50(4). It addresses deployers — the people who use an AI system and publish the result. That is you.
It covers two cases:
Deepfakes. Content that depicts real people or actual events in a deceptively realistic way. Anyone publishing that should disclose that the content has been artificially generated or manipulated, clearly and recognisably.
Text on matters of public interest. Publishing AI-generated text in order to inform the public on such matters should likewise be disclosed.
The second case has an exception that matters a great deal to companies: it does not bite where the content underwent substantial human review or editorial control and a person or organisation holds editorial responsibility for it.
What matters here is how strict that standard is meant to be. Per the European Commission's explanations, spell checking, grammar checking, skimming or a purely formal check are expressly not enough. What is meant is a review of the substance by someone with relevant knowledge.
Two more points that often get muddled:
The machine-readable marking in Article 50(2) is addressed to the providers of AI systems, not to you as a user. That is the duty of OpenAI, Google and Anthropic, not yours.
And: the deployer obligations concern professional and business use in particular. Purely personal, non-professional use by private individuals may fall outside them. The platform rules still apply.
Layer 2: what the platforms require
This is where it gets concrete, because platforms have switches and the law does not.
TikTok
TikTok requires a label in particular for realistic-looking AI-generated or materially AI-altered images, video and audio.
The point almost everyone misses: this includes entirely invented people, places or events where they appear realistic. The fact that the person shown does not exist does not make the content uncritical — that is precisely the case in question.
Obviously stylised or unrealistic content is as a rule outside it. The switch sits in the post settings before you publish. Beyond that, TikTok recommends transparency for other heavily AI-assisted content too.
Instagram and Facebook
Meta runs two mechanisms at once. It detects AI content from technical provenance signals and applies an "AI info" note itself, whether or not you declare anything. And for certain photorealistic video and realistic-sounding audio, Meta expects you to disclose it yourself at upload.
For ads, additional transparency mechanisms apply. And for election advertising and ads about politics or social issues, separate requirements apply, including authorisation, a disclaimer naming who paid, and specific disclosure duties where image or audio material was digitally created or altered.
What does not follow: that running a commercial ad by itself makes every AI image subject to a label.
LinkedIn has so far published no general mandatory label for every upload. Concluding from that that LinkedIn has no position on the topic would be wrong, though.
For certain synthetic or manipulated media, LinkedIn requires clear disclosure of their artificial or altered nature — particularly where real people or events are depicted deceptively. Beyond that, LinkedIn recommends transparency where content is heavily AI-assisted or altered.
Technically, LinkedIn also displays a "Content Credentials" icon where a file carries provenance data under the C2PA standard. That is not a switch you operate; it is something the file brings with it, or does not.
The case everyone gets wrong
Take a photorealistic AI image: a woman in an office, friendly, nice light. The person does not exist. She is entirely invented.
Many people reason: invented, so nobody's personality rights are involved, so no problem.
That is the wrong conclusion. For the platforms' labelling rules, what matters is not whether the person is real but whether the content could be taken for a real recording. A photorealistic human can be, whether or not they exist.
At the statutory layer it runs the other way: through the deepfake concept, Article 50(4) attaches to depictions of real people and events. An invented but realistic scene is not automatically covered.
That is exactly why the layers have to be kept apart. The same post can require a label on TikTok while the statutory position stays open.
What about captions
The short answer: most captions need no label.
The labelling rules of Instagram, Facebook, TikTok and LinkedIn attach to image, video and audio content, not to text alone. A product caption written with AI triggers nothing there.
At the statutory layer it depends on what the text is about. Product, brand and corporate communication typically falls outside Article 50(4), because the provision attaches to matters of public interest. For a political, journalistic or socially relevant text it looks different — and then the question of human review becomes the one that matters.
And filters?
Cropping, exposure, colour correction, light retouching and upscaling without changing the content are generally not treated as AI-generated content.
The line runs where the edit materially changes how someone or something looks. An AI filter that noticeably changes a face, a body or other essential features is no longer harmless retouching. A face swap certainly is not.
One rule of thumb that holds
If you only remember one question, make it this one:
Could someone take this for a real recording?
If the answer is yes and AI was materially involved, labelling is at least worth checking. If the answer is no because the content is obviously illustrated, animated or exaggerated, the position is more relaxed.
And when you are unsure: a voluntary, transparent label is usually the lower-risk option. It costs you little and damages nothing — the belief that labelling ruins your reach is not one the platforms themselves share.
We turned this into a tool
Because these questions keep repeating and the answer depends on four factors, we put the rules into a free check.
It asks a few questions — what the AI did, how realistic the result looks, whether real people appear, where you are publishing — and then shows separately what the AI Act says, what the individual platform requires, and what makes practical sense. Each section links the official primary source so you can read it yourself.
It decides nothing for you and assesses no individual case. It shows you which rules come into play at all.
Finally, a word about this post
This post reproduces what the platforms have published and what the cited provisions cover, as at 9 September 2026. It is general orientation and not legal advice. We are not lawyers, and we are not assessing any specific case here.
Platform rules also change without notice. The current version at each provider always prevails. For commercially or legally sensitive publications, get competent advice when in doubt.
Sources: AI Act, full text (EUR-Lex) · European Commission, FAQ on the Article 50 transparency obligations · Bundesnetzagentur · TikTok Help Center · Meta on labelling AI content · LinkedIn Help