Article 50: A guide to when you need to say AI is involved
Article 50 of the EU AI Act is the transparency rule, and it applies from 2 August 2026. It is the part of the Act that reaches what most of us actually do day to day: it decides when you owe someone a plain word that what they are looking at, listening to, or talking to, was made by AI.
I went looking for a plain explanation of it, something I could hand to a working creative, a marketer, the person running a chatbot, or an ordinary person scrolling their own camera roll, and have them understand it in five minutes. What I found instead was either the legal text, exact and unreadable, or a summary so smoothed over it was simply wrong. So I worked through it myself. This is the version I wish someone had handed me.
Handled properly, the effort needed is a sentence and a label, not a compliance project.
Article 50: The misunderstanding
Many get this wrong in one of two ways, without noticing which one they have picked.
- Treating anything an AI tool has touched as suspect. Colour-correct a photo, crop it, run a paragraph through spellcheck, and by that logic you owe the world a warning. Work that way and you bury the disclosures that matter under a hundred that don’t.
- Assuming the rule only catches content built end to end from a prompt, so a real photograph rebuilt by AI goes through untouched. This is the more dangerous of the two: the realistic fake is exactly the thing this rule exists to catch.
Which tool you used does not matter. What matters is what came out the other end, and that is the test every section below comes back to.
Article 50: Two rules, one name
Article 50 does two separate jobs, and they belong to two different people. Missing that distinction is the single most common way to misread the rule.
- Marking. The tool that generates the content buries a machine-readable watermark inside it, something other software can find later. That is the maker’s job, not yours. Use a mainstream tool and you inherit the mark for free. You never have to touch it.
- Disclosure. If you publish AI content that could pass as real, you say so, out loud, at the point where a person actually meets it. That one is your responsibility.
The key point is one the Commission has addressed directly: a hidden watermark does not count as disclosure. The logic is simple: nobody notices a watermark while they are looking at a picture. So “the tool already marks it” is not an argument. The mark is for machines to read. The disclosure is for people to read, and the people part is yours.
Two questions, and you have your answer
When you are unsure whether something needs a visible label, you are really asking two things about the result, and neither question is about the software.
- Did the AI substantially make or change what this represents? Tidying does not cross that line: colour work, cropping, retouching a blemish, lifting a stray cup off a table, fixing grammar. Building the thing does: generating an image from a prompt, swapping a face, adding a scene that never happened.
- Would it pass for real? If an ordinary person could mistake it for a genuine photo, a real recording, something that actually happened, it is what the Act calls a deepfake, and it must be disclosed. Intent doesn’t count, which surprises people. A perfectly innocent, realistic image still counts.
Let’s use Photoshop to give an example
The same tool sits on either side of the line, depending on what it is asked to do.
- Use generative fill to remove a coffee cup and you have tidied a real photo. Nothing to disclose.
- Use the same feature to drop in a crowd that was never there, or place a real person somewhere they never stood, and you have built something realistic and false. That must be disclosed.
The software stayed exactly the same in both cases. What changed was the result, and that is what decides if you should add a disclosure or not.
Two edges are worth keeping in mind
- Content that is obviously unreal is not a deepfake, so a picture of you riding a dragon needs no disclosure (unless you live in Westeros and your last name is Targaryen).
- A realistic image of a person who does not exist, the flawless synthetic model, counts and must be disclosed, even with no real face involved.
What about editorial accountability?
Text is treated differently in article 50 from images, audio and video. A real human editor who reviews the text and takes responsibility for it can lift the obligation there.
For images, audio and video, that door does not open. Signing off on a realistic AI image does not clear you from the responsibility. The only relief sits with work that is plainly artistic, satirical or fictional, where you disclose in a way that does not spoil it, for example a credit rather than a stamp across the middle.
Putting this in the context of who you are
Article 50: If you make content for a living
The disclosure responsibility sits with you here, not with your tool. Post a realistic AI image or video of yourself, an avatar, a voice clone, a face-swap, a photo of a moment that never happened, and it must be labelled where the viewer meets it, not three lines into the caption.
If your audience is talking to an AI version of you, they must know that from the first message.
Both ordinary beautifying and evident fantasy need no disclosure, but do not lean on your tool’s watermark to cover the rest, because it doesn’t.
Article 50: If you run marketing or sales
Watch the realistic asset dressed up as proof. An AI customer testimonial and a synthetic spokesperson who reads as a real person both must be disclosed, and both carry a second exposure under consumer-protection law that may be the bigger problem than the label. A realistic AI product shot that could pass for a photograph sits on the disclose side too.
Your written copy is mostly safe: ordinary campaign writing is not treated as a matter of public interest. The exception is content on health, the environment, sustainability, science or public affairs, where AI-generated text needs disclosing unless a named person genuinely reviewed and stood behind it. A quick sign-off does not clear that bar.
Disclose at the point of contact, never in the small print.
Article 50: If you run operations or customer service
Your exposure is the chatbot and the voice agent. From 2 August, a customer must learn they are dealing with AI immediately, in plain language.
A line buried in the terms and conditions does not count. Calling it an “assistant” is too vague to lean on. A voice line needs an audible cue, not a sentence hidden in a privacy policy.
If your agent acts on its own and contacts other people on your behalf, the disclosure goes with it, because those third parties are owed the same clarity as your own customer.
Design the disclosure for the person who will actually meet it, including the elderly and the less tech-confident, not for the sharpest user you can picture.
Article 50: What about me just posting on social media for fun?
This is the group most of the coverage has ignored, and it is the one most people actually belong to. A joke face-swap with your mates, an AI filter selfie, a video of your dog “talking”, a fun deepfake made for an evening’s entertainment and nothing else.
Here the news is good. The AI Act has a carve-out (Article 2(10)) for exactly this. It does not apply to the obligations of deployers who are natural persons using an AI system in the course of a purely personal, non-professional activity.
Posting for fun, with no business behind it and no income attached to it, sits inside that carve-out. Legally, the Article 50 disclosure responsibility is not reaching you.
BUT that relief stops short of a blank cheque
- The line moves the moment posting stops being purely personal. An influencer with sponsorships, a creator monetising a following, a small business owner posting from a personal account, all slide back into the section above on making content for a living, because the activity is professional even when the account looks casual. If money, a brand deal, or building an audience as work is part of why you post, assume the exemption does not cover you.
- The platforms do not care what the Act exempts. TikTok, Instagram and Facebook run their own rules on AI content, independent of European law, and those rules are often stricter and enforced faster than any regulator will. If the app offers you a built-in “AI-generated” label, use it. It costs a tap and it covers you on both fronts at once.
And one thing the law was never going to fix for you. A realistic clip built from a real person’s face or voice, posted without asking them, is not primarily a disclosure problem. Label it correctly and you can still have caused real harm, to their reputation, their image rights, or simply their trust in you.
A two-question check
Even where the law does not force your hand, the same two questions from earlier are a decent gut check:
Did you substantially build or change what this shows?
Would a stranger take it as real?
If both answers are yes and another real person is in the frame, ask them before you post, not after.
The part worth remembering
I believe that regulations can be a gift rather than a punishment if done right, and I think it is worth reminding us what the real purpose of them are. In this case, strip away the mechanics and it comes down to something simpler:
The rule is not asking you to apologise for using AI. It is asking you to be the one who says so first.
Disclaimer
Precision is the whole point of writing this, so three points to call out:
- The detail I have drawn on for marking, disclosure and the two-question test is the European Commission’s guidance, published in final form on 20 July 2026, alongside a Code of Practice assessed by the Commission and the AI Board as adequate. Neither carries the same legal force as the Regulation text itself. It is, however, the yardstick regulators will measure you against, so I would treat it as live and act now rather than wait for the final wording.
- The personal-use carve-out above comes from the Act’s own scope provisions rather than from the Article 50 guidance itself, and it has had less scrutiny in commentary than the disclosure duty proper. Treat it as the right general shape rather than the final word on any one situation, particularly where money or an audience is involved.
- This is a general explanation of Article 50, not legal advice. Before you act on a specific piece of content, that is a question for a lawyer who can see the detail.
