You created an image with Midjourney for your LinkedIn post. You edited a promotional video with a synthetic voice. You used AI to retouch a product photo. Well: from August 2, 2026, this content, in many cases, must be flagged as generated or modified with artificial intelligence.
The word going around is "watermarking." But behind this word there are two very different things, and understanding which one concerns you saves time and embarrassment. Let's look at them.
- Article 50 of the AI Act imposes transparency on content generated with AI. The rules become operational on August 2, 2026. Even where AI Act postponements are discussed, the transparency of article 50 stays in force from that date.
- There are two different obligations: the technical watermark (invisible, machine-readable) which belongs to whoever builds the AI tool, and the visible watermark (the label the user sees) which belongs to whoever publishes the content.
- If you use ChatGPT, Midjourney, Sora & co., the technical watermark is already inserted by the provider. Your job, as a company that publishes, is the visible declaration when the content could deceive.
- On images: visible label, well readable. On videos: disclaimer at the opening or fixed "AI" icon for the entire duration. Not just at the end.
First thing: watermarking doesn't mean one thing only
When the AI Act talks about "marking" content, it means two levels that are often confused.
The first is the technical watermark, or machine-readable marking. It's a signal that a computer can read but the human eye can't see: it can be data hidden in the pixels, signed information in the file's metadata, a cryptographic signature on provenance. It serves so that a machine — a social network, a search engine, verification software — recognizes that file was produced with AI.
The second is the visible watermark, that is, the declaration seen by the person looking at the content: a caption, an "AI" icon, a disclaimer at the beginning of the video. It serves to make a human understand that what they're looking at isn't real, or has been manipulated.
They are two separate worlds, with two different responsible parties. And here's the point that concerns you.
Who does what: provider vs publisher
The AI Act divides the roles in two. There's who provides the AI system (who builds and puts on the market ChatGPT, Midjourney, Sora, a synthetic voice generator) and there's who uses it to create or publish content (you, your company, your agency).
The invisible technical watermark is the provider's responsibility. Whoever develops the tool must mark the outputs in a machine-readable format, so they are recognizable as artificial. This obligation kicks in on August 2, 2026, with a grace period until December 2, 2026, for systems already on the market. In practice: OpenAI, Google, Adobe, and the other big players are already moving, and in most cases the file that comes out of their tool is already marked by itself.
The visible watermark is your responsibility. If you use AI to create or manipulate an image, audio, or video that depicts people, places, or events in a way that seems real — the classic deepfake — you must clearly declare that it's artificial, at the moment the person encounters it. The same applies to certain AI-generated texts published to inform the public on topics of general interest.
You can't govern what you don't see. And your audience can't trust what they don't know.
Translated in practice: you don't have to become a watermarking engineer. But you do need to know when the visible label is required, and how to place it well.
When you really need the visible declaration
Not everything you touch with AI needs to be labeled in the user's face. The visible rule mainly concerns content that can deceive:
- A video with a person (real or lifelike) who says or does things that never happened.
- A realistic photo of an event, place, or product that doesn't exist or didn't go that way.
- A synthetic voice that imitates a real person.
- An image retouched with AI to the point of altering the reality of what it shows.
Usually excluded are cases where AI has only helped marginally, or where it's obvious from context that it's artistic fiction or satire — but even there, a small label never hurts and protects you.
When in doubt, apply the simplest criterion: if someone might believe it real, declare it.
How to mark images compliantly
On images the visible declaration is simple. What matters is that it's clear, readable, and visible immediately, not hidden.
In practice, a badge or caption in a corner works well — for example "AI-generated content," "AI Image," or a small "AI" icon. Common-sense rules that keep you safe:
- Place it in a fixed corner, with good contrast against the background (light on dark or vice versa).
- Don't make it microscopic and don't confuse it with the design: it must be noticed.
- If you publish on a social network, also use the native "AI" label of the platform when available: it's one more signal, not an alternative.
- If you want to stay aligned with the European standard, the EU Commission provides a set of official icons to flag AI content (in light, dark, and semi-transparent versions).
Alongside the visible label, it's worth not removing the file's metadata. Many tools today save the image with the so-called Content Credentials (the C2PA standard): signed information saying with which tool and when it was created. If you export or recompress the image, try to preserve this data: they are part of the technical marking and help you demonstrate good faith.
How to mark videos compliantly
On videos the most common mistake is putting the notice only in the end credits. It's not enough: the person would watch almost the entire video without knowing it's artificial, and the obligation is to inform at the latest at first contact.
You have two paths, and it's enough to pick one:
- Disclaimer at the opening. A screen or caption at the start of the video: "This video was generated with artificial intelligence." Clear, readable, before the actual content starts.
- Fixed "AI" icon. A badge in a fixed position (usually a corner) that stays visible for the entire duration of the video.
Two useful details. For artistic or fictional works, where a permanent badge would ruin the experience, the label can be more discreet, but it must still be clear and recognizable at the moment the person encounters the content — for example an icon in the opening credits or embedded in the image. And the information must be made accessible: if the content has audio, also consider a sound notice or subtitle, because not everyone sees the icon.
As with images, the invisible technical watermark (metadata, pixel-level marking like Google's SynthID) is normally already inserted by the tool you used to generate the video. Your job is to add the visible layer and not erase the technical one.
Technical standards, in two words
You don't need them to comply personally, but it's useful to know them because you'll see them cited everywhere and because choosing tools that support them makes your life easier.
C2PA / Content Credentials. It's the standard by which the file carries with it a signed "ID card": who created it, with what tool, when, what modifications it underwent. If someone alters the file, the signature breaks. It's backed by a huge coalition — Google, Microsoft, Adobe, Meta, OpenAI, Sony, BBC, Amazon, and thousands of other players — and is coming into browsers and search engines.
SynthID and watermarks in pixels. They are invisible signals inserted directly into the pixels (or the audio). Unlike metadata, they don't disappear with a simple screenshot or recompression. Google has already marked in this way over 10 billion pieces of content across text, images, audio, and video.
The practical message: use serious tools, from providers that declare they adhere to these standards. That way a good part of the technical marking is already done for you, and what's left for you is the visible declaration.
The checklist to keep on your desk
Before publishing content generated or modified with AI, ask yourself:
- Could this content seem real to whoever looks at it? → If yes, the visible declaration is required.
- Image: did I place a readable "AI" label, in a corner, well contrasted?
- Video: do I have a disclaimer at the opening or a fixed icon for the entire duration? (Not just at the end.)
- Did I choose a tool that marks outputs according to recognized standards (C2PA / pixel watermarks)?
- Did I avoid deleting metadata when exporting or recompressing the file?
- If I publish on a social network, did I also activate the platform's native "AI" label?
If you answer yes to these questions, you're in a solid position.
An honest clarification
The most detailed practical reference today is the Code of Conduct on the transparency of AI-generated content, published by the European Commission on June 10, 2026. It's a voluntary tool: it's not the law, it's a concrete way to demonstrate you're respecting article 50. Adhering to it, or following its indications, is the simplest way not to get it wrong.
And remember the fundamental distinction, because it's the one that saves you from compliance anxiety: the invisible technical watermark is placed by whoever builds the tool; you, who publish, are responsible for the visible label when the content can deceive. A few clear rules beat a thousand prohibitions.
Sources
AI Act — Article 50 (Regulation EU 2024/1689): artificialintelligenceact.eu/article/50
EU Code of Conduct on transparency of AI-generated content (June 10, 2026): digital-strategy.ec.europa.eu
Official EU icons for labeling AI content: digital-strategy.ec.europa.eu — EU icons
C2PA / Content Credentials (provenance standard): c2pa.org
SynthID (Google DeepMind): deepmind.google/models/synthid
