What do disclosure labels look like?
What a good AI disclosure actually looks like in practice - across images, video, audio and text.
The law does not prescribe one giant watermark pasted across everything. The better way to think about it is: the disclosure must be clear, distinguishable and there at first exposure, without ruining the content itself.
What the guidance says a disclosure should do
The draft European Commission guidance does not prescribe a single wording or design for AI disclosures. Instead, it sets out a number of principles that disclosures should follow to ensure people can recognise AI-generated or AI-manipulated content clearly and at the right time.
- Be clear and distinguishable – The disclosure should be easy to identify and clearly separate from the surrounding content. It should not be hidden in terms and conditions, documentation or nested menus.
- Appear at first exposure – People should see the disclosure when they first encounter the AI-generated or AI-manipulated content, not after they have already viewed, watched or listened to it.
- Be easy to understand – The disclosure should make the artificial origin of the content obvious to the intended audience using clear, straightforward language.
- Be accessible – Where accessibility requirements apply, the disclosure should be available in an accessible format so it can be understood by the people it is intended for.
How labels usually appear by content type
Images
A small visible overlay is the normal answer. Best practice is a clear icon or label placed directly on the image or in an equivalent on-image overlay.
Video
The label should appear at the start and, for longer or interrupted video, reappear at intervals. Hiding it only in end credits is too late.
Audio
If there is no visual surface, use a short audible disclaimer at the beginning in plain, simple language. If a screen is available, adding a visual label helps too.
Published text
For public-interest text that needs disclosure, the notice should be visible when the reader first opens or encounters the publication, not buried in a footer or policy page.
The best-practice starting point
The draft code of practice supports a straightforward visual approach: use AI as the main visual cue, then make clear whether the content was generated from scratch or modified from something real. That is why the clearest two-label approach is:
For wholly synthetic content
Use when the image, clip, voice or published text was created by AI rather than altered from a real original.
For real content changed by AI
Use when a real photo, video, voice or text was materially altered by AI in a way that affects what the audience is seeing, hearing or reading.
Why these two labels matter
Draft EU guidance says labels with explicit wording such as generated and modified are easier for people to understand than a bare symbol alone. Clarity beats cleverness.
Where should the label go?
The draft code of practice gives a practical placement direction: for visual media, put the label in an appropriate visible place with no competing overlay elements - for example, the top-right corner of an image or video frame.
Image placement
Put the label directly on the image, or as an overlay that clearly appears to sit on the image. Leave enough spacing so it does not clash with other badges or UI chrome.
Video placement
Show the label at the beginning, then again through the video where appropriate. This matters because viewers may join mid-stream or see clipped fragments later.
Audio placement
Use a short audible disclosure before or at the start of the deep fake audio itself. If a player interface is present, pairing it with visual disclosure is stronger.
Text placement
The safest pattern is a visible disclosure near the headline, deck, byline or opening lines so the reader sees it as part of the first reading experience.
What can you customise?
Quite a lot. Best-practice guidance allows labels to vary in size, colour, contrast and typography as long as they remain clear, accessible and recognisable. In other words, you can make the label fit your brand, but you should not make it so subtle that people miss it.
Customisation should change style, not meaning
A good custom label still tells people the same essential thing: this content is AI generated or AI modified. Brand polish is fine; ambiguity is not.
What not to do
Do not hide it in metadata only
Machine-readable markers are useful, but they do not replace the visible or audible disclosure a person should get at first exposure.
Do not leave it to end credits
If someone has already watched or listened before learning it was AI, the disclosure arrived too late.
Do not make it microscopic
A label can be compact without becoming unreadable. If normal users would miss it, the design is too weak.
Do not let branding overpower the disclosure
If the label looks like decorative UI or a generic feature badge, people may not understand what it means.
The practical takeaway
Good labels are modest but unmistakable. They do not need to dominate the page, but they do need to be there clearly, early and consistently. For most businesses, the right answer is a small on-content label for image and video, a short spoken notice for audio, and a visible notice near the start of any public-interest text that actually falls within the rule.
See how disclosure would look on your own site
Scan your site free and see what may need disclosure - no signup required.
- Instant report
- No sign up required