How to label AI-generated content in 2026

Knowing how to label AI-generated content is now part of publishing video. YouTube, Instagram, Facebook and TikTok ask you to disclose realistic content created or altered with AI, and in the European Union the transparency obligations of the AI Act have applied since 2 August 2026. This guide sums up which content has to be disclosed, what each platform asks for, what changes with Article 50 and how to build it into your workflow so that nothing is forgotten.

What content you need to disclose

The platforms share an underlying criterion: you have to disclose content that looks real and that someone could take for real. YouTube’s help page on how to disclose the use of generative AI spells it out with three cases:

• Making a real person appear to say or do something they did not say or do.
• Altering footage of a real event or place.
• Generating a realistic scene that did not happen.

In 2024, YouTube introduced this obligation as the disclosure of altered or synthetic content; today its help pages talk about disclosing the use of generative AI, with the same criterion. On the other hand, YouTube does not ask you to disclose content that is clearly unrealistic, such as animation or fantasy, minor adjustments (colour, filters, sound enhancement) or the use of AI as a production aid, for example to write scripts, brainstorm ideas or come up with titles.

A practical way to decide: imagine a viewer who watches your video without reading the description. If they could believe that the person said it, that the place looks like that or that the event happened, you have to disclose it.

Disclosing and being original are different things. A properly labelled piece can still be ineligible for monetisation if it looks as though it came off a template: we explain this in inauthentic content on YouTube.

What each platform asks for

This table sums up the essentials for the three platforms that offer a disclosure tool at the time of publishing. Menu names change often; the source for each row is linked.

X works differently. Its synthetic and manipulated media policy allows it to label or remove posts with media that has been deceptively altered or fabricated and that may cause confusion or harm. If you publish on X, apply the same transparency criterion and say so in the post itself.

Where the label appears and when the platform adds it

Disclosing does not always mean the same thing for the viewer. Where the label appears depends on the type of content and the platform, and platforms also add labels without you asking:

• YouTube. According to its May 2026 update, for realistic content the label moves to a more visible place: below the player on long-form videos and overlaid on the video in Shorts. For less realistic content it stays in the expanded description. If you do not disclose and its systems detect significant photorealistic use of AI, the label is applied automatically. YouTube states that the label on its own does not change how the video is recommended or whether it can earn money.
• Meta. The label was first called “Made with AI” and was renamed “AI info” in 2024. According to its explanation of how it labels content, for content that has only been edited with AI the label moves into the post’s menu, while on AI-generated content it stays in view.
• TikTok. Your disclosure appears as an AI-generated content label. If the platform detects C2PA content credentials, it adds the label itself and it cannot be removed.

What the EU AI Act requires

Article 50 of Regulation (EU) 2024/1689 sets transparency obligations for those who develop AI systems and for those who use them. According to the Commission’s FAQs, it has applied to both since 2 August 2026.

These are the two obligations that most affect a video channel:

• Providers (Article 50(2)). Anyone who offers a system that generates synthetic audio, images, video or text must mark the outputs in a machine-readable format, so that they can be detected as artificial. Systems that were already on the market before 2 August 2026 have until 2 December 2026 to comply with this part. It is the only postponement in Article 50.
• Deployers (Article 50(4)). Anyone who uses the system under their own authority, for example a company that publishes videos, must disclose deepfakes: images, audio or video that have been generated or manipulated, that resemble real people, places or events and that could appear authentic. For an evidently artistic, satirical or fictional work, it is enough to disclose it without spoiling the work. They must also disclose AI-generated text that is published to inform the public on matters of public interest, unless it has gone through human review or editorial control and someone holds editorial responsibility for it.

Three nuances matter in practice. The deployer’s disclosure has to be visible or audible: the technical marking added by the tool is not enough. Content created before 2 August 2026 does not have to be labelled retroactively, although the Commission recommends doing so where possible. And the Commission has published guidelines on Article 50 and a voluntary code of practice that help to interpret it.

If you produce videos for your channel with third-party tools, you will normally be the deployer of those tools. Each Member State designates its own supervisory authorities, so check your own country’s rules as well.

Metadata: C2PA and IPTC

Besides the visible label, a file can carry information about its origin. This is the basis of machine-readable marking and what platforms read in order to add labels themselves.

• C2PA. An open technical standard that attaches a signed manifest to the file with its origin and its edits, known as content credentials. If someone alters the file, the signature no longer matches.
• IPTC DigitalSourceType. A metadata vocabulary that indicates where an image or a video comes from. For example, it distinguishes content created with generative AI (trainedAlgorithmicMedia) from content edited with generative AI (compositeWithTrainedAlgorithmicMedia).

Metadata has a limitation: it can be lost when the video is re-encoded, edited in another application or screen-captured. That is why it does not replace disclosure on the platform, or a visible notice when one is needed.

How to build disclosure into your workflow

Disclosure fails when it depends on someone remembering. Make it one more step in production:

1. Take an inventory for each piece. Note which parts are generated (image, video, voice or text) and with which tool.
2. Decide using each platform’s criteria. Does it look real? Does a real person appear? Does it alter an event or a place?
3. Disclose when uploading. It is more reliable than correcting afterwards, and some labels cannot be changed once applied.
4. Add a visible notice when a recreation could mislead, for example on-screen text or a line in the description.
5. Avoid recreating real people. If it is essential, get their consent and disclose the piece.
6. Keep a record of what was generated, with what and when, alongside the piece.
7. Check after publishing that the label appears where you expect it.
8. Review the policies regularly, because they change often.

How CLAQON handles it

CLAQON automates the mechanical part of disclosure on YouTube and keeps a record of what goes into each piece. Disclosure on each platform remains the responsibility of whoever publishes, and we tell you what it does not cover today.

• Synthetic content disclosure on YouTube. If any scene is a generated image or video, the piece is disclosed as synthetic content when it is uploaded. A synthetic voice on its own does not trigger it: this is a CLAQON criterion for narration that does not imitate the voice of any real person. If disclosing it makes sense in your case, do so from YouTube Studio.
• Private upload until the disclosure is accepted. Production can upload the video as private and switch it to public once YouTube has accepted the disclosure.
• Follow-up check (beta). For 7 days it reviews what has been published and alerts you, among other cases, to a pending disclosure.
• Provenance record for each piece. A downloadable document with licences, providers, credits and the IPTC digital source type. Anything not on record is left blank: it is never filled in.
• Real-people guard. It detects proper names in generated scenes and warns you. If you enable it, it diverts those scenes to stock clips or to non-photorealistic illustration.

What it does not do: it does not write C2PA or XMP into the final video, it does not apply Meta’s “AI info” label and it does not add an AI label on X. You will find the details on our compliance record and AI transparency pages.