Disclosure and Watermarking in AI Content: Which Content Must Companies Label?

Updated: 3 September 2026 · Category: AI Transparency

Disclosure and Watermarking in AI Content: Which Content Must Companies Label?

Article 50 of the AI Act requires user information for certain AI systems, machine-readable marking of AI-generated or manipulated content, and disclosure for content such as deepfakes. Not all AI content needs the same label; distinguish roles and use cases.

Contents

Short answer
Why does it matter?
What should companies do?
How to measure it?
Sources
Frequently asked questions

Short Answer

Why Does It Matter?

Both “everything needs a watermark” and “AI generation never needs to be disclosed” are too broad. Article 50 provides different transparency mechanisms for different actors and system types. A provider’s responsibility to mark synthetic content in a machine-readable format differs from a deployer’s responsibility to disclose deepfakes to users.

First determine the content type, the company’s role and the context of use. Treat user-visible disclosure, technical marking and internal evidence records as three separate control layers.

What Should Companies Do? Disclosure decision tree

Does the AI system interact directly with humans?
Does the system produce synthetic audio, images, video, or text?
Is the company the system’s provider or deployer?
Does the content depict a real person or event as a deepfake?
Is machine-readable technical marking provided?
Is visible disclosure also required for the user?
Has Disclosure been checked for location, time, language, and accessibility?
Are the final content and the transparency decision linked to the same evidence record?

Proposed Measurement Model for B10

Transparency layer	Measurement
Role	Provider/deployer classification coverage
Technical marking	Detectability test pass rate
Visible Disclosure	Disclosure implementation rate
Placement	Visibility/accessibility test pass rate
Version	Final-output match rate
Exception review	Documented exception decision rate

Do not assume that a technical watermark automatically removes the need for visible disclosure. Decide separately according to the applicable Article 50 obligation and use case.

Expert Assessment

Companies should not expect creative teams to interpret legislation for every piece of content. A predefined disclosure decision tree based on the production tool and content type improves decision quality and consistency.

Watermarking and disclosure are different

The first disclosure decision-tree question concerns the company’s role and output type, rather than the tool used. Technical watermarks or metadata can provide machine-detectable signals. User-visible disclosure is a separate communication layer that helps people understand that content was artificially generated or manipulated.

Avoid automatic conclusions such as “the file has metadata, so no disclosure is needed” or “the image has an AI label, so technical marking is unnecessary”. Record separately which obligation each layer satisfies in the decision record.

The Location of Disclosure is also Part of Control

Where a disclosure appears matters as much as its presence. It may fail operationally if users cannot notice it when seeing the content, if the text is too small or if it is hidden on another screen. Different placement standards are therefore useful for social media, video, websites and interactive AI experiences.

How can technical testing be done?

For content with machine-readable marking, test whether resizing, compression, video conversion or platform uploads after production damage the mark. Do not rely solely on the production tool’s documentation; verify the capability on sample files at the end of the actual publication pipeline.

Versioned test records show which platforms or post-processing operations preserve or damage marking information. This gives the technical team a concrete checkpoint in disclosure governance.

Disclosure QA and Exception Register

Sampling disclosure decisions for quality checks improves consistency. Teams classifying similar AI content differently can create conflicting practices within one brand over time. A central team can periodically reassess selected content and measure where interpretations of the decision tree diverge.

Keep an exception register for cases outside the standard workflow. Clearly record the reason, decision owner, basis and validity period so temporary exceptions do not become permanent practice. Track both how many items are labelled and how many cases depart from the standard transparency model, and why.

The register also reveals training needs. If recurring misclassifications cluster in a particular content team or channel, use targeted guidance and control updates instead of general training.

Digital Risk and Compliance Analysis

B10 can design a transparency workflow that embeds provider/deployer roles, technical marking and visible disclosure decisions as automated checkpoints in AI content production.