AI Act and AI-generated content labeling in 2026: a practical Article 50 guide

AI-generated content labeling has become an operational requirement rather than only a future policy discussion. Article 50 transparency obligations under the AI Act started applying on 2 August 2026 for covered cases. At the same time, the European Commission has published guidelines and a dedicated Code of Practice on the transparency of AI-generated content. For a business, the challenge is translating those rules into a workflow that works across campaigns, WordPress, ecommerce, video, audio, social media, and editorial publishing. The key point is that there is no single rule saying 'add an AI label.' Article 50 distinguishes roles, content types, and different obligations. A strong implementation starts by classifying each case and ends with evidence showing which decision was taken and why.

1. What changed on 2 August 2026

Article 50 transparency obligations started applying on 2 August 2026. The Commission published guidelines clarifying scope and a Code of Practice to help providers and deployers of generative systems implement marking and disclosure measures. That means organizations can no longer treat the topic as something to review in a future regulatory phase.

Companies need to identify which AI uses fall within scope and document that analysis. The presence of generative AI in an organization does not automatically mean every asset needs identical treatment. An image generator used for internal moodboards, an avatar published in a campaign, and generated text informing the public on a matter of public interest present different contexts.

• Key date: 2 August 2026.
• Guidelines clarify scope; the Code of Practice provides an operational path.
• Use-case classification should come before disclosure.

2. Provider and deployer do not have the same obligation

The Code of Practice divides its commitments into two sections. The first is aimed at providers and focuses on marking and detection of AI-generated or manipulated content. The second is aimed at deployers and addresses disclosure of deepfakes and certain text publications. This distinction prevents applying the same technical obligation to an organization merely using an external tool and to one developing and offering a generative system.

A company may be a deployer in one workflow and a provider in another. For example, a marketing team using an external image generator usually occupies a different position from a company embedding its own generator in a product offered to third parties. The AI systems inventory should therefore record the organization's role for each use case.

3. Machine-readable marking: the technical layer

For providers of certain generative systems, Article 50 and the Code of Practice focus on outputs being marked in a machine-readable format and detectable as artificially generated or manipulated. The code refers to solutions that are effective, interoperable, robust, and reliable as far as technically feasible.

For a technical team, this creates a chain of checks: which signal the original system adds, whether the file retains it after export, what happens during editing, compression, or transcoding, and whether the CMS or CDN removes metadata. Transparency does not end at generation time; where required, the signal needs to survive the publishing pipeline.

A visible watermark and machine-readable marking are not synonymous. Machine-readable marking can rely on embedded technical signals—for example metadata, signatures, or robust watermarking—while a visible label serves a separate disclosure function. The appropriate measure should be assessed against the regulated role, content type, and channel.

4. Deepfakes: disclosure for content that could appear authentic

The Commission describes deepfake content as AI-generated or manipulated image, audio, or video resembling existing people, objects, places, entities, or events and potentially appearing authentic or truthful. This category deserves special attention because users may interpret the content as evidence that something actually happened.

Disclosure should be designed for the context in which content is consumed. A label that works on a webpage may not work the same way inside a video, downloaded image, or social media post. The organization needs a rule for each format and channel, together with a record of the asset and published version.

5. Public-interest text: the human-review exception is not universal

Article 50 addresses certain AI-generated or manipulated text published to inform the public about matters of public interest. European guidance identifies an exception where content has undergone human review and is subject to editorial responsibility. That exception should be understood in its specific context and not automatically extended to images, audio, video, or deepfakes.

For a media outlet, corporate blog, or institutional portal, the editorial process becomes a control. The organization should be able to demonstrate which version a person reviewed, who assumed editorial responsibility, and when it was approved. An informal review without a record is far harder to evidence than a workflow with states, owners, and dates.

6. Icons and labels: useful, but not a substitute for analysis

The European Union has created a set of icons deployers may use to label certain AI-generated content. These resources can help standardize communication, but using an icon alone does not answer questions about scope, role, format, or exceptions.

A strong policy separates design from decision criteria. First determine whether content is subject to disclosure and what information should be communicated. Then select the appropriate visual pattern, wording, or icon. This prevents the creative team from turning a transparency obligation into a purely aesthetic choice.

7. WordPress can break the evidence chain if the pipeline is not controlled

WordPress can regenerate images, create thumbnails, convert formats, and send files through optimization plugins or CDNs. Elementor and other builders add another publication layer. If the organization needs to preserve technical marking or demonstrate which file was published, those transformations need to be included in the technical inventory.

A useful practice is to retain the original received from the generative system, record its hash or identifier, store the published version, and connect both to the page where they appear. Where visible disclosure exists, the wording or component used should also be recorded. This evidence makes it possible to detect later whether optimization removed metadata or an edit replaced the asset without updating its status.

8. The Code of Practice is voluntary, but can provide a recognized framework

The Commission and AI Board have assessed the Code of Practice as an adequate voluntary tool for demonstrating compliance with covered transparency obligations. Organizations signing it can rely on its measures as a recognized way to show how marking and labeling are implemented.

That does not make the code the only possible compliance route. An organization may use other means, but it will need to demonstrate that those measures are adequate. For smaller teams, adopting a common reference can reduce ad hoc decisions and help technology, legal, marketing, and content teams use the same vocabulary.

9. Evidence: record the decision, not only the result

A screenshot showing a label proves that a label existed at one moment, but it does not explain why it was applied, which system generated the content, who reviewed the case, or which version was published. Stronger evidence combines the system inventory, asset classification, applied rule, owner, date, and a copy or reference to the published version.

This also helps when the decision is not to display disclosure. If the team documents the reason—for example because the case falls outside the analyzed category or because a specific exception applies—there is a basis for revisiting that decision when content, tools, or guidance change.

10. Operational checklist for a business in 2026

Start with a list of every generative system used in marketing, design, support, product, and content. For each one record the provider, purpose, generated formats, and owner. Then identify published assets and classify them as image, audio, video, or text. Flag cases that may be deepfakes or public-interest text publications and review which specific obligation applies.

Next check whether technical marking exists on the original and survives the pipeline, define visible disclosures where appropriate, record human review and editorial responsibility when relevant, and retain evidence. Finally, add periodic review because tools and technical interpretation evolve. Kairoseth AI Transparency can centralize that operational layer without turning a technical tool into a substitute for legal analysis.