EU AI Act: Do You Have to Label AI Content? (2026)

The short answer

If you publish images, audio, or video that were generated or meaningfully altered by an AI system, and that content reaches users in the European Union, the EU AI Act now requires you to label it clearly. The rules came into force on 2 August 2026 under Article 50 of Regulation (EU) 2024/1689.

The obligation falls on “deployers,” meaning the businesses and individuals who take AI-generated output and publish it to an audience. It is not limited to the AI companies themselves. A marketer who uses Midjourney images in an ad campaign, a blogger who publishes AI-illustrated posts, or an agency producing AI-narrated videos for a client all fall within the scope of the rule.

The label does not need to be large, but it must be perceptible. A discreet corner badge that says “Made with AI” is sufficient. Burying the disclosure in small print far from the content, or relying only on an invisible technical watermark, does not satisfy the requirement.

What EU AI Act Article 50 actually requires

Article 50 of the EU AI Act sits within the chapter on transparency obligations. You can read the full text on the EU Lex database (Regulation 2024/1689) and the European Commission has published supporting guidance on its digital strategy pages.

In plain language, Article 50 does the following:

• Article 50(1) and 50(2): System-level marking. Providers of AI systems that generate or manipulate content (images, audio, video, text) must ensure the outputs are marked in a machine-readable format. This is the technical watermarking layer: think Google's SynthID, or the C2PA metadata standard. The provider of the AI tool is responsible for building this in.
• Article 50(4): Visible human disclosure. Deployers who publish AI-generated or AI-manipulated images, audio, or video to the public must disclose this in a way that a person can clearly perceive. This is a separate, additional obligation on top of the machine-readable marking. It is the visible label that readers and viewers actually see.
• Deepfakes. The rule gives special attention to AI-generated or manipulated images and video that depict real people (deepfakes). For these, the disclosure obligation is particularly strict: a clear and prominent label is required. Satire and parody involving real people are subject to a lighter treatment, but cannot escape disclosure entirely.

The key point for anyone publishing AI content today: the regulation does not only target the AI companies. It reaches down to the businesses and individuals who use AI tools to produce content that they then share publicly. If you are the one hitting “publish,” you carry the disclosure obligation.

Two separate obligations: machine marking and human disclosure

A common misconception is that the invisible technical watermark already embedded by the AI tool you use is enough. It is not. Article 50 creates two distinct layers of obligation that work alongside each other.

If you generate an image with Midjourney or DALL-E, the AI company may embed an invisible marker. But that does nothing for the person looking at your Instagram post or your website. The visible label is your responsibility as the publisher.

The practical implication: any business or creator who publishes AI-generated images, audio, or video to a public audience needs a workflow for adding a visible label before distribution. That is true even if the AI tool says it already watermarks the output invisibly.

Who is affected in practice

The EU AI Act casts a wide net when it comes to who counts as a “deployer” for the purposes of Article 50. Essentially, any professional or business that takes AI-generated content and makes it publicly available is caught by the rule. Here are the most common categories.

Marketing teams and agencies

AI-generated visuals have become routine in digital advertising. A team producing social media creatives with Midjourney, generating product mockups with Adobe Firefly, or creating AI-narrated video ads would all need to label that content before publishing. The same applies to agencies producing this material on behalf of clients: the obligation falls on whoever publishes the content, not just whoever created it.

Social media managers and creators

A brand account that posts AI-generated images, a travel influencer who uses AI to fill in missing photos, or a content creator who publishes AI-generated voice-over content all fall within the scope of Article 50. Even a single post featuring AI-generated imagery published to a public account would, in principle, require a disclosure.

Ecommerce sellers using AI product images

Many online sellers now generate product visuals using AI tools rather than photography. A product listing that shows an AI-generated image of a bag, a garment, or a piece of furniture displayed in a setting that never existed is AI-generated content. Publishing that image without disclosure would put the seller in breach of Article 50, particularly when the image could be mistaken for a photograph of a real product.

Publishers and blogs using AI illustrations

Many publications, from large news outlets to individual bloggers, now use AI-generated images as article illustrations. If you publish an AI-generated illustration as the hero image of a blog post, or as a social media graphic to promote an article, the image needs a label. The same logic extends to AI-generated podcast thumbnails, newsletter headers, and any other public-facing visual.

Businesses with AI-generated audio or video

Article 50 covers audio and video, not just images. A business that uses an AI voice generator for customer-facing audio content, or that publishes videos where faces or voices have been AI-altered, must disclose this. The disclosure rule is particularly strict for deepfakes, meaning content where a real person's likeness has been generated or manipulated by AI.

Exemptions and nuances

Article 50 is not a blanket rule with no exceptions. Three notable carve-outs exist, and understanding them helps you calibrate your compliance effort correctly.

1. Editorial responsibility exception

Content published under genuine editorial responsibility is treated differently. Where a natural person holds editorial accountability for the content and where a human editorial process applies, the automatic disclosure obligation may not apply in the same way. This is primarily aimed at traditional media and journalism. Importantly, this exception is narrow: the editorial process must be substantive and genuine, not a rubber-stamp review. Simply having a human who technically approved the post is unlikely to qualify.

2. Law enforcement exception

Law enforcement, national security, and public security uses fall outside the standard transparency requirements in specific circumstances. This is a narrow professional exception that will not apply to commercial or content-publishing uses.

3. Artistic, creative, satirical, or fictional works

For content that is clearly artistic, creative, satirical, or fictional in nature, the disclosure obligation is lighter rather than absent. You must still disclose the AI origin, but you may do so in a way that does not materially spoil the enjoyment or intended effect of the work. A small label in the corner, or a caption line, satisfies this. Removing the obligation entirely by calling something “art” is not what the regulation permits.

Timeline: when do the rules apply?

The EU AI Act was published in the Official Journal of the European Union on 12 July 2024 and entered into force on 1 August 2024. Different parts of the regulation apply on a phased schedule.

If you are publishing AI-generated content right now, today is already inside the compliance window for Article 50. The visible labeling obligation is live. Providers of AI tools have until December 2026 to retrofit machine-readable marking into products that launched before August 2026, but that does not give deployers more time on the visible label. That obligation applies to content you publish from 2 August 2026 onward.

Penalties: why this matters beyond compliance

The EU AI Act sets out a tiered penalty structure, and the fines can be substantial. Non-compliance with the rules for general-purpose AI systems and with transparency obligations can result in fines reaching into the millions of euros, depending on the category of violation and the size of the company. The AI Act follows a similar philosophy to the GDPR in that regulators have meaningful enforcement tools available and have stated publicly that they intend to use them.

For businesses, the risk goes beyond the direct fine. A finding of non-compliance under the AI Act creates reputational exposure, can trigger follow-on investigations under other EU laws, and can result in civil liability to individuals whose likeness was used without proper disclosure (particularly for deepfake-related violations).

The practical counterpoint is that adding a visible label costs almost nothing. The compliance effort is genuinely low, and the risk of not complying is real. This is one of those situations where the cost-benefit calculation is straightforward.

How to comply in practice for images

The regulation does not specify an exact visual format for the disclosure. It requires that the label be “clearly perceptible” to a human viewer. In practice, the following approaches all qualify:

• A text overlay directly on the image reading “Made with AI,” “AI-generated,” or “AI-assisted.”
• A small badge or watermark in a corner of the image, clearly readable at normal viewing size.
• A caption immediately below the image stating that the image is AI-generated.
• For video and audio: a spoken disclosure at the start or a persistent on-screen label throughout.

What does not qualify on its own: an invisible technical watermark, a disclosure buried three paragraphs below the image in small type, or a generic site-wide policy statement that says “some of our images may be AI-generated.” The disclosure must be tied to the specific content and perceptible at the point where a viewer encounters it.

Step-by-step: adding a compliant label with SammaPix

The SammaPix AI Label tool is free, runs entirely in your browser, and requires no account or file upload. Here is how to use it:

1. 1Open the AI Label tool at sammapix.com/tools/ai-label.
2. 2Drop your AI-generated image into the tool or click to select it from your device.
3. 3Choose your label text. Options include 'Made with AI', 'AI-generated', 'AI-assisted', or custom text.
4. 4Select the position: top-left, top-right, bottom-left, or bottom-right corner, or centered.
5. 5Choose a label style: a solid badge, a semi-transparent overlay, or an outlined text format.
6. 6Preview the result and adjust if needed.
7. 7Download the labeled image. The file never left your browser. Everything processed locally.

For teams publishing at volume, you can also use SammaPix Stampit for batch watermarking across large image sets. The workflow is the same and the output is compliant with a visible disclosure on every image.