EU AI Act Labelling Rules for AI-Generated Content

If your product generates images, video, audio or text for a European audience, a new set of obligations just landed on your roadmap.

Since 2 August 2026, the transparency rules in Article 50 of the EU AI Act apply. Deepfakes have to be disclosed. AI-generated text published on matters of public interest has to be disclosed. And the European Commission has released an official set of icons you can use to do it.

Here's what actually changed, what you need to build, and where detection fits in.

What Article 50 requires for AI-Generated Content

Article 50(4) targets deployers: the organisations that professionally use a generative AI system and put its output in front of people. Two categories of content are covered:

1. Deepfakes. AI-generated or AI-manipulated image, audio or video content that resembles real people, objects, places, entities or events, and that would plausibly appear authentic to someone viewing it.

2. AI-generated or manipulated text published to inform the public on matters of public interest, but only where the text did not go through human review or editorial control, and where no natural or legal person took editorial responsibility for it.

The goal is straightforward: someone encountering the content should be able to tell, clearly and immediately, that it was artificially generated or manipulated. It's a misinformation and trust measure, not a restriction on using AI.

What is not covered

Plenty of AI output falls outside the rule. Three exceptions matter most in practice:

• Creative and satirical work. If a deepfake is part of an evidently artistic, creative, satirical or fictional work, the disclosure obligation is limited, it has to be done in a way that doesn't spoil the display or enjoyment of the work. A watermark across the middle of a film frame is not the intended outcome.
• Law enforcement contexts. The obligation doesn't apply where the use is authorised by law to detect, prevent, investigate or prosecute criminal offences.
• Human-reviewed text. If AI-generated text went through genuine editorial review and a person or organisation holds editorial responsibility for publishing it, the text disclosure obligation doesn't apply.

That last one is important for most content teams: an editor who actually reviews and signs off on AI-drafted copy changes the compliance picture.

The official EU icons for AI-Generated Content

The Commission has published a set of icons that anyone can use to label AI-generated content. There are three:

Icon	When to use it
	Basic icon — AI was involved in creating deepfake content or published text, or when you pair it with a custom text label or an interactive second layer
	Fully AI-Generated — the entire piece of content is AI-generated with no human-created elements and no human editorial control beyond prompting
	Partially AI-Modified — pre-existing, human-made content was partially modified with AI, turning it into a deepfake or into text on a matter of public interest

Each comes in four variations: black, white, and both at 50% transparency, downloadable as SVG and PNG from the Commission's page.

They're free. No attribution to the Commission or the AI Office is required.

One finding worth acting on: the Commission user-tested these icons, and recognition improved across every measure when the icon was paired with a short text label such as "modified". Don't ship the icon alone if you can help it.

Placement rules in brief

• The label must be perceivable and distinguishable at the latest at the moment someone is first exposed to the content.
• Place it where no overlay elements sit on top of it.
• Embed it directly into the content or use an equivalent alternative like a UI overlay. Critically, it must survive resharing and downloading.
• Make it accessible: visible size, plain language, alt text or ARIA labels so assistive tech announces it, and enough on-screen time to be read if the disclosure is temporary.

One caveat worth repeating

Using the icons is optional. The labelling obligations under Article 50 are not. Slapping an icon on your output does not by itself establish legal compliance, you remain responsible for making sure your disclosure actually meets the requirements.

The timeline

Content generated before 2 August 2026 does not have to be labelled retroactively.

There's also a voluntary Code of Practice on Transparency of AI-generated Content, assessed as adequate by both the Commission and the AI Board. Signing it gives you a recognised route to demonstrating compliance. Choosing not to sign means demonstrating compliance by other, equally adequate means.

The part nobody talks about: you have to know what's AI-generated first

Here's the operational gap. Labelling your own outputs is the easy case: you generated it, so you know. But most real content pipelines aren't that clean:

• A marketplace receives product images from thousands of third-party sellers.
• A news platform ingests wire photos, freelance submissions and UGC.
• A publisher commissions articles from external writers.
• A social product lets users upload video.
• A moderation team has to decide which of 50,000 daily uploads is a synthetic impersonation of a real person.

In every one of these, you're a deployer or distributor sitting on content you didn't create, and you have no idea which parts of it are synthetic. You can't label what you can't identify, and you can't sensibly assess deepfake risk on inbound media without an automated first pass.That's a detection problem, and it's exactly the layer Eden AI covers.

Building a compliance-ready content pipeline

Detection becomes genuinely useful when it's chained. A realistic flow for a publisher or platform:

1. Ingest: content arrives from a user, seller, freelancer or wire feed.
2. Detect: text goes to AI content detection; images and video frames go to AI image / deepfake detection.
3. Branch on score:high confidence synthetic → label automatically; ambiguous → human review queue; clean → publish.
4. Moderate:run explicit content and moderation checks in the same pass.
5. Label: apply the appropriate EU icon (basic, fully AI-generated, or partially AI-modified) with a plain-language text label, embedded so it survives resharing.
6. Log: keep the detection result and the labelling decision. If a supervisory authority ever asks how you assessed a piece of content, that record is your answer.

Eden AI's workflow builder lets you compose steps 2–4 as a single pipeline rather than as separate integrations, and every call is tracked and costed in one dashboard.

Detecting AI-generated text with Eden AI

Eden AI's AI Content Detection API gives you a single endpoint that routes to multiple specialised detection providers: Originality.AI, Winston AI, Sapling, GPTZero and others without creating an account with each one.

import requests

url = "https://api.edenai.run/v3/universal-ai/"
headers = {
    "Authorization": "Bearer <your-api-key>",
    "Content-Type": "application/json"
}
payload = {
    "model": "text/ai_detection/sapling",
    "input": {
        "text": "This is AI generated"
    },
    "show_original_response": False
}

response = requests.post(url, headers=headers, json=payload)
data = response.json()

if data["status"] == "success":
    print("Output:", data["output"])
    print("Cost:", data["cost"])
else:
    print("Error:", data.get("error")) 

‍

You get a probability score back. What you do with it is a product decision: route to a human editor, block the submission, require the contributor to declare AI use, or automatically attach the appropriate EU label.

Why route through several providers? AI text detection is probabilistic and detector accuracy varies significantly by model, language and text length. Calling two or three providers and comparing scores gives you a far more defensible signal than trusting one, and with Eden AI that's a parameter change, not three integrations.

Detecting AI-generated images and deepfakes

The same pattern applies to visual media through the AI Image Detection and Deepfake Detection features:

import requests

url = "https://api.edenai.run/v3/universal-ai/"
headers = {
    "Authorization": "Bearer <your-api-key>",
    "Content-Type": "application/json"
}
payload = {
    "model": "image/deepfake_detection/resemble",
    "input": {
        "file": "<file_id_from_upload_endpoint>"
    },
    "show_original_response": False
}

response = requests.post(url, headers=headers, json=payload)
data = response.json()
print(data)

‍

Setting providers to all runs the image through every available detection engine at once: useful during evaluation, when you're deciding which provider to standardise on, and useful in production when a false negative is expensive. 

FAQs - EU AI Act Labelling Rules for AI-Generated Content