EU AI Act: the Commission publishes the AI content labeling playbook — August 2, 2026 deadline, what this concretely changes for businesses

EU AI Act: the Commission publishes the AI content labelling playbook — deadline August 2, 2026, what concretely changes for businesses

🔎 15 weeks before the deadline, the EU has just published its code of practice for marking AI content

On August 2, 2026, Article 50 of the AI Act enters into application. This means that any image, video, audio, or text generated by AI and disseminated in the EU will have to be labelled. No grey area, no exemptions for small structures.

The European Commission has just published its Code of Practice on marking and labelling of AI-generated content. It is a voluntary playbook, designed to guide providers and deployers of generative AI systems towards compliance. According to AI News, this code was selected by the AI Office as the main reference.

The context is tense. As Thomas Eriksen on LinkedIn points out, the Omnibus deal clarified the timeline: prohibited practices have applied since February 2025, GPAI obligations since August 2025, and transparency with watermarking arrives in August 2026. Those who thought the Digital Omnibus would push back Article 50 are mistaken — Cookiebot confirms it: Article 50 was not touched by the Omnibus package.

The Essentials

• On August 2, 2026, Article 50 of the AI Act makes the labeling of all AI-generated content in the EU (text, image, audio, video) mandatory.
• The Commission has published a voluntary Code of Practice to guide compliance: machine-readable watermarking, metadata, automatic disclosure.
• Fines can reach 35 million euros or 7% of global annual turnover.
• Limited-risk systems (chatbots, generative tools) are directly affected by these transparency obligations, according to DS Solutions.
• Most current DAMs (Digital Asset Management) are structurally incapable of applying watermarking at scale.

Recommended Tools

What Article 50 exactly says — no room for debate

Article 50 imposes a clear transparency obligation: users must know when they are interacting with an AI system or consuming AI-generated content. Period.

BuildThisNow summarizes the situation: starting August 2, 2026, anyone exposed to AI content in the EU must be able to identify it. This covers deepfakes, AI-written articles, generated images, and synthetic voices.

The regulation applies uniformly to the development, provision, and use of AI systems on the European market. There is no company size threshold for this specific transparency obligation.

Those who want to dive deeper into the AI Act européen : ce qui change concrètement pour les devs en 2026 will find the detailed technical implications. But here, we stay focused on labeling.

What counts as "AI-generated content"?

The definition is broad. Any output produced by a generative AI system: text (articles, posts, descriptions), images (logos, photos, illustrations), audio (synthetic voices, music), video (deepfakes, generated sequences).

Pasquale Pillitteri analyzes precisely what the AI Act requires for AI-written articles. The comparison with US and UK approaches is enlightening: the EU is the strictest, with a disclosure obligation that leaves no room for interpretation.

The Code of Practice: this voluntary playbook that will become the de facto standard

The code published by the Commission is not legally binding in itself. But it is the roadmap validated by the AI Office. In practice, if you don't follow it and you are audited, you will have no credible defense.

The code covers three pillars: machine-readable watermarking, standardized metadata, and automatic disclosure visible to the end user.

Machine-readable watermarking

This is the most technical topic. On August 2, 2026, machine-readable watermarking of AI-generated visuals becomes legally mandatory.

The problem: most legacy Digital Asset Management (DAM) systems were designed to store and distribute assets, not to modify them on the fly with invisible signatures. There is a structural design gap. Companies relying on a traditional DAM will either have to replace it or add a middleware layer dedicated to watermarking.

The watermarking must be resistant to compression, cropping, and common transformations. It is a demanding technical standard that is not yet uniform across the industry.

Metadata and automatic disclosure

Beyond invisible watermarking, the code provides for metadata attached to the content: model used, generation date, parameters. And a visible disclosure — a label, a tag, a mention — that the end user can read.

For companies that automatically generate content with AI, this means rethinking the entire publishing chain. Each piece of content must bear its mark of origin.

Who is concerned exactly — Spoiler: almost everyone

The short answer: if you publish AI content in the EU, you are concerned. The long answer is more nuanced but nothing to be reassured about.

Model providers (OpenAI, Google, Anthropic, etc.)

They must integrate labeling mechanisms into their models. When GPT-5.5 generates an image or Claude Opus 4.7 writes an article, the system should theoretically be able to apply a watermark and metadata to it.

In practice, providers deliver APIs. The responsibility for labeling is shared between the provider (who must make the feature available) and the deployer (who must activate it).

AI startups and SaaS tools

This is where it pinches. A startup that builds a content generation tool with the Gemini 3.1 Pro API must ensure that the final output is labeled before serving it to the user. You cannot say "OpenAI/Google handles that".

The strategic guide from AI in Europe details the concrete steps for companies: auditing all touchpoints where AI content is published, setting up labeling pipelines, and compliance testing.

User companies (not just tech)

An e-commerce site that uses AI to generate product descriptions. A communications agency that produces articles with GPT-5.4. A marketing agency that creates visuals with AI. All are "deployers" in the sense of the AI Act.

DS Solutions points out that chatbots and generative content are classified as "limited risk", which does not exempt them: on the contrary, their main obligation is exactly this transparency.

What devs need to implement technically

Concretely, here is what your tech stack must be able to do by August 2026.

Watermarking pipeline for visuals

Every generated image or video must go through a watermarking module before storage and distribution. This module must insert a machine-readable signal into the file that is resistant to transformations.

If you are using a DAM, check right now if it supports programmatic watermarking on ingestion. If the answer is no, start looking for an alternative. Most current DAMs will not be able to do this because of their architecture.

Standardized metadata injection

Every generated piece of content must carry structured metadata. The exact format is not 100% locked in yet, but the code of practice gives clear directions: model identifier, timestamp, AI generation indicator.

For devs, this means a middleware between your AI generation layer and your publishing layer. Nothing technically revolutionary, but it requires thinking about it now rather than in July 2026.

Frontend disclosure

On the end-user side, a visible label is required. An "AI Generated" badge, a mention in the footer, a visual indicator on images. The exact form is left to your discretion, but the obligation for clarity is not.

BuildThisNow insists: the disclosure must be "easily accessible" and "understandable to a non-technical audience". No size 6 disclaimer at the bottom of the page.

The model landscape: where does native labeling stand?

Not all models are equal when it comes to this requirement. Here is where the main generalist and agentic models available as of June 2026 currently stand.

Generalist models — scores and context

Google has a clear advantage with SynthID. For all others, the labeling work falls on the deployer. This is a critical point to consider when choosing a model, especially if you are targeting the European market.

Agentic models — an additional challenge

Agentic models pose a specific problem: they generate content autonomously, often in multiple steps, sometimes without direct human supervision. How do you label each output of an agent that produces dozens per minute?

Agents like the one described in Qu'est-ce qu'OpenClaw ? L'agent IA qui change tout perfectly illustrate the problem: the more autonomous the agent, the more the labeling chain must be automated and robust.

Fines and risks: €35M or 7% of global revenue

The AI Act's sanctions are among the toughest in the world for tech regulation. For the transparency obligations of Article 50, the cap is 35 million euros or 7% of annual global revenue, whichever is higher.

euaiact.com points out that the EU's unified framework targets safe, transparent, traceable, and non-discriminatory AI systems. Transparency is not a cosmetic add-on — it is a pillar of the regime.

In practice, national supervisory authorities (which are just starting to become fully operational) will have the power to conduct investigations, request audits, and impose these fines. The first likely targets: large providers and content distribution platforms, not isolated micro-enterprises. But the law makes no size distinction for the obligation itself.

Comparison with other regulations

The movement is global. While the EU moves forward on labeling, New York sent 7 AI laws to the governor, including a ban on surveillance pricing. In the US, regulation is done state by state, without a unified federal framework. At the international level, discussions continue — as evidenced by the historic meeting at the G7 summit in Évian with Altman, Amodei and Hassabis, a first of its kind.

The European approach remains the most structured and the most advanced in terms of concrete implementation.

Compliance Timeline — 15 weeks, that's short

If you are reading this article in June 2026, there are about 15 weeks left before the deadline. Here is a realistic timeline.

Weeks 1-3: Audit

Map all the points where AI content is generated and published in your organization. This includes obvious uses (generating articles, images) but also hidden uses (automatic suggestions, translations, auto-summaries).

Weeks 4-7: Technical Architecture

Define your labeling pipeline. Choose your watermarking tools. Integrate metadata injection into your publishing workflows. This is where your hosting choice matters — a flexible infrastructure like Hostinger can make it easier to rapidly deploy custom solutions.

Weeks 8-11: Implementation and Testing

Develop, integrate, test. Verify that the watermarking survives transformations. Validate that the metadata is indeed present in the final files. Test the frontend disclosure on different devices.

Weeks 12-15: Deployment and Documentation

Go live. Write your compliance documentation. If an auditor comes asking how you comply with Article 50, you must be able to demonstrate it both technically and organizationally.

❌ Common mistakes

Mistake 1: Thinking the Code of Practice is optional and therefore ignorable

The code is voluntary, yes. But it is the AI Office's reference document. Not following it amounts to saying "I have no compliance method" in the event of an audit. The distinction between voluntary and mandatory is purely theoretical when facing a regulator.

Mistake 2: Relying on your current DAM for watermarking

This is the most technical and costly mistake. Legacy DAMs have a structural gap. They store and distribute; they do not modify files upon ingestion with invisible signatures. Check now, not in July 2026.

Mistake 3: Focusing solely on images

Article 50 covers all types of AI content: text, audio, video, image. If you label your visuals but forget articles generated by GPT-5.4 or synthetic voices, you are not compliant. Pasquale Pillitteri specifically details the obligation for texts.

Mistake 4: Waiting until the last minute

15 weeks is short for a compliance project that impacts technical architecture, editorial processes, and team training. Companies that start in July 2026 risk being non-compliant by the deadline without being able to invoke any additional grace period.

Mistake 5: Believing that the Digital Omnibus pushed back the deadline

Cookiebot is definitive: Article 50 was not affected by the Omnibus package. The August 2, 2026 deadline stands. This confusion is widely circulated and can lead to fatal inaction.

❓ Frequently Asked Questions

Is a personal blog that uses AI for its articles concerned?

Yes. The AI Act does not make a distinction based on company size for the transparency obligations of Article 50. If you publish AI content in the EU, you must label it, even without revenue.

Must the watermarking be visible or invisible?

Both. The AI Act requires machine-readable watermarking (invisible, detectable by a tool) AND visible disclosure to the end user (label, mention). The Code of Practice details both layers.

What happens if my model provider does not offer native labeling?

It is your responsibility as a deployer. You must add your own labeling layer downstream of the model. This is why the choice of model should integrate this criterion — Google with SynthID has a clear advantage on this point.

Must content generated before August 2, 2026, be taken down or retroactively labeled?

The obligation applies to the dissemination of content. In practice, content already published before the deadline does not seem to be subject to a takedown obligation, but authorities have not yet formally ruled on this specific point. Caution is advised.

Are self-hosted open-source models like Kimi K2.6 or GLM-5 concerned?

Yes. Hosting the model yourself does not exempt you from the obligation to label outputs. On the contrary, it is entirely your responsibility since you are both the provider and the deployer.

✅ Conclusion

August 2, 2026 is approaching fast, and the Code of Practice published by the Commission has just transformed an abstract obligation into a concrete playbook. Machine-readable watermarking, metadata, visible disclosure: the three technical pillars are clear. Now it's time to implement them. If you haven't yet audited your AI content pipelines, now is the time to start — le guide stratégique d'AI in Europe is a good starting point to structure your approach.