EU AI Labels Are Live From 2 August

BySarah Fraser
eu ai act transparency obligations: Printed EU AI Act document on a desk with office items nearby.

Article 50 of the EU AI Act started applying on 2 August 2026, and that means AI systems serving users in the EU now have live transparency duties. The core change is practical: providers and deployers must now show chatbot notices, add machine-readable markers to AI-generated content, and label deepfakes and certain public-interest AI text under the official Article 50 text and the European Commission’s guidance.

For product teams, this moved from roadmap work to shipping work this week. If you run a chatbot, an image or video generator, or an AI-assisted publishing tool in the EU, Article 50 is now less about policy decks and more about UI copy, metadata pipelines, moderation rules, and audit trails.

Article 50 transparency duties now apply from 2 August 2026

The European Commission says the AI Act’s transparency obligations began applying on 2 August 2026. Article 50 covers four main buckets in practice:

  • Chatbot transparency: users must be informed they are interacting with an AI system, unless that is already obvious from context, under Article 50(1).
  • AI-generated content marking: providers of generative AI systems must ensure outputs are marked in a machine-readable format and detectable as artificially generated or manipulated, under Article 50(2).
  • Deepfake disclosure: deployers must disclose when image, audio, or video content constitutes a deepfake, with limited exceptions, under Article 50(4).
  • Public-interest AI text disclosure: deployers must disclose AI-generated or AI-manipulated text published to inform the public on matters of public interest, under Article 50(5).

The Commission’s 20 July 2026 guidance is interpretive rather than dispositive, and enforcement will still be carried out by market-surveillance authorities. But for companies shipping into Europe, the direction is now plain enough: if your product emits synthetic content or impersonates a human conversation, you need visible disclosure and, in some cases, structured marking under the hood.

Who is covered depends on both role and product shape. The Commission’s quick-facts page frames the obligations around providers and deployers of certain AI systems. Providers are the ones placing systems on the market or putting them into service; deployers are the ones using them in a professional context.

A simple split helps. Providers are on the hook first for system-level marking and product design choices. Deployers are on the hook for the user-facing disclosure that happens in context. If you ship a foundation model API that generates images, your job includes machine-readable output marking. If you run a newsroom tool or customer-service bot on top of that stack, your job includes the notice the user actually sees.

That division is why Article 50 looks less like a single compliance checkbox than a chain of controls. The model layer needs markers. The app layer needs labels. The publishing layer needs rules for when a disclosure is shown, suppressed, or attached to exported content.

Who must label AI content and inform users first

For chatbots, the immediate requirement is the easiest to understand and probably the easiest to miss in edge cases. Users must be told they are interacting with AI before or during the interaction unless the fact is already obvious, according to Article 50(1). A bot named Support Assistant with a visible AI badge is one thing; a human-sounding sales or support flow that never identifies itself is the problem Article 50 is aiming at.

That means teams should now instrument at least four things:

  • a persistent disclosure in the chat UI
  • a fallback disclosure in voice or multimodal interactions
  • logging that shows when and where the notice was presented
  • identity and access checks where impersonation risk is material, especially in workflows that touch accounts or approvals, a pattern already visible in products tightening AI identity and access checks

For generative media, the requirement is more technical. Article 50(2) says providers must ensure outputs are marked in a machine-readable way and detectable as AI-generated or AI-manipulated “to the extent this is technically feasible.” The Commission’s guidelines and code FAQ point toward implementation patterns such as metadata, provenance markers, and related output-labeling mechanisms.

In practice, that means exported files should not just carry a visible badge in the app. They should also carry a marker another system can read later. A watermark that disappears on screenshot is not the same thing as structured provenance data in the file.

Deepfakes and public-interest publishing push the duty further downstream to deployers. If you publish or distribute deepfake audio, image, or video, you now need disclosure at the point people encounter it under Article 50(4). And if you use AI to generate or manipulate text on matters of public interest, disclosure is also required under Article 50(5), unless the content has undergone human review or editorial control and a natural or legal person holds editorial responsibility.

That editorial-control carveout matters for publishers. It narrows the duty for AI-assisted writing that has actually passed through accountable human hands. It does not create a blanket exemption for “AI helped somewhere.”

The same operational logic now hitting labels also shows up in adjacent compliance work such as AI product privacy compliance: the law is not asking for a vague commitment to transparency, but for product behavior that can be shown, logged, and reviewed.

Providers are on the hook first for system-level marking and product design choices. Deployers are on the hook for the user-facing disclosure that happens in context.

The Commission’s code and grace period through 2 December 2026

The most important narrow caveat is timing. The transitional period to 2 December 2026 applies only to the Article 50(2) machine-readable marking duty for certain generative AI systems already placed on the market before 2 August 2026. It does not broadly delay all Article 50 duties.

So if you already had a generative model in market before the deadline, you may have limited extra time for the marking requirement. If you run chatbots, deepfake tools, or public-interest publishing workflows, the rest of Article 50 is already live.

The Commission has also published a voluntary Code of Practice on Transparency of AI-Generated Content, and in a 9 July 2026 opinion said the code adequately covers implementation of Articles 50(2), 50(4), and 50(5). The code can help companies standardize disclosures, labels, and optional EU icons.

But the code is still voluntary, and the Commission’s own FAQ says it does not by itself constitute conclusive evidence of compliance. It is closer to a shared implementation playbook than a safe harbor.

That is the practical takeaway for engineering and product teams this week. A reasonable Article 50 readiness pass now includes: user-facing chatbot notices, export-safe provenance markers, deepfake labels in playback and publishing surfaces, editorial-control decision rules for public-interest text, and retention of evidence showing those controls fired when they should. If your output can cross app boundaries, assume the label must survive the trip.

The Commission’s quick-facts page also notes that penalties can apply for non-compliance. The next step is not another policy memo; it is testing the actual product flow a user will see.

The current implementation framework is the Commission guidance published on 20 July 2026, while enforcement decisions will come from national market-surveillance authorities as cases arise.

Key Takeaways

  • Article 50 of the EU AI Act started applying on 2 August 2026, making its transparency duties live for services offered in the EU.
  • Chatbots must inform users they are interacting with AI, unless that is already obvious from the context.
  • Providers of generative AI systems must add machine-readable markers to AI-generated or AI-manipulated outputs, subject to technical feasibility.
  • Deployers must disclose deepfakes and certain AI-generated public-interest text, with a narrow editorial-control exemption for some published text.
  • The Commission’s Code of Practice is voluntary, and the limited grace period to 2 December 2026 applies only to some pre-existing systems under Article 50(2).

Further Reading