netzstrategen AI Operations.
Insights Glossary Resources Deutsch Self-Check
← All posts
Governance

AI Content Disclosure: What Article 50 of the AI Act Requires Since 2 August 2026

Published on 8/3/2026 · André Hellmann

AI Act Disclosure Compliance Transparency Governance

The AI Act transparency obligations have applied since 2 August 2026 (Source: European Commission, 2026). Companies publishing AI-generated text must disclose it under certain conditions. Most corporate websites have changed nothing. This article explains AI content disclosure from the deployer’s perspective: the four cases, when the exemption applies, and what the notice actually looks like.

Where do you stand?

Discuss your next step in a free diagnosis call. Book a slot →

One note upfront: this article is not legal advice. The wording examples are meant as orientation. Before going live on your own site, the exact text belongs on a lawyer’s desk.

Contents

The four cases under Article 50

Article 50 of the EU AI Act governs transparency, not risk. The obligations apply regardless of a system’s risk class. Four constellations are covered (Source: European Commission, Quick Facts on transparency rules, 2026):

CaseWho owes the dutyExample
Direct interaction with an AI systemProviderChatbot or voicebot on a website
Synthetic content (text, image, audio, video)Providermachine-readable marking of generated content
DeepfakesDeployerAI-generated image of a real person
Text on matters of public interestDeployerAI-written article, press text, newsletter

For most companies, the fourth row is the relevant one. Anyone using ChatGPT, Claude or Copilot to write articles is a deployer — and therefore on the hook for deepfakes and for text on matters of public interest.

“Public interest” is deliberately broad: socially, politically, economically or culturally relevant topics, published through generally accessible channels. An article on AI investment in the mid-market falls under it in case of doubt. An internal email to three colleagues does not.

Provider or deployer: who owes what

The role question determines the effort. It is often answered wrongly.

Providers develop AI systems, or have them developed, and place them on the market under their own name. Their duty: apply a machine-readable mark to synthetic content and enable its detection. Exempt are systems performing an assistive function for standard editing, or that do not substantially alter the input data or its semantics (Source: European Commission, 2026).

Deployers use third-party AI systems under their own responsibility. Their duty: disclose deepfakes and mark text on matters of public interest with clear and perceivable labels.

The practical consequence is convenient: the watermarking work sits with OpenAI, Anthropic, Google and Microsoft — not with the company using the tools. Anyone consuming models through an API does not need to generate C2PA metadata themselves. The deployer’s duty is editorial, not technical.

Watch one edge case: a company that has a system built by a third party and runs it under its own name can itself become a provider. Then the provider obligations apply on top.

The exemption for editorially reviewed text

This is the point that decides the workload in practice. The disclosure obligation for text does not apply where the text has undergone a process of human review or editorial control and a natural or legal person holds editorial responsibility for the publication (Source: European Commission, 2026).

Two conditions, both must be met. And both are stricter than they sound.

First: genuine review. A spelling or formal check is not enough. What is required is a deliberate substantive review of accuracy, plausibility and sources — with a real option to change or reject the text. Reading an AI draft and waving it through is not reviewing.

Second: named responsibility. There must be a person or organisation accountable for the publication. “The editorial team” does not suffice. A name does.

Do I have to disclose this text? Decision path for text under Art. 50(4) AI Act Generated or altered by AI, in whole or in part? Published for the public, on a matter of public interest? Without human review and without named editorial responsibility? No disclosure obligation under Art. 50(4) Disclosure is mandatory Exemption applies — disclosure is voluntary no no yes no Simplified · Source: European Commission, Art. 50 AI Act, 2026 · not legal advice netzstrategen
The decision path for text: only when all three questions apply is disclosure mandatory. Where the editorial exemption applies, disclosure stays voluntary — and for most companies that is the normal case.
For presentations:

What matters is what follows from the exemption: it is not a free pass but a process requirement. Using it means having a documented flow — AI draft, substantive review, sign-off by a named person. Without that record, the exemption is worth little in a dispute.

The exemption does not require a mistake to be found. It requires a real option to reject the text.

Implementation in articles and blogs: three examples

The legal position turns into a design question. Three variants have proven workable, from lean to detailed.

Variant 1 — the short line. For pieces where AI only assisted:

Produced with AI support. Editorial responsibility: First Last.

Variant 2 — the separated notice block. For formats that regularly use AI:

This article was produced with AI support. Substantive review and approval rest with a human. Editorial responsibility is held by First Last.

Variant 3 — the itemised breakdown. For publications that make transparency a principle: the notice block states separately which steps sat with the human and which with the AI — for instance topic selection, source selection, fact-checking and approval on one side; research, drafting, diagrams and publishing on the other. Plus a link to a page describing the process in full.

We use the third variant on this hub. How the process behind it works in detail is covered in a dedicated article.

What counts in all three cases: the notice must be clear, unambiguous and perceivable — not buried in body copy, not in small print, not on a subpage. For text, the place is the beginning or a clearly visible line at the end. And it belongs together with a statement of who holds editorial responsibility.

Where your company stands on AI governance overall can be mapped in minutes with the free Self-Check.

The EU icons and how to place them

Since June 2026 the Commission provides three icons for labelling AI-generated content: a basic icon, one for fully AI-generated content and one for partially AI-modified content. Each comes in four variants — black, white, and both at 50 percent transparency — as SVG and PNG, free to use without attribution (Source: European Commission, EU Icons, 2026).

One user-testing result matters in practice: the icon performs markedly better when accompanied by a text label. A symbol alone is misread more often.

Three placement rules apply (Source: European Commission, 2026):

  • The notice must be clearly perceivable at the latest on first exposure to the content.
  • It must not be covered by intervening overlay elements.
  • It must remain visible when the content is reshared or downloaded.

Accessibility recommendations come on top: sufficient size, plain language, no abbreviations other than “AI”, alt text or ARIA labels for screen readers. Anyone offering a second layer with details must make it navigable with assistive technologies.

Two things should not be confused. The icons are optional — the disclosure obligation is not. And using an icon does not by itself establish legal compliance; responsibility for an adequate notice stays with the deployer.

What the Digital Omnibus moved — and what it did not

This is where the biggest misunderstanding is forming. The Digital Omnibus moved deadlines in the AI Act — the requirements for high-risk systems shifted back. Article 50 is not affected. The transparency obligations have applied since 2 August 2026 as originally planned.

There is one relief, and it concerns the provider side: for generative AI systems placed on the market before 2 August 2026, a grace period for the marking obligation under Article 50(2) runs until December 2026 (Source: European Commission, 2026). For deepfakes created before 2 August 2026 there is no mandatory retroactive labelling — though it is encouraged.

Deployers publishing text cannot invoke any of these grace periods. For the obligations under paragraph 4, the date is simply 2 August 2026.

Enforcement runs through the national market surveillance authorities, the AI Office and — for EU institutions — the European Data Protection Supervisor. Fines reach up to €15 million or 3 percent of total worldwide annual turnover, with proportionality applied to SMEs (Source: European Commission, 2026).

Conclusion: a process question, not a copy question

Disclosure looks like a wording task. It is an operating task.

Using the exemption requires a demonstrable flow with named responsibility. Not using it requires a visible notice on every affected piece of content — consistent, accessible, and intact when shared. Neither is a copy decision. Both are questions of roles, sign-offs and documentation.

The pragmatic route combines the two: set the process up so the exemption applies, and disclose anyway. That over-delivers on the obligation, leaves nothing to attack, and builds trust with readers who will ask regardless. For European companies that is not a drawback but an argument — precisely the point of digital sovereignty.

That puts disclosure where governance belongs: in continuous operations, not in a one-off project.

Frequently asked questions

Does every AI-written text need to be disclosed?

No. The obligation under Article 50(4) covers text published to inform the public on matters of public interest. Internal communications and text without that connection are out of scope. And the editorial exemption can remove the obligation entirely.

Is it enough if someone reads the AI text before publishing?

No. What is required is a deliberate substantive review of accuracy, plausibility and sources — with a real option to change or reject the text. A spelling check does not qualify. In addition, a named person must hold editorial responsibility.

Are the EU icons mandatory?

No. The icons are optional; the disclosure obligation is not. They make presentation consistent but do not establish compliance on their own. User testing shows an icon is understood far better when a text label accompanies it.

Did the Digital Omnibus postpone the disclosure obligation?

No. What was postponed were the high-risk requirements. Article 50 has applied unchanged since 2 August 2026. Providers get a grace period until December 2026 for the marking obligation on systems placed on the market before that date — deployer obligations under paragraph 4 do not.

How should a company set this up?

Through the process, not the copy: named responsibility, documented review before publication, consistent notices in the CMS. We map out where the gaps sit in a free diagnosis call.

Sources

What's next