AI Labeling Requirement 2026
From 2 August 2026, AI-generated content in the EU must be labeled transparently. This guide explains in plain terms who is affected, what exactly must be labeled, which exceptions apply – and how to comply without effort.
What is the AI labeling requirement?
The AI labeling requirement stems from Article 50 of the EU AI Act. It requires that people can recognise when they interact with AI or face AI-generated content. The goal is transparency: no one should unknowingly take a machine's output for genuine.
Four areas are covered: AI chatbots (50(1)), synthetic media and deepfakes (50(4)), AI text on matters of public interest (50(4)), and emotion recognition / biometric categorisation (50(3)).
Who is affected?
The duty applies regardless of company size – from sole traders to corporations. What matters is not how big you are, but whether you use AI whose outputs reach people in the EU.
The distinction between provider (develops/markets AI) and deployer (uses AI) is key. Most website operators are deployers – the disclosure duties of 50(1), (3) and (4) apply to them, not the providers' technical marking duty under 50(2).
Label chatbots (Art. 50(1))
If you run a chatbot or AI assistant that talks to users, they must be able to recognise they are chatting with AI – unless it is obvious. The notice must appear at first interaction in the chat itself; a line in the privacy policy is not enough.
This is the clearest and fastest duty to meet: no exception debate, a clear statement.
Images, video & deepfakes (Art. 50(4))
For deepfakes – realistic depictions of real people or events – a disclosure duty applies. For artistic, satirical or fictional works it is relaxed.
Ordinary AI images without a deepfake character do not automatically trigger this media duty for you as a deployer. The machine-readable marking inside the image (watermarking) is the AI provider's duty, not the website operator's.
AI text – and the crucial exception
AI-generated text only needs labeling if published to inform the public on matters of public interest (e.g. news, health, politics).
And even then the duty drops away if the text underwent human editorial review and a person holds responsibility. An ordinary company blog or product text with human control is therefore usually not affected – important to avoid over-labeling.
Deadlines: what applies when?
2 August 2026: The Art. 50 transparency obligations become applicable. This deadline was not postponed by the Digital Omnibus (finalised June 2026).
Only the high-risk obligations were postponed (to December 2027 / August 2028). The only nuance for Art. 50: a transition period until 2 December 2026 for the machine-readable provider marking of systems placed on the market before 2 August 2026.
Fines – and the cease-and-desist risk
Violations of Art. 50 can be fined up to €15 million or 3% of worldwide annual turnover (for SMEs the lower amount). The often-cited €35 million only applies to prohibited practices under Art. 5, not to labeling.
At least as relevant in practice: in Germany the duty likely qualifies as a market-conduct rule (§ 3a UWG) – meaning competitors and associations can send cease-and-desist letters, often faster than an authority. Documented labeling protects on both fronts.
How to comply – without effort
- With one script tag (or the free WordPress plugin), AIActify labels AI content automatically – with the official EU AI icon.
- The scoping assistant checks per content item whether a duty exists at all, and documents exceptions verifiably.
- Chatbot notices appear automatically, AI systems land in the inventory, and a compliance report proves everything, ready for authorities.
Six myths about AI labelling
Plenty of half-truths are circulating around 2 August – some of them in press articles. Here is what the law actually says.
"Violations can cost up to €35 million or 7% of turnover."
For breaches of the Art. 50 transparency obligations the ceiling is €15 million or 3% of worldwide annual turnover (Art. 99(4)(g)). The €35 million / 7% figures apply solely to the prohibited practices under Art. 5 – mixing these up is the most common error in the market.
"Every AI-written text has to be labelled."
No. The text duty only applies to publications informing the public on matters of public interest – and it falls away when a human has substantively reviewed the content and holds editorial responsibility. Ordinary marketing and product copy under human control is usually out of scope entirely.
"The deadline has been postponed."
Only the high-risk obligations moved (Annex III to December 2027, Annex I to August 2028). Art. 50 applies unchanged from 2 August 2026. The widely cited December 2026 date concerns only machine-readable marking by providers of pre-existing AI systems – not the duties of website operators.
"A watermark or metadata is enough."
Not for deployers. Disclosure under Art. 50(4) must be perceptible without technical aids – visible on the content itself. Machine-readable marking is a separate obligation and applies to the providers of the AI systems.
"A note in the terms or footer will do."
The Commission's final guidelines (20 July 2026) explicitly call that insufficient. The notice must appear at the first interaction or first exposure to the content – for chatbots that means inside the chat, not in the privacy policy.
"All existing content must be labelled retroactively."
For images, audio and video the creation date governs: content produced before 2 August 2026 does not need retroactive labelling. Careful with texts on matters of public interest – there the publication date is what counts.
Frequently asked questions
When does the AI labeling requirement apply?+
The transparency obligations under Art. 50 EU AI Act apply from 2 August 2026. This deadline was NOT postponed by the Digital Omnibus (finally adopted in June 2026) – only the high-risk obligations were pushed to late 2027/2028.
Does the EU AI Act also apply to small businesses?+
Yes. The Art. 50 labeling requirement applies regardless of company size. For SMEs and start-ups there are only fine reductions (the lower amount applies), but the obligation itself stands.
Do I have to label AI-generated text?+
Not always. The labeling duty for AI text (Art. 50(4)) only applies to text published to inform the public on matters of public interest – and it drops away when a human has editorially reviewed the content and holds responsibility. Ordinary marketing or product copy with human control is therefore usually not affected.
Does a chatbot have to be labeled as AI?+
Yes. Under Art. 50(1) users must be able to recognise that they are interacting with AI – unless that is already obvious. The notice must appear at first interaction in the chat itself; a line in the privacy policy is not enough.
What happens if I don't label AI content?+
Violations of Art. 50 can be fined up to €15 million or 3% of worldwide annual turnover (for SMEs the lower amount). In Germany there is also a realistic risk of cease-and-desist letters from competitors and consumer associations, as the duty likely qualifies as a market-conduct rule.
Do I have to label AI images?+
As a deployer, the media disclosure duty under Art. 50(4) targets deepfakes specifically – realistic depictions of real people or events. Ordinary AI images without a deepfake character do not automatically trigger this duty for you as a mere deployer. The machine-readable marking inside the image is the AI provider's duty, not the website operator's.
This guide is for general information and is not legal advice. For a binding assessment in your specific case, consult a specialised law firm.
Ready for the EU AI Act?
The AI labeling obligation (Art. 50) applies from 2 August 2026. Start implementing now – in just 2 minutes.
No credit card required · 15 days free trial