Contentoren

Compliance & funding

AI Act Article 50: Labelling Chatbots, Websites & Ads

AI Act Article 50 explained: what applies since 2 August 2026 to chatbots, AI images, ads and text, with a labelling matrix and copy-paste disclosures.

20 min readby Contentoren Editorial Team
Share
A designer marks AI-generated images on printed campaign proofs while the art director compares two prints.

Key takeaways

In short:

  • Article 50 of the EU AI Act sets out transparency obligations that, according to the European Commission, have applied since 2 August 2026. The Digital Omnibus did not postpone them.
  • Chatbots, WhatsApp bots and voice AI must say clearly, at the latest at the start of the conversation, that they are AI. A line in your terms or privacy notice is not enough.
  • Visible labels are only required for deepfakes (realistic AI images, video and voices) and for AI text on matters of public interest that nobody has reviewed. Not every AI image or text needs a label.
  • Machine-readable marking (watermarks, metadata) is the job of AI providers. Systems already on the market before 2 August 2026 have until 2 December 2026 (Regulation (EU) 2026/1744).
  • Below you will find a labelling matrix for twelve common use cases and disclosure wording you can copy.

Two myths have spread since August 2026. One says every AI image now needs a warning. The other says the rules only affect big tech. Neither is true. This guide turns Article 50 and the Commission guidelines of 20 July 2026 into concrete decisions for your website, customer service and marketing. Status: November 2026. This article is not legal advice. It does not replace a review of your specific case by a lawyer.

What is the AI labelling obligation under Article 50?

The AI labelling obligation is the duty to tell people openly when they are talking to an AI, are exposed to AI-generated content, or are subject to certain AI analysis. The legal basis is Article 50 of Regulation (EU) 2024/1689, the AI Act. It contains four obligations addressed to two different roles:

Paragraph What is covered? Who is obliged? What must be done? Main exception
Art. 50(1) AI systems interacting directly with people (chatbot, voicebot, AI agent, avatar) Provider Design the system so people are informed they are interacting with AI The AI nature is obvious to a reasonably well-informed person
Art. 50(2) AI-generated or manipulated text, images, audio and video Provider Mark outputs in a machine-readable format and make them detectable Assistive standard editing; no substantial alteration
Art. 50(3) Emotion recognition, biometric categorisation Deployer Inform the people exposed to the system Use permitted by law for criminal investigations
Art. 50(4) Deepfakes, and AI text informing the public on matters of public interest Deployer Visibly disclose that the content is AI-generated or manipulated Art, satire, fiction (lighter regime only); text with human review and editorial responsibility

Article 50(5) applies to all four: the information must be clear and distinguishable, given at the latest at the first interaction or exposure, and accessible (guidelines, section 7).

Provider or deployer: which role do you have?

  • A provider develops an AI system, or has one developed, and places it on the market or puts it into service under its own name. This applies whether it is free or paid and whether the company sits inside or outside the EU. Important for businesses: if you build a chatbot in-house and use it under your own name, the Commission treats you as a provider. The same applies if you modify someone else's generative system, for example with your own training data, and run it under your brand.
  • A deployer uses an AI system professionally under its own authority. A company remains the deployer even when freelancers do the work on its behalf. Individual employees are not separate deployers.
  • Agencies: a company that merely commissions an advertising agency, without deciding whether or how the agency uses AI, is not a deployer according to the guidelines (para. 14). The agency then carries the duty. If you commission AI images specifically or create them yourself, you are the deployer.
  • Platforms such as Instagram or YouTube that only distribute your content are not deployers in that respect. They may, however, give you labelling tools (see below).

In practice, many businesses are both: provider of their self-built website chatbot and deployer of the image generators used in marketing.

When does it apply, and what did the Digital Omnibus change?

Article 50 has applied since 2 August 2026. The Digital Omnibus on AI, Regulation (EU) 2026/1744 of 8 July 2026 (published in the Official Journal on 24 July 2026), mainly postponed the high-risk obligations. For Article 50 it added only a narrow transitional rule:

Date What applies
10 June 2026 Final Code of Practice on Transparency of AI-generated Content published
20 July 2026 Commission guidelines on Article 50 published
2 August 2026 Art. 50(1), (3), (4) and (5) apply to everyone; Art. 50(2) applies to newly placed systems
2 December 2026 Grace period ends for machine-readable marking (Art. 50(2)) by generative systems placed on the market before 2 August 2026 (new Art. 111(4), Reg. 2026/1744)

Three points are often misreported:

  1. The grace period does not cover chatbots. A system that both converses and generates content benefits only for the marking duty. The in-conversation AI disclosure has been mandatory since 2 August 2026 (guidelines, para. 153).
  2. No retroactive labelling. Content generated and published before 2 August 2026 does not have to be labelled after the fact, although the Commission encourages it (Q&A).
  3. One exception to that: AI text on a matter of public interest that was generated before the deadline but published afterwards must be labelled if it was not reviewed by a human (guidelines, para. 154).

For the full AI Act timeline to 2028, see our EU AI Act timeline 2026–2028 after the Digital Omnibus.

Chatbots, WhatsApp and voice bots: what must the disclosure look like?

Article 50(1) covers any AI system capable of a genuine two-way exchange that communicates directly with people, in writing, by voice or as an avatar. The Commission explicitly lists customer-service and e-commerce chatbots, voice assistants, AI hotlines and AI agents. Out of scope are rule-based auto-replies such as traditional out-of-office emails, recommender systems, spam filters, and pure translation or transcription tools (guidelines, section 3.1.1).

Hybrid set-ups are covered too. If AI and human answers are mixed in one support channel, the AI answers must be identifiable. The only exception is when humans properly review the AI answers and send them themselves. An AI tool that only drafts suggestions for your service team is not covered by Article 50(1).

What the Commission says is not enough

  • a disclosure only in terms and conditions, a privacy notice or documentation
  • metadata or invisible watermarks alone
  • vague labels such as "your digital assistant", or a human profile photo that could mislead
  • blanket statements like "This website uses AI"
  • purely technical statements such as "powered by LLMs"
  • an audio tone on the phone without a spoken statement

What works

The guidelines recommend a clear notice before or in the first message, ideally combined with a persistent badge next to the input field. Voice AI needs a spoken statement at the start of the call, plus reminders in longer conversations. In sensitive contexts such as finance, insurance, legal or health questions and complaints, the Commission expects repeated notices. If someone asks "Am I talking to a human?", the system must always answer truthfully.

AI agents that write emails, make bookings or negotiate on your behalf must disclose that they are AI and on whose behalf they act (guidelines, para. 31).

The "obvious" exception is to be read narrowly. The Commission's examples are a code assistant used only by developers, or an internal tool for trained staff. It does not apply to a chatbot on a public website, where older people, children and users with little AI experience also write in. For internal tools, the AI literacy obligation under Article 4 matters most: only trained staff recognise AI reliably.

How a chatbot qualifies enquiries, and which GDPR points come on top, is covered in AI chatbot for website leads. This article is only about labelling.

Which AI images, video and text need a visible label?

A deepfake, under Article 3(60) of the AI Act, is AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear to be authentic. The Commission reads this broadly: it is enough that what is shown could plausibly exist. A photorealistic AI "customer" who does not exist can therefore be a deepfake. Talking mice, dragons or a sphinx flying over the Eiffel Tower are not.

The examples in the guidelines (section 6.1) draw clear lines for marketing:

  • Label: an AI video of a realistic synthetic influencer testing a real product; an AI avatar of the CEO congratulating staff; an AI product image that makes the product look better, higher quality or different from reality; teleshopping-style spots with simulated customers.
  • Usually no label: a real product in front of an AI background, as long as the background does not mislead about the product; colour correction, cut-outs, background clean-up; clearly fictional or cartoon-style motifs.
  • Art, satire, fiction: a disclosure that does not spoil the work is enough, for example in the credits. Advertising, however, only counts as "evidently creative" in specific cases. Where content mixes informative and creative elements, the informative character prevails.

The German advertising federation ZAW, for example, advises advertisers to label all AI depictions that viewers could take as real. According to the ZAW, AI backgrounds that do not mislead about the product need no label.

AI text: label only for public-interest topics without review

For text, Article 50(4) only applies if three criteria are met together: published, intended to inform the public, on a matter of public interest such as politics, health, the environment, consumer safety or finance. Even then, no label is needed if

  1. a knowledgeable person reviewed the substance, including a fact check (a spell check or a quick sign-off does not count), and
  2. a person or company holds editorial responsibility and names it in an easily findable place, such as the legal notice or imprint.

The Commission explicitly lists advertising copy and product descriptions as out of scope, as long as they contain no claims about health, consumer safety or sustainability. One trap: if a text is substantially rewritten by AI after sign-off, the exception is lost (guidelines, para. 136). An unreviewed AI article on diet and disease, by contrast, would need a label. The Commission uses exactly that example.

Emotion recognition and biometrics (Art. 50(3))

Article 50(3) rarely matters for ordinary websites. It covers, for example, cameras that sort visitors by age, or software that infers emotions from biometric data such as faces or voices. If you use such tools, you must inform everyone exposed, for instance with a sign at the entrance or a pop-up. Emotion recognition in the workplace and in education is generally prohibited under Article 5. Text analysis without biometric data is not covered.

Labelling matrix: what applies to your use case?

The matrix summarises the guidelines for twelve typical cases. It does not replace a case-by-case review. Borderline cases are flagged.

Use case Duty? Who? Recommended approach
AI chatbot on your website Yes (50(1)) Bot provider; you, if built in-house Notice in the greeting plus a persistent "AI" badge on the chat window
WhatsApp or messenger bot Yes (50(1)) As above First message with AI notice, "AI" in the profile name
Phone or voice bot Yes (50(1)) As above Spoken statement in the first seconds, reminder on transfer
AI agent emails customers Yes (50(1)) Provider AI notice at the top of the email, plus the principal
Employee uses AI drafts and sends them personally No – No label needed; document internal sign-off
Illustrative, clearly artificial AI graphic No – Keep metadata; optional "Illustration: AI"
Photorealistic AI team photo or AI "customers" Yes, deepfake Deployer (you) Visible label in the image, e.g. EU icon with "AI-generated"
AI product shot that alters or enhances the product Yes, deepfake Deployer Label directly in the image
Real product photo in front of an AI background Usually no – Label only if the background could mislead about the product
AI video ad with realistic people, voices or avatar influencers Yes, deepfake Deployer or agency Label from the first second; spoken at the start for audio
Blog posts, newsletters, product copy with AI help, human-reviewed Usually no – Document review process and responsible person
Unreviewed AI text on health, finance, politics etc. Yes (50(4)) Deployer Notice at the start of the text

Borderline case, website chatbot: who is liable if you embed a SaaS vendor's chatbot? The vendor must make the disclosure technically possible. But if you configure the bot under your own brand, with a human name and avatar, and switch the notice off, responsibility can quickly shift to you. Our advice either way: switch the notice on, get it confirmed in the contract and keep a screenshot as evidence.

Disclosure wording you can copy

The Commission does not prescribe any wording. What matters is clarity, placement and timing. In our reading of the guidelines, these texts meet the requirements:

Website chat (greeting):

Hi! I'm an AI assistant, not a person. I can answer questions about our services. If you'd rather talk to our team, just type "human".

Chat badge: "AI assistant" or "Automated answers by AI"

WhatsApp (first message):

You are chatting with [Company]'s AI assistant. Replies are generated automatically. For personal advice, reply "team".

Voice AI (opening line):

Hello, this is [Company]'s AI assistant. You are speaking with an artificial intelligence. To talk to a member of staff, just say "agent" at any time.

Email from an AI agent (first line):

This message was written and sent by an AI assistant on behalf of [Company].

Image or video (label in the visual): "AI-generated"; for partly altered footage "Modified with AI"; for audio spots, spoken: "This ad contains AI-generated voices."

Unreviewed text on a public-interest topic: "This text was created with AI and has not been editorially reviewed."

For images and video, the EU provides free icons: a basic icon, one for "fully AI-generated" and one for "partially AI-modified", in black and white variants. Using them is optional. According to the Commission, the icons performed clearly better in user testing with a short text label. Place the label in the content itself, keep it visible when the content is shared or downloaded, and make it readable for screen readers via alt text or an ARIA label. If you publish in several EU languages, use the local language: "KI-generiert" in German, "Généré par IA" in French.

Are the labels from Meta, TikTok and Google enough?

The platforms now label content themselves, but not always where Article 50 requires it:

Platform What the platform does What you should do
Meta (Facebook, Instagram) "AI info" label in "About this ad" behind the three-dot menu, including automatic detection of third-party tool signals (Meta, update 1 June 2026) For deepfake visuals, add a label inside the visual as well
TikTok Requires labels on realistic AI content, offers creator labels, reads C2PA metadata and is testing invisible watermarks (TikTok Newsroom) Use the creator label; do not strip C2PA data on export
Google Ads (Search, YouTube, Discover) "How this ad was made" section in My Ad Center since 9 July 2026; automatic disclosure for Google's AI tools, advertiser control for third-party tools; depending on local requirements, also a label on the ad itself (Google) Declare AI use correctly in the campaign

The Commission allows deployers to rely on labelling tools from very large platforms if those tools provide a clear and distinguishable disclosure (guidelines, para. 126). A label that only appears after a click into the three-dot menu is unlikely to qualify: section 7 of the guidelines explicitly says that information "hidden under layers of menu options" is not clear. The ZAW also says the label must be visible at first glance and must travel with the visual. Our advice: use the platform label and, for deepfakes, add one inside the visual.

What are the penalties, and who enforces?

  • Fines: up to €15 million or 3% of worldwide annual turnover, whichever is higher. For SMEs and start-ups, the lower of the two applies (guidelines, para. 152).
  • Enforcement: national market surveillance authorities enforce Article 50, on their own initiative or after complaints. The AI Office only steps in for systems built on a general-purpose AI model from the same provider or integrated into very large platforms (Q&A). In Germany, for example, the Federal Network Agency (Bundesnetzagentur) is the central contact and complaints point, while the state media authorities remain responsible for media (BNetzA, 29 July 2026). Check which authority your Member State has designated.
  • Unfair competition: Article 50 can have a second enforcement route through national unfair-competition law. In Germany, the ZAW considers it possible for competitors and qualified associations to send cease-and-desist letters (Abmahnungen) for breaches of Article 50(4) and (5) under section 3a UWG. We are not aware of any court rulings on this as of November 2026. For unlabelled deepfake ads, however, this risk is probably more concrete than a fine.
  • Other law still applies: an AI label does not make misleading advertising lawful. Where real people are depicted, personality rights, the GDPR and copyright also apply. The Unfair Commercial Practices Directive continues to apply alongside Article 50.

Checklist: implementing AI labelling in 7 steps

  1. Inventory: list every AI touchpoint: website chat, WhatsApp, phone, email agents, image generators, video tools and writing assistants. Note who provides each tool.
  2. Clarify roles: for which systems are you the provider (in-house build, own brand), and for which the deployer?
  3. Check conversational systems: is the AI notice in the first message or greeting? Is the badge visible and accessible? Does the bot answer "Are you human?" correctly?
  4. Classify content: ask the deepfake question for every AI image and video. Does it look real, and could it mislead? If so, label it in the visual.
  5. Define the text process: substantive review with fact check before publication, no AI rewrites after sign-off, responsible person named in the imprint.
  6. Bind agencies and vendors by contract: who decides on AI use, who labels, who delivers C2PA metadata?
  7. Document and train: keep screenshots of notices, classifications and sign-offs. Train your team as part of the AI literacy obligation. After 2 December 2026, check that your tools mark their output in a machine-readable way.

Conclusion: label precisely, not everywhere

Article 50 is a transparency rule with a clear core. If an AI talks to people, it must say so straight away. If you publish realistic AI images, video or voices that could mislead, you must label them visibly. For most businesses this means three things: give chatbots, WhatsApp and voice bots proper disclosures now, label deepfake visuals in marketing consistently, and set up a documented review process for text. Blanket AI warnings on every page help no one.

If you want to introduce a chatbot or phone assistant with correct AI disclosure, see our AI chatbots & phone assistants service. For assessing your processes under the AI Act and the GDPR, our privacy, GDPR & EU AI Act consulting can help (day rate €1,200 incl. VAT). It does not replace legal advice. Book a no-obligation call.

Sources

All sources accessed 6 November 2026.

  • Regulation (EU) 2024/1689 (AI Act), Article 50: EUR-Lex
  • Regulation (EU) 2026/1744 (Digital Omnibus on AI), OJ L of 24 July 2026, recital 38 and new Art. 111(4): EUR-Lex
  • European Commission: Guidelines on the transparency obligations under Article 50 (20 July 2026): overview, PDF
  • European Commission: Questions & answers on Article 50: digital-strategy.ec.europa.eu
  • European Commission: Quick Facts on transparency rules: digital-strategy.ec.europa.eu
  • European Commission: Code of Practice on Transparency of AI-generated Content: digital-strategy.ec.europa.eu
  • European Commission: EU icons for labelling AI-generated content: digital-strategy.ec.europa.eu
  • Future of Privacy Forum: The AI Act implementation timeline under the AI Omnibus: fpf.org
  • Bundesnetzagentur: press release on the AI Act and the KI-MIG (29 July 2026, German): bundesnetzagentur.de
  • ZAW: guidance on Article 50(4) and (5) AI Act (20 July 2026, German): zaw.de
  • German Act against Unfair Competition, section 3a UWG: gesetze-im-internet.de
  • Directive 2005/29/EC (Unfair Commercial Practices Directive): EUR-Lex
  • Meta: Expanding GenAI transparency for Meta's ads products (update 1 June 2026): about.fb.com
  • TikTok: More ways to spot, shape and understand AI-generated content: newsroom.tiktok.com
  • Google: Expanding AI transparency in ads (9 July 2026): blog.google

All information without guarantee. This article is not legal advice. The Commission guidelines are non-binding; only the Court of Justice of the EU can give an authoritative interpretation of the AI Act.

Frequently asked questions

When did the Article 50 transparency obligations start to apply?

Article 50 of the AI Act has applied since 2 August 2026. The only grace period concerns the machine-readable marking of AI-generated content under Article 50(2): generative AI systems placed on the market before 2 August 2026 have until 2 December 2026. Chatbot disclosure has no grace period.

Do I have to label every AI-generated image on my website?

No. A visible label is only required for deepfakes: realistic AI images, video or audio that resemble existing people, places, objects or events and would falsely appear authentic. Clearly fictional illustrations, or a real product shown in front of an AI background, are usually not deepfakes according to the Commission guidelines. An AI product shot that makes the product look better than it really is can be one.

Do AI-assisted blog posts need an AI label?

Usually not. The duty only covers published text that informs the public on matters of public interest, and it does not apply if a knowledgeable person reviewed the substance and someone holds editorial responsibility. The Commission explicitly lists advertising copy and product descriptions without health, safety or sustainability claims as out of scope.

Who is responsible: the chatbot vendor or my company?

The provider of the AI system is responsible for the chatbot disclosure and for machine-readable marking. If you build a chatbot in-house or put it into service under your own name, you can be the provider yourself. The deployer, meaning the business that uses AI professionally and publishes the content, is responsible for labelling deepfakes and certain text.

What are the penalties for missing AI labels?

The AI Act allows fines of up to 15 million euros or 3% of worldwide annual turnover; for SMEs the lower of the two applies. National market surveillance authorities enforce the rules, and anyone may lodge a complaint. In some Member States, such as Germany, competitors may also be able to bring unfair-competition claims.