Regulation (EU) 2024/1689 · Artificial Intelligence Act ● Annex III: — days

AI Chatbot Disclosure Requirements Under EU AI Act Article 50

Article 50 transparency duties have been live since 2 August 2026. This is what a customer-facing chatbot, a generated image and a deepfake each have to disclose, and where the wording must appear.

Renomira Team · 9 min read

If a customer cannot tell they are talking to a machine, the EU AI Act says you have to tell them. Article 50 has been applicable since 2 August 2026, and unlike the high-risk regime, it did not wait for 2027.

A public notice board — disclosure is a notice, not a footnote
Photo: Acabashi / CC BY-SA 4.0 via Wikimedia Commons

Article 50 is the most immediately visible part of the Act: it governs how AI systems interact with people, how synthetic content is marked, and what has to be disclosed around emotional inference and deepfakes. It is cheap to comply with and expensive to ignore, because the evidence of non-compliance sits on your own homepage.

The four disclosure duties

Article 50 splits into four practical obligations. Most mid-market companies touch at least two of them.

Duty Trigger Who must do it
50(1) — Interaction disclosure A system interacts directly with people Provider
50(2) — Synthetic content marking Audio, image, video or text that is synthetic or manipulated Provider
50(3) — Emotion recognition / biometric categorisation disclosure The system infers emotion or categorises biometric data Provider
50(4) — Deepfake disclosure Content that is materially altered or fabricated Deployer

Notice the last row: deployers carry the deepfake and manipulated-content duty, not only the vendor. If your marketing team publishes a synthetic video, the obligation is yours even though you did not build the model.

1. “You are talking to a machine”

Article 50(1) requires that people are informed they are interacting with an AI system, unless that is obvious to a reasonably observant person — and where the system generates or manipulates content that looks like a person’s voice or image, disclosure is mandatory regardless of obviousness.

What this means in practice:

  • Website and in-product chatbots: a persistent line in the chat header — “You’re chatting with an AI assistant” — meets the bar. A one-off greeting that scrolls away does not.
  • Voice bots and IVR: an announcement at the start of the call, repeated if the call is transferred.
  • Email auto-responders: “This reply was drafted by AI and reviewed by a person” is the safe formulation.
  • Obvious cases: an offline help article is not an AI interaction. A chat widget named after your assistant, with an avatar, is not “obvious” enough to skip disclosure.

The standard is what a reasonable person would notice, not what your product team knows.

2. Marking synthetic content

Article 50(2) requires providers of systems generating synthetic audio, image, video or text to ensure outputs are marked in a machine-readable format and detectable as artificially generated or manipulated — and, for the deployer, that the marking is retained and clearly legible where fraud risk is highest.

Two layers here:

  • Machine-readable marking sits in the metadata — C2PA content credentials, watermarking, provenance signals. Ask your vendor for it; if they have none, that is a procurement finding.
  • Human-visible labelling applies where a reasonable person could be misled: your landing page hero image, an AI-generated testimonial video, a synthetic voiceover in an ad.

Text has a narrower scope than images. Article 50(2) text disclosure targets content published to inform the public on matters of public interest — not every AI-assisted blog paragraph. Images and video carry the heavier expectation, because that is where the deception risk lives.

3. Emotion recognition and biometric categorisation

If a system infers emotional states from face or voice, or categorises people by biometric data (skin tone, age, sex, ethnicity), Article 50(3) requires those uses to be explicitly disclosed to the natural persons exposed — subject to narrow exceptions for medical or safety reasons.

Separately, emotion inference in the workplace and in education is prohibited outright under Article 5(1)(f). Disclosure is not a compliance route for that one; the system has to go. This is the single most common confusion in vendor conversations: a “sentiment analysis” plugin on your meeting platform may be either a disclosure duty or a banned practice, depending entirely on whether it reads emotion from faces or voice at work.

If you are unsure which side of the line your tool sits on, that is exactly what the free scan flags first.

4. Deepfakes and manipulated media

Article 50(4) obliges the deployer of an AI system that generates or manipulates content — a deepfake — to disclose that the content has been artificially generated or manipulated. The exception is where the content is plainly artistic, creative, fictional or satirical and disclosure would spoil it — a narrow carve-out, not a general marketing exemption.

The practical rule for a mid-market marketing team:

  • Synthetic person on screen, synthetic voiceover, AI-composited product scene → label it, in the video description and on the page.
  • Clearly fictional or artistic content where the label ruins it → document why you judged it exempt. Keep that note.
  • AI-generated customer quotes, reviews, case-study quotes → do not publish them at all. That is a deception problem before it is a transparency problem.

Where the wording goes

Disclosure fails on placement far more often than on wording. Three placements cover almost every surface:

  1. At the point of interaction — in the chat window, in the voice prompt, beside the generated output.
  2. On the page hosting the content — directly adjacent to the image or video, not buried in a footer link.
  3. In the metadata — machine-readable marking retained and not stripped on upload.

Then write it down: a single disclosure register mapping each surface to its wording, its article reference and its owner. That register is what an inspector asks for, and it is also what your enterprise buyers ask for when they run a security questionnaire.

What “compliant” looks like in one sentence

A visitor can tell, within the first screen, that a machine is speaking; a reader can tell, next to the image, that it was generated; and you can produce the list showing where each of those statements lives.

Frequently asked questions

Does Article 50 apply to our website chatbot right now?

Yes. Article 50 has been applicable since 2 August 2026. If your chatbot interacts directly with customers, they must be informed they are dealing with an AI system unless that is genuinely obvious — and where the system imitates a human voice or likeness, disclosure is required in every case.

Do we need a disclaimer on AI-generated marketing images?

If a reasonable person could be misled, yes — mark it visibly on the page and keep machine-readable provenance where your tool provides it. Purely artistic or clearly fictional content can qualify for the exception, but only where disclosure would undermine the artistic purpose, and you should record that reasoning.

Is every AI-written blog post subject to Article 50?

No. The text-disclosure duty in Article 50(2) is focused on content published to inform the public on matters of public interest, alongside the general interaction and synthetic-media duties. Images, audio and video carry the stronger marking expectation because the deception risk is higher.

What is the difference between Article 5 prohibition and Article 50 disclosure for emotion AI?

Emotion inference from faces or voice in the workplace and in education is prohibited under Article 5(1)(f) — disclosure does not cure it. Article 50(3) disclosure applies to emotion recognition and biometric categorisation in contexts that are not prohibited, such as approved medical or safety uses.

How do we prove we complied?

A disclosure register: every AI-facing surface, the exact wording, where it appears, which article requires it, who owns it and when it was last reviewed. Add the machine-readable marking evidence your vendor supplies and the register answers an inspection in one document.

Next step

Put this guide to work on your own estate.

Run the free exposure scan to see which of these articles actually apply to your company — 12 questions, four minutes, no card. Confirm your email and the full report unlocks.

  • Free forever plan
  • No credit card
  • EU-hosted · GDPR-native
Run a free scan Sign up free

Build a live AI register in minutes — upgrade only when you need the audit pack.

EU AI Act Article 50 disclosureAI chatbot disclosure requirementsAI chatbot legal notice EUmark AI generated content EU lawdeepfake labelling requirementArticle 50 transparency obligationsAI Act chatbot website compliancesynthetic content marking machine readable

Keep reading