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

EU AI Act Deepfake Rules: What Article 50 Requires for Labelling AI-Generated Content

Article 50 splits transparency duties between providers and deployers. This guide covers machine-readable marking, deepfake disclosure, chatbot notices, exemptions and timing.

Renomira Team · 7 min read
AI-generated portrait illustrating synthetic media that must be labelled under EU AI Act Article 50
Photo: YandexART / Public domain via Wikimedia Commons

The EU AI Act deepfake rules are not one obligation. They are four duties split between two different parties, and confusing that split is the most common failure we see in mid-market readiness reviews. If your company generates marketing images, runs a customer-service chatbot or publishes AI-assisted copy, at least one lands on you.

Article 50 of Regulation (EU) 2024/1689 sets out the article 50 transparency obligations ai act in full. Providers — the firms building AI systems or placing them on the market — carry the interaction and marking duties. Deployers — the firms using those systems — carry the deepfake and emotion-recognition duties. Together they form the ai generated content labelling requirements that decide who marks synthetic output, where disclosure must sit, which uses are exempt and when enforcement begins.

What the EU AI Act deepfake rules actually require

Article 50 runs four duties, not one, and they split by role. Provider duties: systems intended to interact directly with people must say so (Article 50(1)), and outputs of systems generating synthetic audio, image, video or text must be marked in a machine-readable format and detectable as artificially generated or manipulated (Article 50(2)). Deployer duties: people exposed to emotion recognition or biometric categorisation must be informed of the system’s operation (Article 50(3)), and deep fakes must be disclosed as artificially generated or manipulated (Article 50(4)), with a parallel limb for text published to inform the public on matters of public interest.

Placement rule. Under Article 50(5), that information must be clear and distinguishable at the latest at first interaction or first exposure, and meet applicable accessibility requirements. Article 50 sits alongside the Chapter III high-risk duties rather than replacing them (Article 50(6)), and the AI Office must encourage Union-level codes of practice on detecting and labelling artificial content (Article 50(7)).

Machine-readable marking under Article 50(2)

Machine-readable means a downstream detector can identify the output as artificial — not that a human sees a caption. The Regulation names no technology: solutions must be effective, interoperable, robust and reliable as far as technically feasible, weighing the content type, cost of implementation and the state of the art reflected in relevant technical standards.

How capable providers meet it. Layer the signals: provenance metadata such as C2PA content credentials, an invisible watermark embedded in the pixels or audio stream, and detection-side fingerprints. Each layer fails differently — platform re-encoding and casual screenshotting strip metadata, so metadata alone is a weak watermarking AI content position under EU law. A watermark that survives compression is what keeps working after distribution.

Who carries the duty. If your agency calls a third-party image API, that provider owes Article 50(2); your deployer duty begins at Article 50(4). The commercial risk remains yours: without vendor marking, nothing a detector can check stands behind your disclosure. A practical test of how to label AI generated images in the EU: generate an image, download it, upload it to a social platform, download it again. If no marker survives the round trip, the marking is decorative.

Deepfake disclosure under Article 50(4)

The definition is broad. A deep fake is AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful. A voice clone in a call recording, a synthetic executive in an internal video, a face-swapped advertisement — all qualify. No deceptive intent is required.

The duty. Disclose that the content has been artificially generated or manipulated. Article 50(5) sets the standard — clear and distinguishable, at the latest at first exposure, accessible — but states the outcome rather than the design:

  • Video: an on-frame notice from the first frame, not a line five seconds in or only in the description field.
  • Audio: a spoken or equally clear statement at the point the synthetic voice is first heard.
  • Image: an in-frame label or a caption that travels with the content.
  • Public-interest text: a clear, distinguishing label.

Metadata cannot discharge this duty. The disclosure must reach a natural person, and nobody reads an IPTC field.

Example. A 300-person marketing agency produces a synthetic spokesperson clip: the model provider owes the machine-readable marking, the agency owes the visible disclosure as deployer, and the client owes it again if it posts the clip itself.

Chatbots, emotion recognition and the rest of Article 50

The chatbot notice (Article 50(1)). The duty sits with the provider: the system must be designed so users are told they are dealing with AI, and the only relief is obviousness — a stretch for anything customer-facing. What passes is a persistent “AI assistant” indicator, an opening line such as “You’re chatting with an AI assistant, not a human agent”. A statement buried in terms and conditions is not clear and distinguishable at first interaction. Where the provider has done this well, the deployer’s job is simply to leave it in place.

Emotion recognition (Article 50(3)). This one falls on the deployer: inform the natural persons exposed to the system. Mid-market cases are easier to picture than the legal text — a retail analytics camera estimating customer mood, an interview tool scoring candidate affect, a kiosk categorising visitors. A notice on screen or in the room counts; a line in a privacy policy nobody reads does not. GDPR processing rules apply alongside.

Why both converge. These duties apply whatever the risk tier of the system. A low-risk marketing chatbot still needs the interaction notice, and a workforce tool that never touches Annex III still triggers Article 50(3) if it performs emotion recognition. That is where the ai act chatbot and deepfake labelling duties meet, which is why they belong in one remediation sprint and one register entry.

The disclosure obligation matrix

Situation Who must disclose Legal basis What compliance looks like
Chatbot or assistant in direct interaction Provider Article 50(1) AI notice visible at first interaction
Synthetic audio, image, video or text output Provider Article 50(2) Machine-readable marking detectable as artificial
Deepfake image, audio or video you publish Deployer Article 50(4) Visible disclosure in or alongside the content
AI text published to inform the public on matters of public interest Deployer Article 50(4) Distinguishing label, unless editorial control applies
Emotion recognition or biometric categorisation in use Deployer Article 50(3) Exposed people informed the system is operating
Assistive editing that does not alter input data Neither Article 50(2) carve-out No marking required; rationale kept on file
Evidently artistic, satirical or fictional work Deployer, reduced duty Article 50(4) Existence disclosed without hampering enjoyment

Exemptions and safe harbours

Assistive editing. Article 50(2) steps aside where a system performs an assistive function for standard editing, or does not substantially alter the deployer’s input data or its semantics. Background removal, exposure correction, denoising and spellcheck-grade rewriting sit inside the carve-out; face swaps, voice clones and generative fill that introduce new subjects sit outside. Grey-zone tools — style filters, de-ageing, synthetic backdrops — need one recorded line of reasoning each.

Law-enforcement mandate. Paragraphs 1 to 4 disapply where use is authorised by law to detect, prevent, investigate or prosecute criminal offences, subject to safeguards — a mandate private companies cannot borrow.

Evidently artistic works. Where content forms part of an evidently artistic, creative, satirical, fictional or analogous work, the deepfake duty shrinks to disclosing that such content exists, in a way that does not hamper enjoyment of the work. Evidently carries the weight: calling a campaign creative does not qualify it.

Editorial control and obviousness. The public-interest text limb falls away where human review or editorial control applies and a natural or legal person holds editorial responsibility for publication. Separately, Article 50(1) yields only where AI involvement is obvious to a reasonably well-informed, observant and circumspect person — assume you must state it.

Timing, enforcement and what is still moving

When the duties bite. Article 113 sets the Regulation’s general application date at 2 August 2026, and the Article 50 obligations come in with it. Verify any date you plan against the consolidated text before it enters your programme.

The penalty tier. Article 99(4) expressly lists breach of the Article 50 transparency obligations: administrative fines of up to EUR 15,000,000 or, for undertakings, 3% of total worldwide annual turnover for the preceding financial year, whichever is higher — the lower of the two figures for SMEs. Labelling failures are priced like any other operator obligation, not as a technicality.

Who enforces, and what is moving. National market surveillance authorities, through information requests and investigations; in practice most mid-market firms first meet Article 50 in a customer or insurer questionnaire rather than a regulator letter. Codes of practice under Article 50(7) and the supporting technical standards are still settling, so treat your marking configuration as version-controlled and revisit it as they land.

How to check your own position

Run this in under an hour with tools you already have. First, list every place synthetic content leaves your organisation: marketing image generation, product descriptions, chatbot replies, video ads, internal comms avatars. For each, tag whether the output is published content (Art. 50(4) disclosure falls on you as deployer) or a model you place on the market (Art. 50(2) marking is on you as provider). Second, check your generation stack: does the tool you pay for already emit C2PA credentials or visible notices? Log the answer per tool — “our image tool marks output” is a fact you will be asked to show. Third, add two lines to your existing content workflow: a reviewer check that the disclosure survived cropping and re-encoding, and a sign-off field where the publisher confirms the notice appears in the interaction context. Fourth, sweep your back catalogue for AI-assisted posts published before you had a process — retrofit notices where Art. 50(4) applies rather than deleting history. The commonest failure is drift: the notice survives on the website and dies in the newsletter and the social carousel.

Frequently asked questions

Do deepfakes have to be labelled in the EU?

Yes. Article 50(4) requires deployers to disclose that content is artificially generated or manipulated when it is publicly accessible and forms an image, audio or video — deepfake is the Regulation’s own term for it. The disclosure must be clear, distinguishable and delivered at the latest at the first exposure or interaction. The duty sits with the party publishing the deepfake, not only with whoever built the model.

Who has to watermark AI-generated images, the provider or the deployer?

Providers of generative AI systems carry the machine-readable marking duty under Article 50(2) for outputs that are text, image, audio or video. Deployers publishing AI-generated content carry the disclosure duty under Article 50(4). If you both build and publish, you owe both. When you use someone else’s generator, verify their marking behaviour rather than assuming it — your disclosure duty does not disappear because the tool forgot to watermark.

Does Article 50 apply to AI chatbots as well as deepfakes?

Yes, on a separate limb. Article 50(1) requires that people are informed they are interacting with an AI system unless that is obvious to a reasonably observant person. A clearly branded bot with a persistent “AI assistant” label is usually defensible; a human-looking agent name with no indicator is not. Chatbot disclosure is conversational and in-context; deepfake disclosure is per piece of content plus machine-readable marking.

What is a machine-readable marking under the AI Act?

A technical signal embedded in the output itself — metadata such as C2PA content credentials, invisible watermarks or provenance signals — that lets software identify the content as AI-generated without human inspection. Article 50(2) requires it “where feasible” and points to the developing technical standards for exactly how. It complements, never replaces, the visible disclosure to humans.

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 deepfake rulesai generated content labelling requirementsarticle 50 transparency obligations ai acthow to label ai generated images eumachine readable marking synthetic contentdeepfake disclosure requirements europewatermarking ai content eu lawai act chatbot and deepfake labelling

Keep reading