EU AI Act Article 50: A Guide to AI Content Labeling, Watermarking and Chatbot Disclosure

TL;DR

Article 50 of the EU AI Act covers the most visible uses of AI: chatbots, image/audio/video generators, deepfakes and AI-written text. Unlike high-risk systems, there is no conformity assessment or quality management system here; but there is a duty to tell people when they are dealing with AI, and it has applied since 2 August 2026.

Four duties, two roles:

PROVIDER (develops the system / offers it under its own name)
  50(1) Tell people they are interacting with AI      → chatbots, voice assistants
  50(2) Mark synthetic content in a machine-readable way → image, audio, video, text generation

DEPLOYER (uses the system in its business)
  50(3) Disclose emotion recognition / biometric categorisation
  50(4) Disclose deepfakes and AI text that informs the public

Timeline as of October 2026:

Important: This guide is not legal advice. The information was checked against primary sources (the consolidated AI Act text on the AI Act Service Desk, European Commission pages) as of October 2026; sources are at the end. Whether a specific system is in scope is a question for the text itself, the Commission's Article 50 guidelines and a lawyer. For the broader governance setup, see the AI governance playbook.

Dates and the transition period

The AI Act (Regulation (EU) 2024/1689) entered into force on 1 August 2024. Under Article 113 its general date of application is 2 August 2026, and Article 50 follows that general date; it was never separately postponed.

What the Digital Omnibus changed. Regulation (EU) 2026/1744 ("Digital Omnibus on AI"), which amends the AI Act, is dated 8 July 2026; it was published in the Official Journal on 24 July 2026 and entered into force on 27 July 2026 (the third day after publication). We checked the provisions below against the Official Journal text (the Publications Office Cellar copy) and the consolidated text on the Service Desk. For Article 50 the change is narrow:

• New Article 111(4): providers of AI systems, including general-purpose AI systems, generating synthetic audio, image, video or text that were placed on the market before 2 August 2026 must comply with Article 50(2) by 2 December 2026.
• The transition covers only 50(2). Chatbot disclosure (50(1)), emotion recognition notices (50(3)) and deepfake disclosure (50(4)) have no transition; they have all applied since 2 August 2026.
• A new generative system placed on the market after 2 August 2026 gets no transition either.
• The Commission had proposed six months (to 2 February 2027); the adopted text gives four (Cuatrecasas' note on the legislative history and the AI Act Service Desk FAQ).
• The Omnibus also amended Article 50(7) on codes of practice (Regulation 2026/1744, Article 1(20)): the Commission's power to approve the codes by implementing act was removed and replaced with a Commission assessment of whether adherence is adequate for compliance with 50(2) and 50(4), taking utmost account of the AI Board's opinion. The power to adopt an implementing act with common rules if a code is inadequate remains. Recital 41 of the Omnibus states that these codes have limited legal effect and do not grant a presumption of conformity.

Older content. According to the Commission's "Quick Facts" page, retroactive labelling of deepfakes generated before 2 August 2026 is not mandatory, but encouraged.

Supporting documents — status:

Note: according to the Commission's opinion, signing the code is not conclusive evidence of compliance, but it is an EU-wide recognised route to it. Providers and deployers who don't sign must demonstrate compliance with the marking and labelling obligations through other, equivalently adequate means.

Provider or deployer?

Article 50 splits duties by role. The definitions are in Article 3:

• Provider (Article 3(3)): a person or body that develops an AI system or general-purpose AI model, or has one developed, and places it on the market or puts it into service under its own name or trademark, whether for payment or free of charge.
• Deployer (Article 3(4)): a person or body using an AI system under its authority, except in the course of a personal non-professional activity.

In practice:

The Commission's Article 50 guidelines explain with detailed examples which party carries which duty along the value chain; check them first if you are in a borderline case.

Recital 133 says marking techniques can be implemented at model level, which makes things easier for the downstream provider using that model. But the duty sits with the system provider: if your model vendor embeds a watermark, document it; if not, the duty falls on you.

Article 50(1): telling people they're talking to AI

Whose duty: the provider's.

What it requires: AI systems intended to interact directly with natural persons must be designed and developed so that those persons are informed they are interacting with an AI system.

Exceptions:

• Where this is obvious to a reasonably well-informed, observant and circumspect person, taking into account the circumstances and context of use.
• Systems authorised by law to detect, prevent, investigate or prosecute criminal offences (with appropriate safeguards), unless they are available for the public to report a criminal offence.

How and when (Article 50(5)): the information must be given in a clear and distinguishable manner, at the latest at the time of the first interaction, and must meet applicable accessibility requirements. A small grey footnote a screen reader can't reach doesn't meet that bar.

Practical reading: interpret the "obvious" exception narrowly. A bot in a support window that talks under a human agent's name and photo is not "obviously" AI. The safest route is a persistent line at the start of the conversation:

Hi, I'm the [Product] Assistant, an AI system.
Type "agent" any time to talk to a human.

For voice assistants, give the same notice in the first spoken sentence.

Article 50(2): machine-readable marking of synthetic content

Whose duty: providers of AI systems, including general-purpose AI systems, that generate synthetic audio, image, video or text.

What it requires: outputs must be marked in a machine-readable format and detectable as artificially generated or manipulated. Technical solutions must be effective, interoperable, robust and reliable as far as technically feasible, taking into account the specificities and limitations of different content types, implementation costs and the generally acknowledged state of the art as reflected in relevant technical standards.

Exceptions:

• The system performs an assistive function for standard editing (e.g. spell-check),
• It does not substantially alter the input data provided by the deployer or its semantics,
• It is authorised by law for criminal-offence purposes.

Text is included. 50(2) covers text, not just images. If you offer an LLM-based writing assistant, machine-readable marking of its output is your problem. Text watermarking is technically the weakest area; see "Technical implementation" below.

A visible "AI-generated" caption does not satisfy 50(2) on its own; 50(2) asks for a machine-readable mark. The visible label is a separate duty (50(4)) and usually belongs to another party (the deployer).

Article 50(3): emotion recognition and biometric categorisation

Whose duty: the deployer's.

What it requires: deployers of an emotion recognition or biometric categorisation system must inform the people exposed to it that the system is operating, and process personal data in line with the GDPR (2016/679), Regulation (EU) 2018/1725 and Directive (EU) 2016/680, as applicable.

Exception: uses permitted by law to detect, prevent or investigate criminal offences, with appropriate safeguards.

Check the prohibitions first: 50(3) only covers permitted uses. Emotion recognition at work and in education (except for medical or safety reasons) is prohibited under Article 5; a notice doesn't make it lawful. Some biometric systems are also high-risk under Annex III, with those obligations arriving separately from 2 December 2027.

Article 50(4): deepfakes and text informing the public

Whose duty: the deployer's.

Deepfake definition (Article 3(60)): 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.

What it requires (first part): deployers of an AI system that generates or manipulates deepfake content must disclose that the content has been artificially generated or manipulated.

Exceptions and relief:

• Use authorised by law for criminal-offence purposes.
• Where the content is part of an evidently artistic, creative, satirical, fictional or analogous work or programme, the duty is limited to disclosing the existence of generated or manipulated content in an appropriate manner that does not hamper the display or enjoyment of the work. A note in the credits or description can be enough; you don't need a label on every frame.

What it requires (second part): deployers of an AI system that generates or manipulates text published with the purpose of informing the public on matters of public interest must disclose that the text was artificially generated or manipulated.

Exception to this part: no disclosure is needed where the AI-generated content has undergone human review or editorial control and a natural or legal person holds editorial responsibility for the publication. A news text an editor has read and taken responsibility for is out; auto-published AI news summaries are in.

What the visible label should look like: Article 50(5) applies here too: clear, distinguishable, at the latest at first exposure, and accessible. The Commission's summary rules for the EU icons: the icon should be clearly perceivable, have no overlapping elements, and be embedded directly in the deepfake or text (except for creative works). Detailed placement rules are in Section 2 of the code. Non-signatories may use the icons too, but that use should not be read as adherence to the code.

Fines and enforcement

• Article 99(4)(g): breaching Article 50 is subject to fines of up to €15 million or 3% of total worldwide annual turnover, whichever is higher.
• Article 99(5): supplying incorrect, incomplete or misleading information to authorities: €7.5 million or 1%.
• Articles 99(6) and 99(6a): for SMEs (including start-ups), the cap is whichever of the amount or percentage is lower; the Omnibus extended this to small mid-cap enterprises (SMCs) for 99(4) and (5).
• Who enforces: national market surveillance authorities; the AI Office for systems under its supervision; the European Data Protection Supervisor where EU institutions are providers or deployers (Commission "Quick Facts" page).

In practice, an information request and a demand to fix things seem more likely than a fine as a first step; but that is our guess, the text sets no such escalation. Keep an evidence file from day one: which marking, which notice, since when, who approved it.

How Article 50 relates to the GPAI rules

Article 50 and the general-purpose AI rules regulate different things:

Where they meet:

• Article 50(2) explicitly includes general-purpose AI systems. A company that offers its own model through a chat app is both a GPAI model provider and an Article 50 system provider.
• Under Recital 133, marking can happen at model level. If a model vendor embeds a watermark in the model, a system provider using it can benefit, but should confirm this in the contract and documentation.
• The GPAI Code of Practice (2025) and the Code of Practice on Transparency of AI-generated Content (2026) are separate documents; signing one doesn't cover the other.

Technical implementation: options and honest limits

Article 50(2) doesn't name a technology. Recital 133 lists watermarks, metadata identifications, cryptographic methods for proving provenance and authenticity, logging methods and fingerprints as examples, and allows combinations. For the concepts in detail, see AI watermarking and content labeling.

The final code (10 June 2026), Section 1, contains these technical commitments for providers:

• Multi-layered marking (Measure 1.1): for audio, images, video and containerised text, at least two machine-readable layers: digitally signed and (where possible) time-stamped metadata plus an imperceptible watermark. For free-form text, which can't carry metadata, a single layer (the watermark) is enough. A single layer is also accepted for physical products generating content in a closed environment that technically prevents export.
• Very short text: text under 200 tokens is exempt from watermarking (the glossary's "very short text" definition); longer free-form text must be watermarked even if reliability is lower.
• Fingerprinting / logging (Sub-measure 1.1.3): optional; not sufficient on its own.
• Non-removal (Measure 1.2): preserve existing metadata markings on inputs and make best efforts against intentional removal.
• Detection (Commitment 2): the detection tool must be available through an interface suited to the audience exposed to the content (the public, for public content); access to free-form text watermark detection may be limited to verified experts (authorities, researchers, fact-checkers and others).
• Interoperability (Measure 3.4): adopt established standards for metadata; by 2 February 2027, implement a common access method, a publicly readable signpost or a shared detection service for watermark detection.
• The code does not name C2PA; it says "established standards". C2PA is the most widespread standard for signed metadata today, so it is the obvious candidate, but that is our reading.

These commitments bind signatories only; Article 50(2) itself doesn't require a set number of layers. For a provider who doesn't sign, the list is still a good reference for where the "state of the art" bar sits in practice.

Honest limits:

• "No watermark found" does not mean "made by a human"; a detector only recognises its own scheme.
• Non-watermark "AI detectors" don't read a signal, they guess; accusing an employee or student based on them is wrong.
• Academic work argues that strong watermarks can be removed under certain assumptions (e.g. arXiv 2311.04378). That's why Article 50(2) says "as far as technically feasible".
• Your own publishing pipeline may strip metadata: image optimisation plugins, CDN transforms, social media automation. Test it.

A simple implementation pipeline:

generation request
  → model / API (is the vendor's watermark on? document it)
  → output
  → add and sign a C2PA manifest (generator, "AI-generated", date)
  → store a fingerprint (optional)
  → deliver to the user
  → in the publishing pipeline: does the manifest survive? (CDN, compression test)
  → verification: content → detector / manifest reader

Checklist: a company with a chatbot

• Is the bot offered under our name? (If yes, 50(1) is our duty.)
• Is there an "I'm an AI system" notice in the first chat message, and in the first sentence on voice channels?
• Can a screen reader read the notice, and is the contrast sufficient? (Article 50(5))
• Is the bot presented with a human name, photo or "your agent Anna"? If so, the notice needs to be even more prominent.
• Do other channels (email, WhatsApp, phone) carry the same notice?
• If the bot generates images or long text for users, does 50(2) marking apply too?
• Are the notice text and a dated screenshot in the evidence file?
• Have we checked whether the bot performs emotion analysis? (50(3) and Article 5)

Checklist: a company offering image generation

• Was the feature placed on the market before 2 August 2026? If yes, the 50(2) deadline is 2 December 2026; if not, it already applies.
• Does the model vendor embed a watermark? Is that in the contract or documentation? Can it be switched off?
• Does every output get a signed C2PA manifest?
• Does the manifest survive download, sharing and API response flows? (Re-compression and resizing tests)
• Is there a verification path (detector, manifest reader, support process)?
• Have we separated functions that fall under the "standard editing" exception (e.g. colour correction) from those that don't (e.g. replacing the background, generating faces), and documented why?
• Will we sign the code of practice? If not, how do we show equivalent compliance?
• Do our user terms tell users who publish content resembling real people about their 50(4) disclosure duty?
• Do we have measures preventing generation of non-consensual intimate imagery and child sexual abuse material? (Prohibition added by the Omnibus, 2 December 2026)

Checklist: a marketing team using generative AI

• Does the campaign include AI images, audio or video that resemble a real person, place or event and look authentic? If so, visible disclosure is mandatory (50(4)).
• If the content is evidently creative or satirical, is the disclosure placed somewhere appropriate that doesn't spoil the work? (E.g. video description, credits)
• If a celebrity's or employee's face or voice is used, have consent and personality rights been checked separately from disclosure? (That's other law, not Article 50.)
• Do texts that inform the public, such as blog posts and press releases, go through an editor who takes editorial responsibility? If not, disclosure is needed.
• Do we strip metadata and watermarks from tool output when publishing? It isn't prohibited, but deliberately cleaning a provider's mark sends a bad signal.
• EU icon set or our own label? Is the choice and the placement rule written down?
• Are all AI-generated campaign assets in the inventory and approval log? (See the AI governance playbook, Step 1.)

Notes for Turkish companies selling into the EU

Extraterritorial scope. Article 2(1)(a) applies the AI Act to providers placing systems on the EU market wherever they are established. Article 2(1)(c) also covers providers and deployers in third countries where the output produced by the system is used in the EU. So:

• An Istanbul SaaS company offering a chatbot or image tool to EU customers is a provider for 50(1) and 50(2).
• An agency run from Turkey using deepfakes in a campaign aimed at an EU audience can be in scope as a deployer for 50(4).
• There are also authorised-representative rules for non-EU providers placing AI on the EU market; these mainly concern high-risk systems and GPAI models. Clarify your own case with a lawyer.

The KVKK angle. We found no Turkish rule in force that directly mirrors Article 50 for AI content labelling. The Turkish Personal Data Protection Authority (KVKK) has two verifiable documents on the topic, both guidance rather than binding regulation:

• Deepfake Information Note (19 January 2024): what deepfakes are, the threats they pose to personal data, how to spot them, and what individuals and organisations can do.
• Generative AI and Personal Data Protection Guide (in 15 Questions) (24 November 2025).

A person's face or voice is personal data; producing a deepfake from it is processing under the KVKK (Law No. 6698), and the AI Act's disclosure duty does not replace a legal basis for that processing. Bills. The TBMM page for AI bill no. 2/2234, submitted on 24 June 2024, shows its status as "KOMİSYONDA" (in committee) as of October 2026; the lead committee is the Industry, Trade, Energy, Natural Resources, Information and Technology Committee. Turkish media have reported other bills in 2025 and 2026 that would require deepfakes to carry a "generated by artificial intelligence" notice and fine platforms; we have not verified their text or status against TBMM records. None appears to have become law, but don't plan compliance around bills. For the wider picture, see the KVKK section of the AI governance playbook.

Sources

Primary sources (checked as of October 2026)

• AI Act Service Desk: Article 50 (consolidated text as at 27 July 2026)
• AI Act Service Desk: Article 111 (transitional provisions, 111(4))
• AI Act Service Desk: Article 113 (dates of application)
• AI Act Service Desk: Article 99 (penalties)
• AI Act Service Desk: Article 2 (scope) and Article 3 (definitions)
• AI Act Service Desk: Recital 133 and Recital 134
• Regulation (EU) 2024/1689 (AI Act), EUR-Lex
• Regulation (EU) 2026/1744 (Digital Omnibus on AI), EUR-Lex
• European Commission: Code of Practice on Transparency of AI-generated Content
• European Commission: Publication of the final code (10 June 2026)
• European Commission: Opinion on the assessment of the code
• European Commission: Article 50 guidelines (20 July 2026)
• European Commission: Quick Facts — Transparency rules for AI systems
• European Commission: EU icons for labelling AI-generated content
• European Commission: Final text of the code (PDF, 10 June 2026)
• TBMM: AI bill 2/2234 (Turkish)
• KVKK: Deepfake Information Note (Turkish)
• KVKK: Generative AI and Personal Data Protection Guide (Turkish)

Secondary and technical sources

• Cuatrecasas: Digital Omnibus on AI has been published
• Cooley: Digital AI Omnibus delays key deadlines
• C2PA Specifications
• Google: SynthID Text documentation
• Dathathri et al., "Scalable watermarking for identifying large language model outputs", Nature (2024)
• Gowal et al., "SynthID-Image: Image watermarking at internet scale" (arXiv 2510.09263)
• Kirchenbauer et al., "A Watermark for Large Language Models" (arXiv 2301.10226)
• Zhang et al., "Watermarks in the Sand" (arXiv 2311.04378)

On this site

• EU AI Act, AI Risk Classification, AI Governance — the legal framework.
• AI Watermarking & Content Labeling, Image Generation, Diffusion Model — the technical side.
• AI Governance: A Practical Starter Playbook — the full inventory, policy and controls setup.

Frequently asked questions

When does Article 50 of the AI Act apply?

Article 50 has applied since 2 August 2026, the AI Act's general date of application (Article 113). The only transitional exception concerns the Article 50(2) machine-readable marking duty for generative AI systems placed on the market before 2 August 2026; those have until 2 December 2026 (Article 111(4)).

Does AI-generated content have to be labelled?

Not every piece needs a visible label. Providers of generative AI systems must mark all synthetic audio, image, video and text output in a machine-readable way; a visible disclosure by the deployer is only mandatory for deepfakes and for AI text published to inform the public on matters of public interest without human editorial control.

Is a 'you are talking to an AI' notice mandatory for chatbots?

Yes. Under Article 50(1), providers of AI systems that interact directly with people must ensure those people are informed they are interacting with AI. There is an exception where this is obvious to a reasonably well-informed, observant and circumspect person, but relying on it is risky; the notice must be clear and distinguishable at the latest at the first interaction.

What is the fine for breaching Article 50?

Under Article 99(4)(g), up to €15 million or 3% of total worldwide annual turnover, whichever is higher. For SMEs, and for the small mid-cap enterprises (SMCs) added by the Omnibus, the cap is whichever of the two is lower.

Does Article 50 apply to a company outside the EU?

It can. The AI Act applies to providers placing systems on the EU market wherever they are established, and to providers and deployers in third countries where the system's output is used in the EU (Article 2(1)(a) and (c)).

Do I have to sign the Code of Practice on marking and labelling?

No, adherence is voluntary, but the Article 50 obligations are binding. The Commission and the AI Board found the code adequate in July 2026; non-signatories have to demonstrate compliance through other, equivalently adequate means.

Is C2PA or a watermark enough on its own?

Article 50(2) doesn't prescribe a technology; it requires solutions that are effective, interoperable, robust and reliable as far as technically feasible. Metadata can be lost with a screenshot and watermarks can weaken under rewriting or translation, so in practice a layered approach is used.