AI Act23 July 2026DILAIG

EU AI Act Article 50: Transparency Obligations Take Effect on 2 August 2026

Chatbots, deepfakes, AI-generated content: Article 50 of the EU AI Act applies from 2 August 2026, Digital Omnibus notwithstanding. Who is covered, what the text actually requires, and what to do before the deadline.

The Digital Omnibus has created real confusion. Since the provisional agreement reached between the European Parliament and the Council on 7 May 2026, a belief has spread through legal departments and executive committees: "the whole AI Act has been pushed back anyway." That's wrong in principle, and the mistake can be costly.

Article 50 of Regulation (EU) 2024/1689- the transparency obligations for chatbots, synthetic content generators, and emotion recognition or biometric categorisation systems- remains anchored to its original date: 2 August 2026. Only one narrow piece of the article received a targeted delay under the Omnibus, and it covers just part of the obligated parties. For everything else, the deadline is untouched.

If your company deploys a conversational agent, an image or text generation tool, or an emotion-analysis system, the countdown is real. This article breaks down who is covered, what the text actually requires, and what to put in place before the deadline.

What the Digital Omnibus Actually Postponed (and What It Didn't Touch)

The provisional agreement reached on 7 May 2026, and formally endorsed by Parliament on 16 June 2026, changes several AI Act deadlines- notably the timeline for high-risk systems under Annex III (pushed to December 2027) and Annex I (pushed to August 2028). Those are significant delays, but they don't apply to the whole regulation.

Article 50 remains largely unchanged. Its baseline application date stays 2 August 2026, as set by Article 113 of the regulation. Only one adjustment was introduced: systems generating or manipulating synthetic content that were already placed on the market before 2 August 2026 get extra time for the machine-readable marking obligation (Article 50(2)), pushed to 2 December 2026. New systems placed on the market after 2 August 2026 remain subject to the original date, with no grace period.

Every other obligation under Article 50- informing users interacting with a chatbot, informing individuals exposed to emotion recognition or biometric categorisation, disclosing deepfakes and AI-generated public-interest text- gets no postponement at all. They apply from 2 August 2026, full stop.

Why this confusion is dangerous for your business: it pushes teams to delay compliance work that, for the most part, has no slack left. Article 50 is a cross-cutting transparency layer- it applies regardless of the AI system's risk level. A chatbot classified as minimal risk (and therefore outside the compliance obligations of Articles 8-15) is still subject to Article 50 if it interacts with natural persons.

Who Is Covered by Article 50: Four Categories of Obligations

Article 50 sets out four distinct situations, with obligations falling on either providers or deployers.

1. Providers of conversational AI systems

Any company placing on the market an AI system intended to interact directly with natural persons- a customer service chatbot, a voice assistant, a conversational agent built into a product- must inform the user that they are interacting with an AI, unless this is already obvious to a reasonably well-informed person given the circumstances (Article 50(1)).

Concrete example: an SME running a chatbot on its e-commerce site to answer customer questions must clearly signal, from the very first exchange, that the other party is an AI- a mention buried in the terms of service is not enough.

2. Providers of synthetic content generators

Providers of AI systems generating synthetic audio, image, video, or text content must ensure outputs are marked in a machine-readable format and detectable as artificially generated or manipulated (Article 50(2)). This marking has to be technical- a digital watermark, metadata, a cryptographic fingerprint- not just an optional visual disclaimer.

This is the only piece affected by the Omnibus delay: for systems already on the market before 2 August 2026, the marking obligation applies from 2 December 2026. For any system placed on the market after 2 August 2026, there is no delay.

3. Deployers of emotion recognition or biometric categorisation systems

Companies that use an emotion recognition or biometric categorisation system must inform the individuals exposed to it, and process personal data in line with applicable rules- GDPR included (Article 50(3)).

Concrete example: a call centre deploying an emotion-analysis tool on customer recordings to assess satisfaction must inform the individuals concerned, independently of any separate GDPR obligation.

4. Deployers of deepfakes and AI-generated public-interest text

Any company or organisation that generates or manipulates images, audio, or video content constituting a "deepfake" must disclose that the content has been artificially generated or manipulated. A similar obligation applies to AI-generated text published to inform the public on matters of public interest (Article 50(4)).

Concrete example: a media outlet or communications agency publishing an AI-written article on a current-affairs topic must clearly flag it- this obligation is specifically aimed at limiting AI-driven disinformation on matters of public concern.

What the Text Actually Requires: Timing and Form of the Disclosure

The regulation is precise on two points that are often overlooked:

Timing: the information must be provided at the latest at the time of the first interaction or exposure of the individual to the system. This isn't information that can be deferred to terms and conditions read after the fact.
Form: the information must be clear and distinguishable- understandable without particular effort, not buried in dense legal text or a hard-to-read interface.

These requirements echo a logic already familiar from consumer law and GDPR: information has to be useful at the moment it matters, not merely available somewhere.

What to Do Before 2 August 2026

Action	Detail
Map the systems in scope	Chatbots, content generators, biometric/emotion tools- including third-party tools you've integrated (SaaS, plugins, APIs)
Identify your role	Provider or deployer under the AI Act; a company integrating a third-party API can hold both roles at once
Check the technical marking	For any synthetic content generated after 2 August 2026; request written confirmation of compliance if the generator is supplied by a third party
Review user interfaces	Disclosure visible at first contact, not buried in the terms of service
Document the approach	Keep a record of the analysis performed, the marking choices made, and the disclosures put in place
Don't confuse this with high-risk status	A minimal-risk system is still covered by Article 50 if it falls into one of the four categories

Penalties: A Generic Regime, But a Real One

Breaches of Article 50 fall under the AI Act's "generic" penalty regime, set out in Article 99- distinct from the harsher regime that applies to the prohibited practices under Article 5. This generic regime provides for administrative fines that can reach tens of millions of euros or a percentage of global annual turnover, whichever is higher, with reduced caps for SMEs and startups. The exact amounts and their application vary depending on the nature of the breach and the size of the company: checking the consolidated text of Article 99, or seeking legal advice, is recommended to assess the financial exposure specific to your organisation.

Beyond the fine, reputational risk shouldn't be underestimated: a chatbot that doesn't identify itself as AI, or AI-generated content distributed without disclosure, can quickly become a public controversy- well before any supervisory authority gets involved.

FAQ: Article 50 and AI Transparency

Was Article 50 postponed by the Digital Omnibus? Not as a whole. Only the machine-readable marking obligation for synthetic content (paragraph 2) gets a delay, to 2 December 2026, and only for systems already on the market before 2 August 2026. Every other obligation under Article 50 applies from 2 August 2026.

Is a minimal-risk chatbot covered by Article 50? Yes. Article 50 applies regardless of the system's risk level. A chatbot classified as minimal risk is still subject to the obligation to inform users they're interacting with an AI, unless that's already obvious.

Who has to inform the user- the provider or the deployer of the chatbot? The obligation to disclose AI interaction (paragraph 1) falls on the provider of the system. The obligations related to emotion recognition, biometric categorisation, and deepfakes (paragraphs 3 and 4) fall on the deployer.

Does AI-generated content marking need to be visible to the user? The regulation requires machine-readable technical marking (metadata, digital watermarking). It doesn't necessarily require a systematic visible display- but the disclosure obligations under paragraphs 1 and 4 do require clear, distinguishable information for the exposed individual.

Key Takeaways

Article 50 of the AI Act applies from 2 August 2026. The Digital Omnibus postponed only one piece: machine-readable marking of synthetic content for systems already on the market, delayed to 2 December 2026.
It covers four categories: chatbots, synthetic content generators, emotion recognition/biometric categorisation systems, and deepfakes or AI-generated public-interest text.
The transparency obligation applies regardless of the system's risk level.
Disclosure must happen at first contact, clearly and distinguishably.
Penalties fall under the generic Article 99 regime- a real financial and reputational risk, not one to underestimate on the assumption that "everything got postponed."