Table of Contents

Reach SOC 2 Compliance in 6 Weeks or Less.

  /

  / EU AI Act Article 50: AI Transparency Rules from August 2026

EU AI Act Article 50: AI Transparency Rules from August 2026

The EU AI Act’s transparency requirements take effect on 2 August 2026, and most of the companies they cover still think the rules are not their problem. Article 50 applies to any business that publishes AI-generated content or runs an AI system that talks to people in the EU. That includes the marketing team generating campaign images and the support team running a chatbot. It also covers the AI agents you’ve wired into customer email. Penalties reach €15 million or 3% of total worldwide annual turnover, whichever is higher, and you don’t need an office in Europe to be in scope. If your content or your chatbot reaches EU users, the obligations reach you.

In a nutshell: if you publish AI-generated images or video, deploy chatbots or AI agents that interact with EU users, or publish AI-written text on matters of public interest, then yes, the EU AI Act applies, starting 2 August 2026.

A quick word on the “AI Act delay” headlines. The Digital Omnibus package did push the high-risk system deadlines back, in some cases by more than a year, but it did not move the deployer obligations in Article 50. Companies that read those headlines and stood down their AI Act work made an expensive mistake, because the rules most likely to touch an ordinary business are the ones that stayed on the calendar.

What Article 50 Actually Requires

Article 50 of the AI Act sets out transparency obligations in four situations. In plain English:

  • Tell people when they’re talking to AI. Systems designed to interact directly with people — chatbots, voice assistants, and AI agents — must make clear that the user is dealing with AI, unless that’s already obvious.
  • Mark AI-generated content so machines can detect it. Providers of generative AI systems must mark outputs in a machine-readable format, typically through metadata and watermarking, so the content is detectable as artificially generated.
  • Label deepfakes. Anyone deploying AI to generate or manipulate image, audio, or video content that resembles real people, places, objects, or events, and could falsely appear authentic, must disclose that the content is artificial.
  • Label AI-generated text on matters of public interest. Text published to inform the public must carry a label if AI-generated or manipulated, unless a human reviewed it and a person or organization holds editorial responsibility for it.

Article 50 also covers emotion recognition and biometric categorization systems, which carry their own disclosure duties. Far fewer businesses run into those, so this article sticks to the four above.

The distinction running through all of this is provider vs deployer. The provider builds or supplies the AI system. The deployer uses it professionally. Most companies reading this are deployers.

Let Axipro help you build a business continuity plan that's practical, compliant, and audit-ready.

Schedule Your Free Assessment Today

If You Use AI-Generated Images

Realistic AI images sit closer to the deepfake rules than most marketing teams assume. The Act’s definition covers content depicting people, objects, places, and events that could falsely appear authentic to a viewer, which describes a large share of what image generators produce for campaigns, social posts, and landing pages.

So what does “clearly and distinguishably labeled” mean? The threshold is best described by its failures: a tiny disclosure hidden in the website footer doesn’t qualify. Neither does a faint label on an image, a label that flashes for an instant in a video, or a disclosure buried in your terms and conditions.

The label has to be visible right where someone sees the content, and it has to meet accessibility standards so people with disabilities can perceive it too.

The Code of Practice proposes a standardized “AI” visual label, localized per language (“KI” in German, “IA” in French). It also draws a useful line between fully AI-generated content and AI-assisted content, with lighter requirements for the latter. A designer who used AI to extend a background is in a different position from a team publishing a fully synthetic image of a person who doesn’t exist.

Important: The deepfake duty doesn’t care about intent. A flattering, harmless AI image of your CEO at an event that never happened is still a deepfake under the Act. Marketing teams generate this kind of content casually. From August, every one of those images needs a label.

EU AI Act Article 50

If You Deploy AI Agents or Chatbots

The rule itself is simple: people must know they’re dealing with AI. The provider carries the design obligation, but as the deployer you’re the one putting the system in front of your customers, and you’re the one an EU regulator will contact if your branded assistant pretends to be human.

The Act contains an exception for cases where it’s “obvious” the user is talking to AI, judged from the perspective of a reasonably well-informed and observant person. Don’t lean on it. What’s obvious to your product team isn’t obvious to every customer, and the human-sounding voice agents and email-writing AI agents rolling out right now are designed specifically to not feel like software. If an AI agent negotiates a renewal over email or handles a support ticket end to end, disclose it.

Pro Tip: Put the Disclosure at the Start of the Interaction

Put the disclosure at the start of the interaction, in the interface itself: "You're chatting with an AI assistant." A line in your privacy policy doesn't meet the standard, and a disclosure that appears after the conversation ends is worthless. For voice agents, say it up front in the greeting.

What Your AI Vendors Owe You

The machine-readable marking obligation in Article 50(2) sits with providers — the companies supplying your generative AI tools. The final Code of Practice expects providers to apply at least two layers of marking where necessary, such as embedded metadata combined with watermarking, and to offer detection mechanisms so deployers, authorities, and researchers can verify whether a piece of content came from AI.

One timing caveat: the Digital Omnibus gives generative AI systems already on the market before 2 August 2026 until 2 December 2026 to comply with the marking requirement. Every other Article 50 obligation stays on the August date.

The practical consequence lands in your contracts. The hard part of Article 50 is deciding who is the provider and who is the deployer at each link in the supply chain, and your vendor agreements need to say which party holds each duty. Ask your generative AI vendors three things: what marking technology they apply, how you can access their detection tooling, and whether they’ve signed the Code of Practice.

Insider Note: The four-month marking grace period creates an awkward window. Between August and December 2026, some of the tools you use daily won’t watermark their output yet, while your deployer duties apply in full from day one. So don’t build your labeling process on the assumption that vendor watermarks have you covered. Label at your own layer, visibly, and treat vendor-side marking as reinforcement rather than the plan.

The Code of Practice: Voluntary, but the De Facto Standard

The European Commission published the final Code of Practice on marking and labelling of AI-generated content, drafted by independent experts with input from more than 180 stakeholders. It has two sections: one for providers, covering marking and detection, and one for deployers, covering the labeling of deepfakes and public-interest text.

Signing is voluntary. Ignoring it is unwise. The Code is the only official, concrete answer to “what does compliance actually look like,” and regulators will most likely use it as their reference point when they judge whether a company made a good-faith effort. The Commission has also prepared guidelines interpreting Article 50’s scope, definitions, and exceptions, with the final version expected around the application date.

Behind on This? A 7-Day Catch-Up Plan

If today is the first time anyone in your company has looked hard at Article 50, a week of focused work covers most of the exposure.

Days 1 and 2 — Inventory your AI content flows.
List every place AI-generated text, images, audio, or video reaches an EU audience, plus every chatbot, voice assistant, or AI agent that interacts with EU users. Count what marketing, sales, and support produce with AI tools, alongside anything built into your product.

Day 3 — Assign roles.
For each tool and use case, decide whether you’re the provider or the deployer, and note which Article 50 duty attaches. This one exercise clears up most of the confusion.

Day 4 — Fix your visual content.
Add clear labels to realistic AI-generated images and video, and audible disclosures to synthetic audio, following the Code’s guidance on visible, persistent labels.

Day 5 — Fix your interactions.
Add disclosures to every chatbot, assistant, and agent at the start of the interaction, and check they meet accessibility requirements.

Day 6 — Fix your contracts.
Write to your generative AI vendors, request their marking and detection documentation, and update agreements to state who holds each duty.

Day 7 — Document everything.
Document everything — record what you inventoried, the roles you assigned, the labels you applied, and the reasoning behind each judgment call. If a regulator ever asks, this record is your first line of defense.

Let Axipro help you build a business continuity plan that's practical, compliant, and audit-ready.

Schedule Your Free Assessment Today

Where Article 50 Fits in Your AI Governance Program

Nothing in that checklist should live as a one-off project. An AI system inventory, assigned roles and responsibilities, transparency controls, vendor management, and documented decisions are all core requirements of ISO 42001, the international standard for AI management systems. Companies that already run an ISO 42001 program will find that Article 50 mostly falls out of controls they operate anyway. Companies scrambling this week are usually the ones without a governance structure to hang the work on.

That’s the real lesson of this deadline. The EU has more AI rules coming: the marking obligation lands in December, and the high-risk regime arrives later. Treating each deadline as a fire drill doesn’t scale. Axipro helps companies build ISO 42001-aligned AI governance programs that absorb requirements like Article 50 as routine control updates instead of emergencies. If the checklist above turned up gaps you can’t close on your own, that’s the work we do. Talk to us.

The short version: Article 50’s transparency requirements apply from 2 August 2026 to any business whose AI-generated content or AI systems reach EU users. Deployers must label deepfakes and public-interest AI text and make sure people know when they’re talking to AI. Providers must mark synthetic content, from December for tools already on the market. The final Code of Practice explains what good looks like.

Heavy AI use isn’t what puts a company at risk here. The risk comes from never writing down where that AI shows up.

Frequently Asked Questions

Does Article 50 apply to companies outside the EU?

Yes. The obligations follow the content and the users, not your headquarters. A US or GCC company publishing AI-generated campaigns to EU audiences, or running a chatbot that serves EU customers, is in scope.

Partly. The Digital Omnibus moved the high-risk system deadlines back and gave providers of generative AI systems already on the market until 2 December 2026 for the machine-readable marking duty. The deployer obligations — deepfake labeling, public-interest text labeling, and chatbot disclosure — all apply from 2 August 2026 as originally scheduled.

No. The text-labeling duty covers AI-generated or AI-manipulated text published to inform the public on matters of public interest, and it includes a carve-out where a human reviewed the text and someone holds editorial responsibility for it. Ordinary marketing content that goes through human review generally doesn’t need a label. Realistic synthetic images and video face stricter treatment.

No, it’s voluntary. But it’s the Commission-backed description of what compliance looks like in practice, and following it is the most defensible way to show good faith. Most companies should treat it as the standard.

Axipro Author

Picture of Pedro Dias

Pedro Dias

Pedro has been writing online for over 10 years. With experience in all things programming, cyber security, and compliance, he is our editor-in-chief at Axipro.

Blog Highlights

Explore More Articles

The EU AI Act’s transparency requirements take effect on 2 August 2026, and most of the companies they cover still think the rules are not their problem. Article 50 applies to any business that publishes AI-generated content or runs an AI system that talks to people in the EU. That includes the marketing team generating campaign images and the support team running a chatbot. It also covers the AI agents you’ve wired into customer email. Penalties reach €15 million or 3% of total worldwide annual turnover, whichever is higher, and you don’t need an office in Europe to be in scope. If your content or your chatbot reaches EU users, the obligations reach you. In a nutshell: if you publish AI-generated images or video, deploy chatbots or AI agents that interact with EU users, or publish AI-written text on matters of public interest, then yes, the EU AI Act applies, starting 2 August 2026. A quick word on the “AI Act delay” headlines. The Digital Omnibus package did push the high-risk system deadlines back, in some cases by more than a year, but it did not move the deployer obligations in Article 50. Companies that read those headlines and stood down their AI Act work made an expensive mistake, because the rules most likely to touch an ordinary business are the ones that stayed on the calendar. What Article 50 Actually Requires Article 50 of the AI Act sets out transparency obligations in four situations. In plain English: Tell people when they’re talking to AI. Systems designed to interact directly with people — chatbots, voice assistants, and AI agents — must make clear that the user is dealing with AI, unless that’s already obvious. Mark AI-generated content so machines can detect it. Providers of generative AI systems must mark outputs in a machine-readable format, typically through metadata and watermarking, so the content is detectable as artificially generated. Label deepfakes. Anyone deploying AI to generate or manipulate image, audio, or video content that resembles real people, places, objects, or events, and could falsely appear authentic, must disclose that the content is artificial. Label AI-generated text on matters of public interest. Text published to inform the public must carry a label if AI-generated or manipulated, unless a human reviewed it and a person or organization holds editorial responsibility for it. Article 50 also covers emotion recognition and biometric categorization systems, which carry their own disclosure duties. Far fewer businesses run into those, so this article sticks to the four above. The distinction running through all of this is provider vs deployer. The provider builds or supplies the AI system. The deployer uses it professionally. Most companies reading this are deployers. If You Use AI-Generated Images Realistic AI images sit closer to the deepfake rules than most marketing teams assume. The Act’s definition covers content depicting people, objects, places, and events that could falsely appear authentic to a viewer, which describes a large share of what image generators produce for campaigns, social posts, and landing pages. So what does “clearly and distinguishably labeled” mean? The threshold is best described by its failures: a tiny disclosure hidden in the website footer doesn’t qualify. Neither does a faint label on an image, a label that flashes for an instant in a video, or a disclosure buried in your terms and conditions. The label has to be visible right where someone sees the content, and it has to meet accessibility standards so people with disabilities can perceive it too. The Code of Practice proposes a standardized “AI” visual label, localized per language (“KI” in German, “IA” in French). It also draws a useful line between fully AI-generated content and AI-assisted content, with lighter requirements for the latter. A designer who used AI to extend a background is in a different position from a team publishing a fully synthetic image of a person who doesn’t exist. Important: The deepfake duty doesn’t care about intent. A flattering, harmless AI image of your CEO at an event that never happened is still a deepfake under the Act. Marketing teams generate this kind of content casually. From August, every one of those images needs a label. If You Deploy AI Agents or Chatbots The rule itself is simple: people must know they’re dealing with AI. The provider carries the design obligation, but as the deployer you’re the one putting the system in front of your customers, and you’re the one an EU regulator will contact if your branded assistant pretends to be human. The Act contains an exception for cases where it’s “obvious” the user is talking to AI, judged from the perspective of a reasonably well-informed and observant person. Don’t lean on it. What’s obvious to your product team isn’t obvious to every customer, and the human-sounding voice agents and email-writing AI agents rolling out right now are designed specifically to not feel like software. If an AI agent negotiates a renewal over email or handles a support ticket end to end, disclose it. Pro Tip: Put the Disclosure at the Start of the Interaction Put the disclosure at the start of the interaction, in the interface itself: “You’re chatting with an AI assistant.” A line in your privacy policy doesn’t meet the standard, and a disclosure that appears after the conversation ends is worthless. For voice agents, say it up front in the greeting. What Your AI Vendors Owe You The machine-readable marking obligation in Article 50(2) sits with providers — the companies supplying your generative AI tools. The final Code of Practice expects providers to apply at least two layers of marking where necessary, such as embedded metadata combined with watermarking, and to offer detection mechanisms so deployers, authorities, and researchers can verify whether a piece of content came from AI. One timing caveat: the Digital Omnibus gives generative AI systems already on the market before 2 August 2026 until 2 December 2026 to comply with the marking requirement. Every other Article 50 obligation stays on

How Axipro Guided Technovative Solutions & DigiProd Pass to ISO 27001
DPIA vs Risk Assessment Under GDPR

The CNIL‘s screening rule sounds simple: hit two of the nine high-risk criteria, and you owe a full Data Protection Impact Assessment (DPIA). The trouble starts when you hit one or none, because the GDPR never says that skipping the DPIA means skipping assessment altogether. Plenty of processing falls outside the CNIL’s screening rules: operations below the two-criteria threshold, activities on the CNIL’s exemption list, processing already covered by an earlier DPIA, and controllers who answer to a different supervisory authority altogether. In every one of those cases, the Article 35 GDPR DPIA obligation may fall away while the risk assessment obligations under Articles 24 and 32 stay exactly where they were. This article maps the scenarios where CNIL criteria don’t apply and what a defensible assessment strategy looks like when they don’t. DPIA vs General Risk Assessment: Core Distinctions Under GDPR These two assessments get conflated constantly, and the mix-up has real consequences. They rest on different legal bases, serve different purposes, and trigger under different conditions. Article 35 GDPR requires a DPIA where processing is “likely to result in a high risk” to people’s rights and freedoms, and it requires the assessment before processing begins. The DPIA looks outward. It evaluates the necessity and proportionality of the processing and the risks it creates for data subjects: discrimination, identity theft, financial loss, reputational damage, loss of control over personal data. The measuring stick throughout is harm to people. Article 32 GDPR requires controllers and processors to put in place technical and organizational measures (TOMs) appropriate to the risk of the processing. You can’t know what’s appropriate without assessing that risk first, so Article 32 carries an implicit risk assessment duty for every processing operation you run, high risk or not. Its focus is security: the confidentiality, integrity, availability, and resilience of the systems handling personal data. Article 24 completes the picture by making the controller responsible for implementing measures proportionate to risk and able to demonstrate compliance. That’s the accountability principle at work. So risk assessment is universal, and the DPIA is the escalated version you reserve for processing that crosses the high-risk line. The real question is which assessment to run and how deep to go. You don’t need a six-figure budget to be GDPR compliant. You need a clear plan and someone to do the work. Affordable GDPR Compliance Services Book a Free GDPR Consultation The CNIL Criteria: A Quick Recap The Article 35(3) Baseline and the 9 Criteria Article 35(3) names three situations where a DPIA is always mandatory: systematic and extensive automated evaluation of individuals, including profiling, with legal or similarly significant effects; large-scale processing of special categories of data (Article 9) or criminal conviction data (Article 10); and large-scale systematic monitoring of a publicly accessible area. Beyond those, the WP29 guidelines on DPIAs (WP248 rev.01), endorsed by the European Data Protection Board (EDPB), list nine criteria that indicate likely high-risk processing: evaluation or scoring, including profiling; automated decision-making with legal or similarly significant effect; systematic monitoring; sensitive data or data of a highly personal nature; processing on a large scale; matching or combining datasets; data concerning vulnerable data subjects (employees, patients, children); innovative use or application of new technological or organizational solutions; and processing that prevents data subjects from exercising a right or using a service or contract. The “Two Criteria” Threshold Rule The CNIL’s position is that processing meeting at least two of the nine criteria requires a DPIA as a general rule. WP248 leaves room on both sides of that line: a controller can conclude that processing meeting two criteria still isn’t high risk, and in some cases a single criterion is enough to trigger the obligation. Either way, the reasoning has to be documented. Where there’s genuine doubt, the CNIL’s advice is simple: do the DPIA. CNIL’s List of Processing Operations Requiring a DPIA The CNIL also maintains a mandatory list under Article 35(4), adopted through Deliberation No. 2018-327 of October 11, 2018. It names 14 types of processing that require a DPIA outright, including systematic employee monitoring, whistleblowing schemes, profiling that can exclude people from a contract, and large-scale processing of health data. If your processing appears on this list, you can skip the criteria math because the DPIA is mandatory regardless. Insider Note: The CNIL’s sectoral “referentials” do more work than most DPOs realize. If your processing fully complies with an applicable referential, the CNIL accepts the position that residual risk isn’t high, which takes Article 36 prior consultation off the table. Checking for a referential before scoping a DPIA can remove the most painful step of the entire process. When CNIL Criteria Don’t Apply: Key Scenarios Processing Falling Below the Two-Criteria Threshold Most B2B processing lives here. A standard CRM, a newsletter list, routine supplier management: these might touch one criterion (large scale, perhaps) without hitting a second. No DPIA is required, but the screening itself is a compliance artifact. Record which criteria you tested, what you concluded, and why. If the CNIL inspects, the absence of a DPIA is defensible only when the screening decision is on paper. Operations on CNIL’s Exemption List Article 35(5) lets supervisory authorities publish “whitelists” of processing that doesn’t require a DPIA. The CNIL adopted one in 2019 after an EDPB opinion, covering categories such as routine HR management in organizations with fewer than 250 employees (without profiling, biometrics, or sensitive data), badge-based physical access control without biometrics, and time management systems that don’t process biometric data. France is one of only a few member states with a formal whitelist, which matters for cross-border groups: the same HR system can be exempt in France and assessable case by case in Luxembourg. Processing Authorized by Specific Legal Provisions Article 35(10) carves out processing based on a legal obligation or public interest task under Article 6(1)(c) or (e), where the legal basis regulates the specific operation and a general impact assessment was already carried out when that law was adopted. It’s a narrow