The European Commission's guidelines on the transparency obligations in Article 50 of the AI Act were published today in the Official Journal of the European Union, under the reference C/2026/6005. The 43-page communication, organised in 155 numbered points, sets out how the Commission reads the rules on chatbot disclosure, machine-readable marking of AI output, emotion recognition notices and the labelling of deep fakes and AI-written public interest text. For advertisers, one example stands out: a company established outside the EU that uses an AI system to generate a deep fake of a celebrity for an advertisement displayed in the Union is, according to the guidelines, a deployer within the scope of the AI Act.

In Short

The European Commission's rulebook for labelling AI chatbots, deep fakes and AI-written news was formally published today in the EU's Official Journal, nearly four months after the Commission first released it in July. It matters to anyone making ads with AI for European audiences, including companies based outside the EU, because AI-made celebrity endorsements, synthetic influencers testing products and product images that misrepresent the real item all count as deep fakes that need a visible label. If you only commission an agency and have no say over how it uses AI, the labelling duty sits with the agency, and ad copy and product descriptions stay outside the text rule unless they make health, consumer safety or sustainability claims.

What appeared in the Official Journal

The document published today is the formal C-series version of guidelines the Commission first released on July 20, 2026. PPC Land reported that release the same day, when the text carried the internal reference C(2026) 5054 final and arrived alongside the finalised Code of Practice on Transparency of AI-Generated Content. Today's Official Journal publication comes 115 days after that release and 102 days after Article 50 began to apply on August 2, 2026, according to PPC Land's count of the dates in the two documents.

The guidelines say what they are not. "These Guidelines are non-binding," the text states in point 5, adding that "any authoritative interpretation of the AI Act may ultimately only be given by the Court of Justice of the European Union." They were issued under Article 96(1)(d) of the AI Act and, according to the document, were informed by a broad stakeholder consultation, input from Member States in the AI Board and comments on a draft. The Commission opened that first consultation on September 4, 2025, with a deadline of October 2, 2025. A draft of the guidelines followed on May 8, 2026, with comments due by June 3, according to the timeline in PPC Land's coverage of the EU's free AI labelling icons.

One detail in the dated text shows its age. The section on entry into application still says Article 50 "will apply as from 2 August 2026," a date that has passed. The same section describes the AI Omnibus, the regulation amending the AI Act, as "recently adopted by the Union legislature." When PPC Land covered the provisional agreement on that package on May 7, 2026, formal endorsement by the Council and Parliament was still outstanding.

According to an analysis published on July 30 by the law firm Faegre Drinker, the final July text differed from the May draft in several places: it widened the territorial reach for deep fakes, moved AI translations into the standard-editing exemption and added a business-to-business carve-out. All three appear in the Official Journal version.

Four obligations, two sets of actors

Article 50 contains four separate duties, and the guidelines open with a table that splits them between providers, who build and place AI systems on the market, and deployers, who use them under their own authority.

  • Article 50(1) requires providers of AI systems that interact directly with people to design them so that those people are informed they are dealing with AI.
  • Article 50(2) requires providers of systems that generate or manipulate synthetic image, video, audio or text to mark outputs in a machine-readable format and make them detectable.
  • Article 50(3) requires deployers of emotion recognition and biometric categorisation systems to inform people exposed to them.
  • Article 50(4) requires deployers to disclose deep fakes and AI-generated or manipulated text published to inform the public on matters of public interest.

These can stack. Point 8 of the guidelines describes a system that generates images in conversation with a user: its provider may have to meet both 50(1) and 50(2), and if the images are deep fakes, the deployer may also have to label them under 50(4). Article 50(5) then adds horizontal conditions for all four: information must be "clear and distinguishable," given at the latest at the first interaction or exposure, and must meet accessibility rules.

Who carries the duty in advertising

The guidelines answer a question that agencies and brands have been asking since the summer, and the answer turns on control rather than on who pays. Point 14 uses an "advertising company" as its example of a legal person deploying an AI system. Employees acting under its instructions, such as "digital animators, web designers, content creators, journalists," are not separate deployers. The company "remains a deployer even if it involves third parties (e.g. contractors, freelancers)" to operate the system on its behalf.

The client side gets an explicit carve-out. "A company that merely commissions an advertising agency to produce an advertisement, without taking decisions and exercising control over whether and how the advertising agency uses AI in the production process, is not a deployer," the text reads. That aligns with the reading set out by IAB Austria's guide in September, which concluded that the producing agency is normally the responsible deployer and that a client becomes one when it picks the tool, writes prompts or sets parameters. The Dutch trade body VIA Nederland mapped four agency activities that can trigger disclosure duties on September 5.

Geography matters less than the advertiser might hope. Under point 13, deployers established outside the EU are covered where they foresee dissemination of AI outputs in the Union, "including by posting deep fakes on the globally accessible internet." The celebrity advertisement example in point 14 applies that rule directly. A third-country deployer escapes the obligation only where content reaches EU audiences "through channels that are unforeseeable and outside their control."

Distribution intermediaries sit outside the definition. Hosting services, online platforms and broadcasters whose role is limited to transmitting AI content created by others "are not deployers within the meaning of the AI Act," point 16 states, though they are "strongly encouraged" to preserve marks and labels. The position changes when a platform uses AI for itself. According to point 126, providers of VLOPs and very large online search engines become deployers when they use AI under their own authority "for their own professional purposes (e.g. creating visuals to be used as part of marketing material)." Where such platforms offer labelling tools that meet Article 50(4), deployers posting on those services can rely on them, without shifting their own responsibility.

For deployers in long content supply chains, point 12 expects "proportionate measures" to keep labels visible at the point of first exposure, naming contractual conditions with distributing partners and interface settings as examples.

When an advertisement becomes a deep fake

The AI Act defines a deep fake as AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear to be authentic or truthful. The guidelines break that into four cumulative criteria and then read "existing" broadly. It is enough, according to point 113, for simulated subjects to resemble someone or something that "exists, can plausibly exist or could have plausibly existed in reality." A realistic AI-generated human who never lived can therefore qualify. Dragons, flying humans and "elephants driving cars" do not.

Intent is irrelevant. The assessment "is objective and does not require the intention of the deployer to deceive or mislead," point 114 states. Photorealism raises the likelihood of a deep fake finding, "but photorealism alone is not determinative." The audience test is also stricter than the one applied to chatbots. Rather than an average person, deployers must consider the "possible diverse composition of the reasonably foreseeable audience," including children and the elderly. Content shown only in a subscriber-only area of a website or in a corporate newsletter does not have to be assessed as if it were broadly public.

The examples matter more than the definitions for creative teams. Among deep fakes, the guidelines list:

  • an AI-generated video featuring an AI-generated depiction of a celebrity influencer "in an advertising or promotional context";
  • an AI-generated video of a realistic synthetic avatar of a company CEO congratulating employees on the year's results;
  • "an AI-generated image of a product in advertisement or packaging that can affect the audience's perception and mislead as to the actual product appearance, characteristics or use (e.g. making the product appear not identical to the real product, more appealing or with improved quality than in real life)."

Among content that is not a deep fake: an "AI-generated video of mice arguing in human language over the best type of cheese as part of an advertisement campaign for a cheese manufacturer," and a real car shown against an AI-generated background, "as long as the ad is not likely to mislead the audience about the product's actual representation and its characteristics and use."

Routine production work is largely spared. According to point 116, "AI-powered colour correction, background extensions of existing content, adjustments or replacements of backgrounds for clearly aesthetic purposes, compositions and arrangements of existing products, or re-scaling of images applied in product advertisements or packaging is likely to have only a minor impact" on a person's perception of authenticity. Substantial AI editing of background details in journalistic images, by contrast, may cross the line.

The creative carve-out is narrow for ads

Article 50(4) offers a lighter regime for deep fakes in "evidently artistic, creative, satirical, fictional or analogous" works: disclosure is still required, but in a manner "that does not hamper the display or enjoyment of the work." The guidelines insist that "evidently" be read strictly. Content that is "exclusively informative or commercial and is recognisable as such" falls outside, and where content mixes characters, "the informative character should always prevail."

Advertising sits in a grey zone. Point 122 says advertisements or documentaries containing deep fakes "might be regarded as evidently creative or fictional works in certain, specific situations, but not in others." The examples leave little doubt about the commercial formats most exposed. Two cases are explicitly excluded from the creative regime: an AI-manipulated teleshopping-style video with simulated consumers using a product "with the aim of persuading viewers to buy the product," and an AI-manipulated video "featuring a realistic synthetic influencer testing out a sponsored real product, focused solely on displaying the product functionalities." That nuance is worth setting beside PPC Land's July summary that the artistic carve-out does not extend to advertising. The text does not close the door entirely; it makes the exception case-specific.

What a label must look like

Deep fake labels must be "understandable and perceivable by natural persons (e.g. with visible or audible labels), without them needing to rely on any specific technical tools." That has one direct consequence: deployers "cannot rely on the machine-readable marking embedded in the content by the provider." A watermark applied by the image generator does not discharge the advertiser's or agency's duty.

Timing follows Article 50(5). For each output and each person exposed, disclosure must arrive at first exposure. Point 143 adds that where viewers may not see content from its beginning, a label only at the start "does not adequately inform those persons" and should be complemented by later disclosure where possible. Scrolling past an AI-manipulated deep fake on social media is one of the document's examples of first exposure.

A label does not make content lawful. Point 129 says labelling does not permit deep fakes that are unlawful under other rules, naming "misleading advertising" alongside child sexual abuse material and non-consensual intimate images. Deployers depicting real people must also comply with data protection law, intellectual property law and personality rights over images and voices.

Ad copy, product descriptions and public interest text

The text-labelling rule in Article 50(4) is narrower than many marketers assumed. It covers AI-generated or manipulated text that is published, aims to inform the public and concerns matters of public interest. The guidelines place "AI-manipulated text that is part of a company's advertisement or product descriptions" outside that scope, with a parenthetical exception for "any claims related to e.g. health, consumer safety or sustainability." AI-generated fantasy novels, a chatbot news summary seen only by the person who prompted it, and a consultant's compliance memo to a client are also out.

In scope are an AI summary of a newspaper article on a town council decision, AI-manipulated parts of a lifestyle-website article comparing diets, "AI-manipulated corporate reports published on a listed company's website containing investor information," and a meteorological institute's AI-written storm warning on social media.

The exception for human oversight is demanding. Text escapes labelling only if it has undergone human review or editorial control and a natural or legal person holds editorial responsibility. "Fact-checking the accuracy of the content is a minimum requirement," point 134 states. Spell-checks, "the mere existence of an editorial policy," automated review or "cursory editorial approval" do not qualify. Timing can undo compliance: "Any substantive AI intervention occurring after the human review or editorial control process has taken place will therefore cause the exception to become void." The identity and contact details of whoever holds editorial responsibility should be publicly available, for example in website terms or a colophon. One example that fails the test: AI-generated articles "reviewed and edited by another AI system" where a human performs only a superficial grammatical check.

For publishers, the guidelines tie editorial responsibility to the definition in the European Media Freedom Act and allow media service providers to rely on existing editorial processes.

Chatbots, shopping assistants and AI agents

Article 50(1) covers AI systems with "a genuine conversational or responsive character." In scope are chatbots in customer support, complaints handling, e-commerce and finance, AI companions, avatars, AI bots on social networks and coding agents. Out of scope are recommender systems, spam filters, AI search and retrieval "without generating or modifying content," text auto-completion and "AI-assisted product visualisation (e.g., virtual try-on, room staging, product rendering)." Systems where a human customer service agent uses AI behind the scenes are also outside, though products that blend AI responses with human-curated content need disclosure for the AI-generated parts.

AI agents receive specific treatment. Point 31 says they "must be designed and developed in such a way that they disclose both their artificial nature and the person on whose behalf they are acting," whether booking, negotiating, managing correspondence or executing purchases. Where a provider cannot know in advance whether an agent will meet a human, the agent should be built to disclose itself "in every situation where it is reasonably likely" that it may interact with one. Agents should also disclose themselves to the people instructing them at key steps such as authorisation and reporting. A footnote points to electronic attestations under Regulation (EU) No 910/2014, the eIDAS regulation, EU Digital Identity Wallets and the proposed European Business Wallets as ways to verify agent identity.

The guidelines list techniques that do not work on their own: disclosures only in terms and conditions, URLs or documentation; machine-readable marks that users cannot perceive; "generic references to 'assistant'"; general statements such as "Services on this website use AI"; and technical descriptions like "this system uses LLMs." Suggested approaches include a first-turn greeting, persistent badges, spoken statements in voice channels and standardised visual indicators. The text warns against overuse and "banner blindness," but says periodic reminders are likely necessary for vulnerable users and in contexts such as financial advice, insurance, health advice, complaints handling and AI companions.

The exemption for interactions that are "obvious" is to be interpreted restrictively and limited to cases with "almost no doubt left about the nature of the interaction." Code assistants used only by professional developers, internal staff tools for AI-literate employees and non-player characters in single-player games qualify. Helpdesk chatbots on online platforms do not. Consumer law runs in parallel: the guidelines note that where a service is AI-driven, such as a subscription chatbot or AI companion app, the AI functionality "may qualify as an essential characteristic" under the Consumer Rights Directive, regardless of whether the interaction is obvious.

Marking and detection for generative tool providers

For providers of generative systems, Article 50(2) requires two things at once: machine-readable marking and the means to detect it. "Fulfilling only one element," point 70 states, "will not suffice." Recital 133 of the AI Act lists watermarks, metadata, cryptographic provenance methods, logging and fingerprints as possible techniques. Detection must ultimately rely on "publicly-available industry standard detection solutions" that any third party can implement, ideally locally executable. Proprietary or shared detectors are allowed in the interim, provided they interoperate, until harmonised standards emerge.

Advertising production gets a specific clarification. Outputs used only in "closed loop environments in industrial and product development workflows (for example for film, animation, games or advertising production)" need no marking unless they are the final output. Short outputs such as image captions, alt-text and UI labels are excluded, as are source code, JSON and other machine-readable formats, and agent-to-agent traffic. A business-to-business exemption applies only when output is strictly technical, seen by a limited predefined group of professionals and not intended to leave the company.

The standard-editing exception covers grammar correction, AI translation, minor cropping and colour adjustment, red-eye removal and background blurring. Marking is required for AI summaries, rewriting that changes meaning, object or person removal that changes the substance of an image, face replacement, voice synthesis of a specific person and alterations to body shape or skin colour. Point 94 adds a data protection limit: for marking and detection, "the information related to the creator of the content should not be processed."

Industry implementation has moved since July. Google added "How this ad was made" disclosures on July 9, citing SynthID and C2PA as its machine-readable layer, and added an AI attestation field to the DV360 API on July 23. Anthropic began watermarking Claude text and signing files with provenance metadata in August. Yet a study of Meta ads published in October found that only 2 of 3,379 ad files flagged as AI-generated disclosed AI use in the ad copy, and none of 971 files checked carried C2PA credentials.

Emotion recognition and biometric categorisation

Article 50(3) applies to deployers. Anyone operating an emotion recognition or biometric categorisation system must tell the people exposed, including children, at the latest at first exposure. The guidelines note that Article 50(3) applies to any biometric categorisation system, "e.g. age or gender classification based on biometric data," whether or not it is classified as high-risk. Their examples include a pop-up before a computer game indicating that the player's face is recorded to capture emotions, and a visible notice at each entrance to an exhibition room where visitors' faces are captured to assign them to an age group. Shops and train stations are among the deployment settings the text names. Emotion recognition in workplaces and education remains prohibited under Article 5.

Codes, enforcement and the calendar

The guidelines give the Code of Practice on Transparency of AI-Generated Content a central role. Adherence to a code assessed as adequate is described as "the only Union-wide recognised practical framework" for demonstrating compliance with Articles 50(2) and (4). For signatories, the Commission and market surveillance authorities "will focus their supervisory activities" on whether the code was followed, and commitments may count as a mitigating factor when fines are set. Non-signatories are expected to run a gap analysis against the code and "may also be subject to a larger number of requests for information and requests for access." Opting out of code sections forfeits the benefit for those sections. Google signed the code on July 24, two days after the deadline for the presumption of compliance, and Meta followed on July 28. A footnote in today's text dates the Commission's information page on the code to June 10, 2026.

Enforcement sits with national market surveillance authorities, the European Data Protection Supervisor for EU institutions, and the AI Office for AI systems built on a general-purpose model from the same provider. Any person with grounds to consider that Article 50 has been violated may lodge a complaint. Fines for breaching Article 50 reach EUR 15,000,000 or 3% of total worldwide annual turnover for the preceding financial year, whichever is higher. For SMEs and start-ups, the lower of the two applies. EU institutions face up to EUR 750,000.

The calendar is fixed by Article 113 of the AI Act and the AI Omnibus. All in-scope systems on the EU market had to comply with Article 50 from August 2, 2026, regardless of when they were launched. The only grandfathering covers the marking and detection duty in Article 50(2) for generative systems placed on the market before that date, which have until December 2, 2026, 20 days from today. Systems that are partly interactive and partly generative benefit only for the marking element; chatbot disclosure has been due since August 2. The code itself sets February 2, 2027 as the date for interoperability under Measure 3.4.

Retroactivity is limited. Outputs and deep fakes generated before August 2, 2026 "do not need to be marked or labelled retroactively," and the same applies to public interest text published before that date. Text generated before August 2 but published after it must be labelled. Holders of older, unlabelled deep fakes are encouraged to label them without "disproportionate efforts," which the guidelines describe as "such as auditing of pre-existing content databases or modifying already printed product packaging."

Why it matters for marketers

The guidelines move the compliance question away from whether AI was used and toward who controlled it and whether an audience could be misled. For agencies, the text confirms that they usually carry the labelling duty, including when they use third-party tools. For brands, the boundary is the degree of control: choosing tools or directing AI use can turn a client into a deployer. For non-EU advertisers buying European reach, the celebrity example removes any doubt about territorial scope.

The product-image example deserves particular attention in retail and commerce advertising, where AI-enhanced packshots are now routine. An image that makes a product look "more appealing or with improved quality than in real life" is treated as a deep fake, while background swaps and colour work for aesthetic purposes generally are not. Platforms have started to impose their own lines; Etsy now requires at least one unaltered photo in every listing.

There is a commercial cost to labels as well as a legal one. The IAB's second AI disclosure framework, published in August, cited an NYU Stern finding that disclosing generative AI cut ad click-through by 31.5%. The Commission's own text concedes the tension between effective disclosure and user experience, asking for labels calibrated "to remain effective, user-friendly and avoid, to the extent, possible disruptions of the user experience."

The guidelines describe themselves as "a first interpretation" and say the Commission will review them as practical experience, market surveillance cases and rulings of the Court of Justice accumulate.

Timeline

Summary

Who: The European Commission, which issued the guidelines; providers of AI systems, including generative tools, chatbots and AI agents; deployers such as advertising companies, agencies, publishers and listed companies; and the market surveillance authorities and AI Office that enforce Article 50.

What: The Commission's guidelines on the transparency obligations in Article 50 of the AI Act were published in the Official Journal as C/2026/6005. They explain chatbot and AI agent disclosure, machine-readable marking and detection, emotion recognition notices, and the labelling of deep fakes and public interest text, with examples that classify celebrity-likeness ads, synthetic influencers testing products and misleading AI product images as deep fakes, while placing ordinary ad copy and product descriptions outside the text-labelling rule.

When: Official Journal publication today, November 12, 2026, following the Commission's first release on July 20, 2026. Article 50 has applied since August 2, 2026, and generative systems already on the market before that date must meet the marking obligation by December 2, 2026.

Where: The European Union, with reach to companies established outside the EU whose AI-generated deep fakes are distributed to audiences in the Union.

Why: According to the guidelines, the transparency rules aim to reduce impersonation, deception, misinformation, manipulation at scale and fraud. For marketers, the text settles who carries the labelling duty in agency relationships and where common AI production techniques fall, under a penalty regime of up to EUR 15,000,000 or 3% of worldwide turnover.