AI Act 2026: Transparency of AI Content in Marketing

By Million MarketingReading time: 12 min

Since 2 August 2026, an AI-generated product visual, a synthetic voice in an advertising spot or an article optimised for generative AI published without any mention of its artificial nature is no longer merely a matter of editorial ethics: it is a breach of European law, punishable by a fine of up to €15 million or 3% of the company’s worldwide turnover [1][6]. The European regulation on artificial intelligence (AI Act, Regulation (EU) 2024/1689) entered into force on 1 August 2024, but it is its Article 50 — dedicated to the transparency of synthetic content — that is now reshaping the production chains of every digital marketing player using generative AI: agencies, marketing departments of large accounts, mid-caps, content creators and SEO/AIO providers.

This article details exactly what Article 50 says, the application timeline revised by the “Digital Omnibus on AI”, the specific obligations of general-purpose AI model (GPAI) providers, the sanctions and enforcement regime in France, and the practical consequences for agencies and their clients who produce advertising content, visuals and videos, or who run an AI-assisted generative SEO (AIO) strategy.

The new framework: the AI Act and Article 50 on one page

Regulation (EU) 2024/1689 is based on a classification of AI systems by level of risk: unacceptable risk (prohibited), high risk (heavy obligations), limited risk (enhanced transparency) and minimal risk. The generative AI tools widely used in marketing — text, image, video and audio generators, conversational agents — fall into the “limited risk” category and come exclusively under Article 50 [1][2].

The text distinguishes two types of players, and the allocation of responsibilities between them is central to understanding who must do what:

The provider: the entity that develops an AI system and places it on the market under its own name (OpenAI, Google, Midjourney, Adobe…).
The deployer: the organisation that uses this system in the course of its professional activity — an advertising agency, a marketing department, a large account or a mid-cap [2][6].

For an agency like ours, or for our clients, the question is therefore never theoretical: as soon as advertising content, a campaign visual or an SEO text is produced with the help of a generative AI tool, the company that uses it takes on the role of deployer — with direct transparency obligations towards the public.

Application timeline: what changes in 2025, 2026 and 2027

The AI Act’s initial timeline was substantially reworked by amending Regulation (EU) 2026/1744, known as the “Digital Omnibus on AI”, adopted by the European Parliament on 16 June 2026 and in force since 27 July 2026 [4]. This text grants targeted technical grace periods while keeping regulatory pressure on the substantive obligations.

Deadline	Obligation	Scope for marketing
2 February 2025	Ban on unacceptable-risk practices + AI literacy (art. 4)	Companies must train their teams in the use of AI
2 August 2025	Obligations of general-purpose AI model (GPAI) providers	Technical documentation and copyright compliance for LLM publishers
2 August 2026	Full application of Article 50 (transparency)	Mandatory labelling of deepfakes, AI texts of public interest and chatbots
2 December 2026	Machine-readable marking (art. 50.2) for systems already on the market	Technical grace period introduced by the Digital Omnibus
2 December 2027	High-risk AI (Annex III)	Postponed from August 2026 by the Digital Omnibus
2 August 2028	High-risk AI embedded in regulated products (Annex I)	Final phase, outside the pure marketing scope

The key point for any organisation producing content today: the visible labelling obligation has been fully applicable since 2 August 2026. Only the technical requirement for retroactive machine-readable marking benefits from a reprieve until 2 December 2026 [3][4].

Article 50, point by point: who must label what

Article 50 is made up of five paragraphs whose operational scope differs sharply depending on the marketing use concerned.

Article 50.1 — direct interactions (chatbots, conversational agents)

Any system that interacts directly with a natural person must clearly disclose its artificial nature from the very first interaction, unless this is obvious from the context [2]. This directly concerns customer-service chatbots, voice assistants and conversational agents deployed on an e-commerce site or a lead-generation page — a scope we cover directly in our deployments of conversational AI agents for our large-account and mid-cap clients.

Article 50.2 — technical marking, the providers’ responsibility

Providers of generative AI systems must embed machine-readable marking (watermarks, metadata) in the content produced, in a robust and interoperable way [1][2]. An agency is not responsible for developing this technology, but it depends on its reliability to prove its own compliance in the event of an inspection.

Article 50.4 — visible labelling, the deployers’ responsibility

This is the provision that most directly concerns agencies and advertisers. It requires two types of content to be clearly disclosed:

Deepfakes — a deliberately broad definition covering any generated or manipulated content that resembles existing persons, objects, places or entities and could appear authentic to a reasonably attentive person [2][6]. A photorealistic virtual model wearing a collection, or an AI-generated living-room set for a furniture catalogue, both fall within this scope.
AI-generated texts on matters of public interest (health, environment, economy, public safety, democratic debate).

An exemption exists for texts: it applies once a substantial, documented human review has taken place under the editorial responsibility of an identified person. A simple automated spell-check is not enough to trigger it [2][6].

GPAI and the European code of practice

To operationalise Article 50, the European Commission, through its AI Office, published in early 2026 a code of practice on the transparency of AI-generated content, signed by nearly 190 technology organisations by the end of July 2026 [5]. This code, structured into a “providers” section (technical marking, C2PA, watermarks) and a “deployers” section (visible labelling methods, standardised iconography), remains voluntary — but signing up to it confers a presumption of compliance. A company that departs from it must demonstrate for itself that its method is equivalent.

Sanctions and supervisory authorities in France

The AI Act’s sanctions regime approaches the scale of the GDPR’s: up to €15 million or 3% of annual worldwide turnover for a breach of Article 50 or of the GPAI rules, and up to €7.5 million or 1% of worldwide turnover for misleading information supplied to a supervisory authority [1][6].

In France, supervision is shared between several specialised authorities [7][8][9]:

The DGCCRF (consumer protection authority) — national coordinator, competent for misleading commercial practices, Article 50.1 (chatbots) and Article 50.2 (technical marking by providers).
ARCOM (audiovisual and digital communication regulator) — exclusive competence over the labelling of deepfakes and texts of public interest (Article 50.4).
The CNIL (data protection authority) — supervises biometric and emotion recognition (Article 50.3), in continuity with its GDPR role.
PEReN (the French centre of expertise for digital regulation) — provides technical support to audit watermarks and extract provenance manifests.

The technical Achilles’ heel: C2PA, watermarks and the “stripping” problem

The industry has largely rallied behind the C2PA standard (Coalition for Content Provenance and Authenticity, backed by Adobe, Microsoft, OpenAI and Google), which embeds cryptographic metadata in a file — origin, generation tool, edit history [10]. The problem: this so-called “hard binding” method is systematically stripped when content is uploaded to Instagram, X, YouTube, Facebook or WhatsApp, which remove this metadata for bandwidth and privacy reasons. An advertising visual that is perfectly compliant when created can therefore lose all technical proof of compliance at the very moment it is published.

The industry’s answer is “soft binding” (or cloud binding): an invisible pixel-level watermark — like Google’s SynthID system — that withstands compression, cropping and even screenshots, and allows the original manifest to be retrieved from a cloud database. The European code of practice now requires at least two combined layers of marking to meet the robustness criterion [10].

Under the dual pressure of the AI Act and the Digital Services Act, the major platforms have anticipated the move with their own rules: TikTok labelled more than 3 billion videos through its “AIGC” button in 2026 and joined the C2PA steering committee; Meta applies an automatic “Made with AI” label and prohibits the use of its own generative tools for political advertising; Google and YouTube rely on SynthID and require an “altered or synthetic content” box to be ticked in Google Ads. Non-compliance now exposes advertisers not only to regulatory sanctions but also to the risk of algorithmic “shadowbanning” of their campaigns.

The impact on generative SEO: AIO and the citation economy

Article 50 is not only about visuals: it also redefines the rules of the game for text content, at the very moment traditional SEO is shifting towards AIO (Artificial Intelligence Optimization). Since the March 2026 update, Google’s AI Overviews are triggered on nearly 40% of queries, pushing the zero-click rate to around 69% and cutting clicks by up to 58% for some sites that used to rank first [13].

In this environment, being cited — not just indexed — becomes the goal. AIO practitioners now track specific indicators: citation frequency (CF), citation prominence (CP), citation accuracy (CA) and citation share of voice versus competitors (CSoV). These indicators reward the very signal that Article 50 is looking for: expertise and human editorial responsibility, criteria at the heart of search engines’ E-E-A-T principle. Content mass-produced by AI without qualitative review risks a double penalty — algorithmic de-indexing and the absence of a mandatory compliance label. This is precisely one of the workstreams we structure with our clients as part of our AIO services: regulatory compliance and AI citation performance converge on the same requirement for editorial governance.

What the industry says (and doesn’t say): the debate between the ARPP and Brussels

A study by the World Federation of Advertisers (WFA), conducted in April 2026 among 27 multinationals representing $71 billion in advertising investment, shows that 78% of advertisers already use AI-generated or AI-enhanced content — 87% for product visuals, 80% for copywriting [11]. But the same study reveals a tension: 82% of brands see opacity about AI use as a major reputational risk, and 80% are calling for clear global guidelines. There is a strong consensus (91%) on labelling a synthetic human in a central role in an advertisement — but only 4% of brands consider it necessary to label a simple AI-generated background.

In France, the Union des Marques (the French advertisers’ association) has been circulating a “Responsible Generative AI” guide with audit grids for large accounts since early 2025 [12]. The ARPP (the French advertising self-regulatory body), for its part, published on 8 July 2026 a position more nuanced than the spirit of the AI Act: its Advertising Ethics Council opposes the systematic labelling of all advertising visuals, considering that fiction and staging are part of advertising’s tacit contract, and reserves explicit disclosure for cases of real harm through deception — an AI model directly embodying a cosmetics brand’s promise, or a fake seaside destination visual for a tour operator [13]. This position, more permissive than the European text, does not however exempt any company from complying with Article 50, which remains the binding standard.

This debate is not new for luxury and prestige brands: the French law of 9 June 2023 on influencers already required a “Virtual images” notice for any AI-generated face or silhouette in commercial content — a topic we covered in detail in our article on the legal framework of influencer marketing for luxury houses. The AI Act does not replace this obligation: it extends and harmonises it at European level, with its own sanctions regime.

What this concretely changes for agencies and their clients

For the marketing department of a large account, a mid-cap or an agency, 2 August 2026 is not an administrative deadline: it is a change in working methods. Five concrete workstreams stand out:

Map the AI tools in use (Adobe Firefly, Midjourney, integrated OpenAI APIs…) and check, before each licence renewal, that the provider implements C2PA and, in time, soft binding.
Archive proof of compliance in document management systems (DAM/CMS), since social networks strip metadata on publication — keep the master files and log the human review (who, when, what kind of change).
Train teams in AI (literacy obligation, Article 4) and keep training records, especially for teams working with prompts and sensitive data.
Revise agency–advertiser contracts to include clauses on delivering C2PA-certified files, a written statement of the tools used per deliverable, and a clear allocation of responsibilities if labelling is omitted.
Build a proactive visual label into the creative itself rather than relying on the platforms’ automatic detection, which often fails.

This link between creative production and compliance governance is exactly what we build into our AI Studio & Creation: every visual or video generated or retouched with AI for our clients is produced under documented human art direction, with traceability of the tools used and, from now on, the notices required by Article 50 built in from the design stage rather than added afterwards.

Our approach at Million Marketing

In our Le Remix methodology, we treat regulatory compliance not as a constraint downstream of creation, but as a brief parameter in the same way as media targeting or the conversion objective. For our large-account and mid-cap clients deploying generative AI in their advertising, social media or SEO/AIO strategy, this means auditing tools before deploying them at scale, systematically documenting the human review of text content, and building AI visuals with transparency notices integrated into the brand guidelines — not as a badge added at the last minute. At a time when legal compliance and citation performance in generative engines (AIO) demand the same editorial discipline, brands that structure this governance now are gaining a lead that will extend well beyond the 2 August 2026 deadline.

Sources and references

What is Article 50 of the AI Act and who does it apply to?

Article 50 of the European AI Regulation (EU 2024/1689) imposes transparency obligations for synthetic content. It distinguishes providers of generative AI systems (OpenAI, Google, Midjourney…), who must embed machine-readable technical marking, from deployers — agencies, advertisers, large accounts — who must visibly label deepfakes and AI-generated texts on matters of public interest.

Since when has the transparency obligation for AI content applied?

The visible labelling obligation under Article 50 has been fully applicable since 2 August 2026. A technical grace period, introduced by the Digital Omnibus on AI (Regulation (EU) 2026/1744), runs until 2 December 2026 for the machine-readable digital marking of systems already on the market before that date.

Must AI-generated text content always be labelled?

No. The obligation only applies to texts published on matters of public interest (health, economy, public safety, democratic debate). It no longer applies if the text has undergone a substantial, documented human review under the editorial responsibility of an identified person — a simple automated spell-check is not enough to trigger this exemption.

What are the sanctions for non-compliance with Article 50?

Breaches of the transparency obligations of Article 50 and of the rules applicable to general-purpose AI model (GPAI) providers can lead to fines of up to €15 million or 3% of the company’s annual worldwide turnover, whichever is higher. In France, the DGCCRF coordinates the supervisory authorities (ARCOM for deepfakes, CNIL for biometrics), with technical support from PEReN.