Skip to content
Rafael Patron

Note Notes 2026

Watermarking AI-generated content: what the AI Act requires from 2026

Translated from the Italian original.

In May 2026 the advertising group WPP announced that all AI-generated content produced by its agencies for European campaigns will be watermark-protected according to the C2PA standard. The decision anticipates by a few months the full obligations of Article 50 of the AI Act, applicable from August 2, 2026. The group’s message to clients is twofold: compliance with the regulation and competitive differentiation. Mid-sized Italian agencies are asking themselves whether to replicate the approach or wait.

The point is that the choice is not only technical. It is strategic.

What Article 50 actually says

The operational application of the AI Act to digital marketing is the frame of reference for editorial choices. Article 50 of the AI Act imposes transparency obligations in three directions. First, providers of AI systems generating synthetic audio, image, video or text content must ensure that outputs are marked in a machine-readable format and identifiable as artificially generated or manipulated. Second, deployers of deepfakes (synthetic images, audio or video that appreciably resemble real persons, objects or events) must disclose it. Third, deployers of systems that generate or manipulate text published to inform the public on matters of public interest must disclose its synthetic nature.

These obligations apply from August 2, 2026. Exemptions are limited: evidently artistic, parodic or satirical content is exempt from the deepfake labeling obligation. Content for law enforcement purposes has a special regime. Everything else is in scope.

Many Italian brands assume this is a problem only for those who create purely generative content. That is a restrictive interpretation. Even a product image retouched with generative AI in substantial parts (replaced backgrounds, added objects, generated models) could fall within the perimeter depending on context and on the intensity of the manipulation.

Visible watermark vs. cryptographic watermark: the two technical routes

The evolution of generative models directly shapes the characteristics of applicable watermarks. Available marking technologies fall into two families. Visible or overt watermarks: text labels (“AI-generated”), graphic badges, disclosures in EXIF metadata readable by anyone. Cryptographic or invisible watermarks: noise patterns inserted into pixels or sound, detectable by dedicated tools but not perceptible to the end user.

Google DeepMind’s SynthID is the best-known cryptographic watermark, integrated into Imagen, Veo and the Gemini text generation models. OpenAI has developed similar systems for DALL-E and Sora. Adobe Firefly applies Content Credentials metadata according to the C2PA standard. Stable Diffusion includes markers in some versions and not in others, depending on the distribution.

The operational problem is robustness. A well-made cryptographic watermark survives cropping, resizing, moderate recompression. It survives poorly heavy edits, repeated screenshots, some targeted transformations. For Italian deployers the consequence is that the watermark alone is not enough. An upstream traceability system is needed (a register of generated content, historical hashes) plus explicit disclosure at publication.

The C2PA standard: why it is becoming the reference

The Coalition for Content Provenance and Authenticity, an initiative promoted by Adobe, BBC, Microsoft, Intel and others, developed C2PA as an open standard for the provenance of digital content. Adobe made it the basis of its Content Credentials, BBC and other news organizations use it to prove the authenticity of journalistic content, Sony and Nikon integrate it into professional cameras to authenticate original shots.

The idea is to record in signed metadata a history of how the content was created and modified. If it was generated by AI, the metadata says so. If it was modified, the metadata says so. If it is an authentic photo taken with a certified device, the metadata says so and no one can remove it without invalidating the signature.

For Italian advertising agencies, C2PA is becoming the de facto standard for professional content. Working with Adobe Creative Cloud means working with C2PA. Publishing on platforms that read Content Credentials (LinkedIn, for example, is implementing support in 2026) means automatic disclosure.

Practical cases: what changes in three real workflows

Case 1: e-commerce product photo with replaced background. Model photographed in the studio, background replaced with generative AI for catalog consistency. The real object is the one shown, the context is synthetic. Under the emerging framework, labeling is recommended; the artistic content exemption does not clearly apply. Best practice: disclosure in metadata (C2PA), no visible disclosure needed on the published image if the manipulation is limited to the context.

Case 2: commercial with a synthesized celebrity voice. Voice cloning, even when authorized by the celebrity, clearly falls within the definition of deepfake (synthetic audio that appreciably resembles a real person). Explicit disclosure is required. Best practice: a statement in the final frame or in the credits of the spot (“voice generated with AI with the artist’s authorization”). Enterprise clients are already requiring this kind of transparency in licensing contracts.

Case 3: editorial article written in collaboration with AI for an online publication. The article informs the public on matters of public interest and needs disclosure under Article 50. Best practice: a clear editorial statement in the article footer (“This article was drafted with the support of AI systems, editorial review by X”). Many Italian publications are already moving in this direction.

The hidden risk: liability along the chain

What many Italian deployers do not consider is liability along the chain. The provider of the AI system applies a watermark. The deployer publishes the content. A third party modifies or redistributes the content, removing the watermark. Who is liable?

The structure of the AI Act distributes liability. The provider is liable if the watermark is technically inadequate. The deployer is liable if it does not ensure disclosure in its own publication. The third party is liable if it deliberately removes the watermark to deceive. But the evidentiary picture in case of litigation is complex.

For Italian agencies the consequence is to document the entire flow. Which tool generated the content, when, with which version, what the state of the metadata was at delivery to the client, what the client’s policy is on keeping watermarks at publication. AIPIA is developing operational guidelines on this content authenticity supply chain for members working in the creative and media sector.

The real competitive cost of AI transparency

There is a school of thought, especially among the most aggressive marketers, that sees disclosure as a commercial brake. “If we say the voice is AI-generated, the consumer reacts badly.” Recent market research says the opposite. An Edelman Trust 2026 study on European consumers shows that 67% of respondents declare greater trust in brands that openly disclose the use of AI in content, against 23% who declare less trust.

In other words: disclosure is not only an obligation, it is a reputational asset. WPP and other large groups understood this first. Italian agencies that move in time will be able to turn compliance into differentiation. Those that postpone will find themselves doing the legal minimum, without the positioning advantage.

The agenda for the coming months for those operating in Italy

Internal audit of AI-generated content produced in the last 12 months. Definition of a company content policy on AI transparency. Selection of watermarking and disclosure tools (C2PA-compatible where possible). Update of contracts with clients and suppliers to include clauses on AI-generated content. Training of the creative and media team on the subject. A simple checklist, but with a real timeline, to be closed by August 2, 2026.

Those who reach August 3 without these six things done are not automatically non-compliant, but they are exposed to risk. Those who have done them can market the approach as a qualifying element. The regulation is the same for everyone. How it is made one’s own is what distinguishes companies.

In the same section

The whole section

Advisory, talks and contact

Contact Rafael

Advisory, speaking and contact requests are handled on rafaelpatron.it.