Since 2 August, disclosing that content has been generated or manipulated using artificial intelligence has ceased to be a good practice and become an enforceable obligation. Its timeline did not change despite the postponement of the rest of the regulation, the editorial exception is narrower than many assume, and the technical response is already becoming standardised without anyone explicitly imposing it

For almost two years, the debate around artificial intelligence in the European audiovisual sector has revolved around the same question: what can be done? The relevant question this autumn is a different one, considerably less glamorous and much more urgent: what has to be disclosed?

On 2 August 2026, Article 50 of the European Union’s AI Act, which sets out transparency obligations, entered into application. And there is one detail that has been lost in the noise of recent weeks: its timeline did not change. The digital omnibus package approved by the European legislature in July postponed the obligations concerning high-risk systems listed in Annex III until December 2027, creating a widespread impression that “the AI rules have been delayed”. They have not. The transparency requirements stayed where they were, and they are now enforceable.

The two obligations that reach the newsroom

The regulation distinguishes between the provider of an AI system and the entity deploying it. Providers of generative AI systems are required to mark their synthetic outputs in a machine-readable format, allowing a third party to detect that a file is artificial. This is a technical obligation, and systems already on the market before 2 August have until 2 December to adapt.

For deployers — which is what a broadcaster or production company becomes when it integrates these tools into its workflow — two different obligations apply, and they should not be confused.

The first concerns deepfakes, the term used by the regulation for manipulated or synthetic content that constitutes an ultra-realistic impersonation: their artificial nature must be disclosed clearly and distinguishably, in a way that can be understood without special tools, and no later than the first exposure to the viewer. For artistic, satirical or fictional works, the threshold is lower: an appropriate disclosure is sufficient, provided that it does not hinder the exhibition or enjoyment of the work.

The second concerns AI-generated or manipulated text published to inform the public about matters of public interest. Politics, public administration, justice, rights, security, health and the environment all fall within this category. This is where the exception that interests everyone comes in — and where it is most often misunderstood.

The editorial exception exists, but it is narrower than it appears

Informational text is exempt from labelling when it has undergone substantive human review and editorial control, with someone assuming responsibility for its publication. This may sound like a broad exemption, and for a professional newsroom it can be significant. But there are two limits that need to be made clear before considering the matter settled.

The first is substantive. The Commission has clarified that this review requires deliberate examination of the content by people with relevant knowledge, including verification, and that editorial control means having effective authority to approve, alter or reject it. A spell-check does not count. The exemption protects journalism that is actually carried out, not journalism that is merely declared.

The second is its scope, and this is the one that matters most to the audiovisual industry: the exception covers text, not images, audio or video. In practical terms, an AI-generated recreation of an event, a virtual studio simulating a real location or a synthetic likeness of a contributor is not exempt simply because a news editor has approved it. If it constitutes a deepfake, it must be disclosed.

Why the European industry challenged this wording

The fact that European broadcasters have debated the details of these guidelines is not simply a corporate concern, but a technical warning worth considering. On 17 June, the European Broadcasting Union (EBU), together with the main associations representing commercial broadcasters across Europe, signed a letter to the Commission containing three requests: narrow the definition of deepfake so that it captures only content carrying a genuine risk of deception; clarify when disclosure must take place to avoid permanent on-screen labelling; and ensure that marking standards are proportionate and technically feasible.

The central argument is summed up in a sentence from the letter that is worth bearing in mind whenever someone proposes systematic labelling: “If warning messages become ubiquitous, people will stop noticing them, and labels will lose their meaning exactly where it matters most.” The risk of indiscriminate labelling is not an aesthetic inconvenience. It is the devaluation of the warning signal itself.

The technical response is already taking shape

In parallel, on 10 June the Commission published a Code of Practice on Transparency of AI-Generated Content. It is voluntary, but it has a very specific effect: organisations that adhere to it can rely on its measures to demonstrate compliance to any market surveillance authority in the European Union, rather than having to prove the adequacy of their own solution on a country-by-country basis.

What is particularly interesting is the architecture assumed by the code, because it recognises that no single technique is sufficient: cryptographically signed content credentials that travel with the file, combined with an invisible watermark embedded in the pixels themselves or in the audio waveform. Two layers. In practice, this means C2PA combined with watermarking such as SynthID, which is precisely the direction the industry has taken on its own. In May, OpenAI became C2PA-compliant as a generator and incorporated Google DeepMind’s SynthID into its images, alongside a public verification tool. On 31 July, it extended this marking to audio and opened verification through an API. The EBU, for its part, has been advocating the adoption of provenance standards such as C2PA since March.

No European regulation imposes C2PA by name. It simply happens to be the technology that currently meets the criteria described by the regulation.

Spain: the law that is not yet in force, and the one that already applies

The Organic Bill on the Proper Use and Governance of Artificial Intelligence, published in the Official Gazette of the Spanish Parliament on 12 June, remains at the amendment stage in Congress. It designates AESIA as the market surveillance authority and classifies breaches of these labelling obligations as serious offences, with fines of up to €7.5 million or 1% of turnover.

One clarification is worth making, because the wrong figure is frequently reported: the €35 million and 7% figures commonly cited in the press apply to prohibited practices, which are a different matter. Incorrect labelling is not the same as using emotion recognition in the workplace.

In any case, waiting for that law would be a miscalculation. The European regulation has applied directly since 2 August, and its own penalties can reach €15 million or 3% of global annual turnover.

What this changes in the workflow

The operational conclusion is straightforward, and somewhat uncomfortable: traceability has to be designed at the output node, not patched into continuity. A generation workflow that discards metadata between steps, or an export that strips out credentials, forces teams to manually reconstruct a chain of custody that the model had already signed. And reconstructing it manually 30 minutes before going on air is not a plan.

This is precisely where regulation stops being a legal-department issue and becomes a pipeline-design issue. Against this new backdrop, the next issue of TM BROADCAST will examine four tactical approaches to integrating generative AI into real-world production workflows, from local node-level control to collaborative cloud-based canvases, with one surprise that is already becoming apparent: the model that scores best in the rankings is not always the one that can legally be deployed in Europe.

Sources

• European Commission. Transparency obligations under Article 50 of the AI Act: FAQ and Guidelines on transparency obligations (20 July 2026).
• European Commission. Code of Practice on Transparency of AI-Generated Content (10 June 2026).
• EBU, ACT, VAUNET, egta, AER, AKTV, APR and Privatsender. Joint letter to the European Commission (17 June 2026).
• Advancing content provenance (19 May and 31 July 2026).
• Spanish Parliament. Organic Bill on the Proper Use and Governance of Artificial Intelligence. Congress of Deputies, Series A, No. 97-1 (12 June 2026).