Deepfakes in adverts, films and social media – when and how should AI-generated content be labelled?

Artificial intelligence has become a permanent fixture in creative production – from generating graphics and voices to modifying actors’ faces on screen. However, as these tools have become more widespread, a real problem has emerged: audiences are increasingly unable to tell what is real and what has been generated by an algorithm. The EU regulation – the AI Act – addresses this by imposing a specific obligation from 2 August 2026: deepfakes must be clearly labelled.

New rules of the game for AI-generated content – the obligation to label deepfakes from 2 August 2026

From 2 August 2026, any company, creative agency or film producer that uses artificial intelligence to create or modify images, audio and video recordings must clearly inform audiences that they are dealing with content generated or manipulated by AI, provided that such content meets the definition of a deepfake. This new obligation stems from Article 50 of the EU Artificial Intelligence Regulation (AI Act) and applies to any entity creating deepfakes as part of its professional activities – whether it is an advertising agency, a media house, a production studio or an e-commerce platform. Breaches are punishable by an administrative fine of up to €15 million or 3 per cent of total global annual turnover for the previous year (whichever is higher).

What is a deepfake – and when does the labelling obligation apply?

The AI Act defines a deepfake broadly: it is an image, audio or video content generated or manipulated by AI that resembles existing persons, objects, places or events and may be falsely passed off as authentic or genuine. Four elements determine whether content is a deepfake: (1) similarity – a high, though not necessarily identical, degree of resemblance to the original; (2) realism – the content must resemble someone or something that exists or could exist; (3) category – people, objects, places or events; (4) potential to mislead – assessed objectively, regardless of the creator’s intention, taking into account audiences with less digital experience (children, older people).

What constitutes a deepfake (examples from the EC Guidelines): an AI-generated image of an actor or celebrity in a promotional advert; an actor’s face made to look younger or older; a synthetic voice of a real person; a product packshot manipulated to look better than in reality; a video of a politician delivering a speech they never actually gave.

What is not a deepfake: clearly fantastical scenes (dragons, flying sphinxes, talking mice); minor technical adjustments to the background, colour, lighting or recording quality, provided they do not affect the perception of the content’s authenticity; synthetic voices of fictional characters in animations and games; a film in which real actors perform against an AI-generated background, but the scene does not purport to be real.

From 2 August 2026, texts generated or manipulated by AI that deal with matters of public interest (e.g. health, safety, politics, armaments, the environment) should also be labelled, unless the text is fact-checked by a human and subject to genuine editorial responsibility (in which case no such labelling obligation arises). Texts on less significant topics, including typically marketing-related ones, are generally not subject to this regime (with the exception of elements of public interest, e.g. health claims, consumer safety, financial services, etc.).

How to label – general principles

The labelling obligation rests with the AI deployer, i.e. the entity that uses the AI system under its own responsibility and decides how and for what purpose it is used. The labelling must be clear, distinguishable and accessible upon the recipient’s first encounter with the AI-generated content. Technical metadata or a watermark alone are not sufficient. In practice, the EU Code of Good Practice and Guidelines propose:

• in video and image content: a visible icon or text label ‘AI’ / ‘SI’, optionally accompanied by the phrase ‘Generated by AI’ or ‘Modified by AI’, placed in a fixed position (e.g. the top right-hand corner);
• in audio content: a voice-over at the beginning (e.g. ‘This recording has been generated/processed by AI’);
• in texts providing information on matters of public interest: a label above the text, next to the headline or in a prominent place on the interface.

If the AI has generated only a fragment of the material, we mark that fragment. If the entire material comes from the AI, we mark the whole of it.

Feature films, TV series and artistic works – a more lenient regime

The AI Act provides for an important exception for artistic, fictional, satirical and creative works. If the nature of the work is obvious to the audience (feature film, animation, satire), it is sufficient to disclose the use of AI in a way that does not hinder the audience’s enjoyment of the work – e.g. in the end credits, the description on a streaming platform, or accompanying materials. In practice, this means that:

• in feature films, it would suffice to indicate in the end credits that selected scenes or characters have been generated or modified by AI, and if actors are listed in the opening credits – to include an appropriate annotation next to the name of any generated or de-aged actor;
• documentaries, due to their informative nature, do not, as a rule, benefit from this relaxation – re-enactments should be labelled directly whilst they are being shown;
• Advertisements – even if creatively produced – do not, as a rule, benefit from the artistic exception (as they serve a commercial purpose); they require a clear label at the point and time of display.

Practical recommendations

The new regulations apply from 2 August 2026. Content published before that date does not require retrospective labelling; however, content generated before that date but published on or after 2 August 2026 must be labelled. What you should do now:

1. Take stock of AI applications in content production and publication processes – video, audio, graphics and text.
2. Implement an internal content classification policy (deepfakes / standard editing / artistic material) with label templates.
3. Document the process of creating AI-generated material: prompts, intermediate versions, final versions.
4. Set out labelling requirements contractually with subcontractors, distribution partners and clients.
5. In borderline cases – once the Commission for the Development and Safety of Artificial Intelligence has been established – request an individual opinion regarding the application of the AI Act (cost: 150 zł, response time: 30 days; up to 60 days for complex cases).

It is important to bear in mind that the final version of the EU guidelines … was published … and has significantly altered the way deepfakes are viewed compared to previously published draft guidelines. The final text is more liberal in many respects and appears to better reflect the purpose and meaning of labelling AI-generated content. Regardless, the guidelines are often imprecise, so practice will only begin to take shape through the first decisions, guidelines or opinions issued by, amongst others, the aforementioned Commission on the Development and Safety of Artificial Intelligence. Therefore, at present, each case should be assessed on its own merits, with a particular focus on the potential for misleading the public.

Sources:

• Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 on the establishment of harmonised rules on artificial intelligence and amending Regulations (EC) No 300/2008, (EU) No 167/2013, (EU) No 168/2013, (EU) 2018/858, (EU) 2018/1139 and (EU) 2019/2144, and Directives 2014/90/EU, (EU) 2016/797 and (EU) 2020/1828 (the Artificial Intelligence Act), OJ (EU) L 2024, p. 1689, dated 12 July 2024
• Code: https://digital-strategy.ec.europa.eu/en/policies/code-practice-ai-generated-content
• Guidelines: https://digital-strategy.ec.europa.eu/en/policies/guidelines-transparency-ai-generated-content