AI deepfake & AI influencer – transparency labelling according to EU AI Act Article 50 | Word/PDF/Excel

Do you publish AI influencers, realistic AI videos, synthetic voices, or AI-manipulated images? From 2 August 2026, transparency rules under Article 50 of the EU AI Act (EU) 2024/1689 apply to certain AI systems and content. This specialised document package helps companies, marketing agencies, content creators, and e-commerce businesses assess whether specific content is a deepfake, draft the correct disclosure, make publishing decisions, and maintain an audit trail.

Short answer: According to Article 3.60, not every AI image is a deepfake. It refers to AI-generated or AI-manipulated image, audio, or video that resembles existing persons, objects, places, entities, or events and could be falsely perceived as authentic. A professional user publishing deepfake content may need to provide a clear disclosure under Articles 50.4–50.5, at the latest upon first exposure. The package provides auditable documents, not automatic legal classification.

Why is a special template for AI deepfakes needed in 2026?

General AI policies often describe what an organisation may do with AI. Article 50, however, requires a concrete assessment for every relevant publication: what has been generated, who is responsible, could the material be perceived as authentic, what disclosure is required, and has the audience actually seen it? These are the questions the product’s five working documents are designed to address.

Three examples: (1) A company publishes a film featuring a realistic AI reconstruction of a real event. Assess whether it is a deepfake and how the disclosure should be visible upon first exposure. (2) An advertising agency produces a synthetic influencer resembling an existing person. Document actual resemblance, rights, advertising identification, and labelling decisions. (3) A clearly fictional commercial contains AI-generated special effects. Assess whether it is a deepfake within the meaning of the law and whether the artistic provision allows for a limited form of disclosure. Do not make a general automatic assessment.

Complete File List

Excel: 9 tabs for control and traceability

START · Content · Assessment · Labelling · Publications · Checks · Deviations · Overview · READ ME. The company can maintain its own register of AI productions with a version ID, the AI solution used, who is publishing, the deepfake assessment, and selected labelling text. Each actual publishing channel gets a separate row. Mobile, accessibility, and redistribution tests can be reviewed and documented with evidence references.

The Excel overview counts registered content, assessments, labels, publications, completed checks, open deviations, and blocked decisions. The formulas are based on your own data and never provide automatic legal approval. Date columns in the check and deviation tabs count calendar days between two real dates. A test with 3 and 7 November gave four days. The delivered customer file is empty of test data and contains no macros or external data connections.

How Article 50 Works – Seven Key Differences

1. Applies from 2 August 2026. The European Commission clarifies that the transparency obligations of Article 50 began to apply on this date. For certain AI systems already on the market, there is a limited transition period for the provider's machine-readable labelling under Article 50.2 until 2 December 2026. This does not mean that every publisher can postpone their deepfake disclosure under Article 50.4. The Commission's FAQ.

2. Not every AI-generated image is a deepfake. Article 3.60 requires a specific risk of confusion with authentic material resembling an existing person, place, event, or other specified subject. The actual context and audience need to be assessed. The mere use of generative AI is not sufficient for this classification.

3. Companies and individuals may have different roles. Article 3.4 defines users (deployers) and exempts personal, non-professional use. However, a company, agency, or creator that professionally uses an AI system may be a deployer. Who is the provider, deployer, and responsible publisher should not be guessed.

4. Machine-readable labelling is not the same as a visible disclosure. Providers of certain generative AI systems are subject to a technical labelling obligation under Article 50.2. Anyone who professionally uses the system to publish a deepfake may have a separate obligation to clearly disclose AI generation/manipulation under 50.4. Hidden metadata from the provider is not always sufficient for the publisher.

5. First exposure and accessibility. Under Article 50.5, relevant information must be clear and distinguishable at the latest when a person first encounters the content. Placement in the image/video, captions, audio, and mobile interface therefore needs to be checked. Redistribution and previewing can change how the label is visible.

6. Creative or fictional works are not automatically exempt from all disclosure. Article 50.4 contains a special rule for obviously artistic, creative, satirical, fictional, or similar works. It can limit the form of disclosure so that it does not disrupt the experience, but is not a general waiver of transparency. The assessment must be documented.

7. AI-generated text has a different rule. Article 50.4 also has a special provision for AI-generated text published to inform about matters of general interest, with an exception when relevant human review/editorial control and editorial responsibility are present. This text exception cannot be automatically used for realistic deepfake video.

Ready-made but editable examples of AI disclosure

The examples are intended as a starting point and should only be selected when they correctly describe the actual content.

• AI-generated realistic image: “AI-generated image. The subject is not an authentic photograph.”
• Manipulated photo: “AI-manipulated image. Parts of the photograph have been changed with AI.”
• Reconstructed video: “This video material is AI-generated and does not show an authentic recording of the event.”
• Synthetic voice: “The voice in the clip is AI-generated and not an authentic recording.”

Choose words that truly fit; distinguish between fully generated and partially manipulated, and avoid claims about people's identities without verification. EU AI labelling icons can be used voluntarily but are not mandatory and do not in themselves constitute a presumption of legal compliance. The icons are not included as copied third-party files in this package.

Workflow – Six Steps Before Publication

1. Identify the AI production. Collect information about the actual file version, AI system, creator's role, and planned publishing channel.
2. Perform the deepfake assessment. Complete document 01 with the criteria under Articles 3.60 and 50.4, including whether the content is professional and realistic.
3. Determine the disclosure. Use document 02 and select visible or audible labelling that correctly describes the content. Test the first exposure.
4. Check advertising and rights. For an AI influencer or advertisement, complete document 03 and also assess advertising identification, personality rights, and GDPR separately.
5. Assess exceptions and make a decision. Use document 04 when clear fiction/artistic nature is relevant and document 05 for the final publishing decision. Block uncertain cases.
6. Save verifiable evidence. Add actual screenshots, channel URLs, and check results to Excel. If the labelling is deficient, document the action and publish a corrected version if necessary.

Frequently Asked Questions about AI Influencers, Deepfakes, and Article 50

Must all AI-generated ads be labelled as deepfakes?

No, not solely because AI was used. The deepfake assessment requires an assessment against Article 3.60. Other rules about advertising identification, platform information, or misleading marketing may still apply.

Do the rules apply to AI influencers on Instagram, TikTok, and YouTube?

They may be relevant when a professional actor publishes AI material that meets the deepfake definition. Platforms may have their own additional labelling requirements. It is incorrect to claim that all virtual influencers automatically meet Article 3.60.

Is labelling needed if the video already has hidden AI metadata?

For a relevant deepfake, the deployer rule requires a clear disclosure for people. Machine-readable provider labelling under Article 50.2 is not necessarily sufficient for Articles 50.4–50.5.

Is the EU's official AI icon sufficient?

Not automatically. The icon is voluntary support, and the disclosure must still be appropriate, clear, and in the right context at first exposure. The Commission states that the icon itself does not prove compliance.

Is clearly fictional advertising exempt?

An obviously artistic, creative, satirical, or fictional presentation may allow for a limited form of disclosure. A general exemption from informing does not automatically follow. Document 04 helps report the reasons.

Must AI-generated text on a website be labelled?

Article 50.4 has a special rule for AI-generated text about matters of general interest and certain exceptions for human review and editorial responsibility. The rule cannot be generalised to every product text, or directly transferred to deepfake images and video.

What responsibility does the agency have compared to the one who developed the AI system?

Provider and deployer are different roles under the AI Act. An agency may be a deployer when it uses an AI system under its own or the customer's responsibility, while the one who develops and provides a system may be the provider. This must be assessed from the actual relationships.

Is this an officially certified AI Act template?

No. The package contains auditable working documents and legal references. It does not provide an official AI certificate, automatic approval, or guarantee that GDPR, the Marketing Act, or the AI Act is complied with.

Are the letters in English and the PDFs fillable?

The five working documents and one guide are in Swedish. A separate English User Guide describes how they are used. The PDFs are searchable/printable but are not interactive fillable forms; the Word files can be edited.

Related Products for Broader AI Work

For overall company instructions, there is AI Policy for Companies. For governing AI use across the entire organisation, there is the AI Governance Package. For assignments with external content creators, the UGC Creator Agreement may also be relevant. These fulfil different needs than a separate deepfake assessment and documented labelling for every actual publication.

Official Legal Sources and Audit Status

EU AI Act – Article 50 · EU AI Act – Article 3 and definition of deepfake · European Commission Guidelines from July 2026 · European Commission Article 50 FAQ · EU Voluntary Labelling Icons.

Version 1.0 · Jurisdiction EU/Sweden · AI-assisted legal source check 2026-11-07. No external independent legal review, official certification, or guarantee of compliance has been performed. Check the applicable version of EU law, agreements, and publishing context before use. Delivery is provided as a digital ZIP only after potential later publication and separate connection to a digital delivery app; the product is created as a DRAFT.