Deepfake Disclosure Rules for Websites and Media Teams

Deepfakes — AI-generated or manipulated media that falsely appears authentic — are a specific focus of Article 50(4) of the EU AI Act. If your website publishes AI-generated images, audio, or video that resembles real people, places, or events, this guide explains what the law requires. For the broader context,
see our Article 50 explainer and EU AI Act compliance guide.

What Is a Deepfake Under the EU AI Act?

Article 3(60) of the AI Act defines a deepfake as AI-generated or manipulated image, audio, or video content that meets three cumulative criteria (European Commission):

1. Resemblance — The content resembles existing persons, objects, places, entities, or events. There must be a high level of similarity between the deepfake and the simulated subject.
2. Existing subject — The simulated person, object, place, entity, or event resembles something that exists, can plausibly exist, or could have plausibly existed in reality.
3. False appearance of authenticity — The content has the capacity to potentially deceive or mislead a person regarding its authenticity or truthfulness.

All three criteria must be met. If any one is missing, the content is not a deepfake under the AI Act definition.

What Is NOT a Deepfake?

The guidelines identify content that is unlikely to be a deepfake:

• Background scenes — AI-generated backgrounds that don’t resemble specific real places
• Special effects — Standard movie production special effects
• Technical pre/post-processing — Standard editing and enhancement
• Cartoons and illustrations — Unless they realistically resemble real persons or events

For example, a cartoon dragon is not a deepfake — it doesn’t resemble an existing entity. A photorealistic AI-generated video of a real politician saying something they never said is a deepfake.

When Must You Disclose?

Deployers must disclose deepfake content to viewers no later than the person’s first exposure to the content (European Commission).

This means the disclosure must appear before or simultaneously with the viewer seeing the deepfake. You cannot delay disclosure until later in the content or require the viewer to click through to a separate page.

How Must Disclosure Be Made?

Disclosure must be:

• Clear — Unambiguous that the content is AI-generated or manipulated
• Distinguishable — Visibly separate from the main content
• Understandable — Comprehensible to the average person
• Perceivable — Detectable without requiring specific technical tools or dedicated actions by the viewer

Suitable forms include:

• Visible labels — Text labels, badges, or watermarks visible on the content
• Audible labels — For audio or video content, a spoken disclosure

Critically, deployers cannot rely solely on the machine-readable marking embedded by the provider under Article 50(2). The provider’s invisible watermark is not sufficient — you must also provide a visible, human-perceivable label (European Commission).

The Artistic Works Exception

For deepfakes that form part of evidently artistic, creative, satirical, or fictional works (or analogous works), the transparency obligation is lighter. Disclosure must be made in an appropriate manner that does not hamper:

• Display of the work
• Enjoyment of the work

This means you still need to disclose, but the disclosure should be proportionate — it shouldn’t ruin the artistic experience. A small credit-line disclosure may be sufficient for a satirical video, whereas a photorealistic deepfake in a news article would need prominent labeling.

Factors for Assessing False Authenticity

The guidelines allow stakeholders to consider several factors when assessing whether content falsely appears authentic:

• The level of resemblance to the real subject
• The potential substantive message of the content
• The intended deployment context
• The foreseeable deployment context
• The intended audience composition
• The reasonably foreseeable audience composition
• The expectations of the audience

If the intended audience in a particular context does not expect the content to be authentic or truthful, it may not falsely appear authentic. For example, an audience watching a science fiction film doesn’t expect the content to be real, so AI-generated elements within it may not need prominent disclosure (the lighter artistic works exception applies).

Practical Implementation for WordPress

For AI-generated images that resemble real people or places

1. Apply a visible label — Badge, overlay, or caption stating “AI-generated” or “This image has been manipulated”
2. Preserve machine-readable marks — Don’t strip provider-side watermarks from the image file
3. Position the label prominently — Near the image, not buried in a footer
4. Use alt text — Include AI-generation status in the image’s alt text for accessibility

For AI-generated video or audio

1. Add a visible watermark — Burn a “AI-generated” watermark into the video
2. Add an audible disclosure — For audio, include a spoken disclosure at the beginning
3. Add a description — State AI generation in the video description and metadata
4. Label before exposure — The disclosure must appear before or at the moment of first viewing

For AI-generated text that could be mistaken for a real person’s words

If you publish AI-generated text that quotes or paraphrases a real person in a way that could deceive readers, this may qualify as a deepfake (if it resembles the person and falsely appears authentic). Apply clear labels.

Tools for Deepfake Disclosure

• MegaGovern — Provides visible badges, floating labels, and EXIF & C2PA image scanning (Pro tier) to verify provenance. Explore features
• Legalithm — Offers invisible image watermarking that survives re-encoding, plus visible badges in four styles. WordPress.com
• AIM Transparency — Free visible badges, IPTC/XMP provenance embedding, and schema.org JSON-LD. WordPress.org
• Clarivo — Media library scanning with overlay badges and captions. WordPress.org

For a full comparison, see our best AI disclosure plugins guide. For step-by-step labeling instructions, see how to label AI-generated content.

Key Takeaways

1. A deepfake requires three criteria: resemblance to an existing subject, plausibility, and false appearance of authenticity — all three must be met
2. Disclosure must happen no later than the viewer’s first exposure to the content
3. Disclosure must be clear, distinguishable, understandable, and perceivable — visible or audible
4. You cannot rely solely on machine-readable marks — a visible label is also required
5. Artistic, creative, satirical, and fictional works get a lighter-touch disclosure duty
6. Standard editing, special effects, and background scenes are generally not deepfakes

Disclaimer: This article is informational, not legal advice. Deepfake definitions and disclosure requirements involve nuanced legal interpretation. Consult the official European Commission guidance and a qualified legal professional for your specific situation.