Deepfake / Public-Interest Content Label
A transparency and trust pattern ensuring that deployers publishing deepfake image, audio, or video content, and AI-generated text published to inform the public on matters of public interest, clearly label that content as artificially generated or manipulated. It defines when the obligation applies, what the label must communicate, how to surface it across content types and channels, and how each of the three exemptions works. It is the home of the mandatory human-facing disclosure duty, complementing Article 50(2) machine-readable marking.
The Problem
Deepfakes and unlabeled AI-generated public-interest text erode the basic ability of an audience to tell what is real. Unlike a chatbot disclosure, which protects a single person mid-conversation, this pattern protects a public that may encounter the content stripped of context (reshared, screenshotted, clipped, or quoted well after publication). The harm compounds because deepfake content is built to look authentic, and public-interest text is trusted precisely because it appears to inform rather than persuade. The deeper risk is the erosion of a shared baseline: the public's ability to assume that content presented as authentic actually is.
- A manipulated image, audio clip, or video circulates with no indication it was artificially generated, and is taken at face value.
- AI-generated text on a matter of public interest is published without disclosure, lending it unearned authority.
- A label exists at first publication but is stripped as the content is re-shared, embedded, or clipped.
- The artistic/satirical exception is used to justify no disclosure at all, rather than a less prominent one.
- The editorial-review exemption is claimed on the strength of someone clicking publish, with no real review or named responsible person.
- The label is present but not perceivable: an on-image overlay with no alt text, or a banner conveyed only by colour and position.
Intended Outcome
Deepfake / Public-Interest Content Label is working when an audience can tell synthetic from unaltered without investigating.
- Labeled at consumption, not just at publication: every deepfake carries a label wherever the audience actually meets it.
- Text disclosed before reading: AI-generated public-interest text is disclosed ahead of the substance, not in a byline tooltip.
- Perceivable, not only present: the label reaches assistive technology and is never conveyed by colour or position alone.
- Exemptions applied honestly: the artistic exception reduces prominence rather than removing disclosure; the editorial exemption rests on a real logged review.
- Persistent where possible, honest where not: labels survive controlled redistribution paths and residual risk is documented.
- Consistent across formats: every content type within a piece is covered, not just the most visually obvious one.
Scope & Boundaries
- The legal baseline, two limbs: Article 50(4) deepfake disclosure (image, audio, video) and AI-generated public-interest text disclosure. Both are deployer obligations with different exemptions.
- Article 50(5) delivery: clear, distinguishable, accessible, no later than first exposure. The accessibility limb lands here in full.
- A human-visible labeling mechanism per content type, distinct from any 50(2) machine-readable mark.
- An editorial responsibility record: a logged human review and a named person holding editorial responsibility for the text limb.
- A label application log and three separate exception registers (artistic, editorial, law-enforcement).
- Machine-readable marking: an Article 50(2) mark is invisible to a human audience. The two obligations are complementary, not interchangeable.
- Guaranteed persistence: labels do not reliably survive screenshots, re-uploads, or clips, so residual risk is documented rather than assumed away.
- A publish click as editorial control: the exemption needs a real review step and an identifiable responsible person.
- Treating 'public interest' as a fixed category: it is a judgement the deployer must define and document.
How it Works
At the point of publication, the content pipeline determines whether an asset is a deepfake or AI-generated public-interest text, resolves which exemption (if any) applies, and attaches a human-visible label appropriate to the content type and channel before the content reaches an audience.
Identify all content types and channels in scope: image/audio/video generation or editing, AI-assisted or fully AI-generated article pipelines, and every distribution channel (web, app, social embeds, syndication, email).
Resolve the limb and the exemption before the label: check the editorial-review exemption for the text limb, the artistic exemption (reduced prominence) for the deepfake limb, and law-enforcement for either.
Define the label trigger and format per content type: trigger point, rendered format, copy, persistence through export/syndication, and assistive-technology exposure.
Log label application and exemptions: content ID, type, obligation limb, format, exemption category and basis, editorial reviewer, publication date.
Document exemptions by category: artistic, editorial (named responsible person), and law-enforcement.
System Interaction
Placement & Triggers
- Synthetic media (image, audio, video) depicting real or realistic people, events, or statements.
- AI-generated articles, summaries, or explainers published to inform the public on current events, politics, or civic matters.
- Content likely to be redistributed outside the platform or context where it was originally labeled.
- Election periods, public-health communications, or other moments where public-interest content carries outsized influence.
Supported Architectures
Experience Surface
The label should be legible without overwhelming the content. The goal is an informed audience, not a defaced asset. Prominence varies with the exemption; presence does not. A satirical deepfake still says it is AI-generated, still says it in alt text, and still says what it is rather than gesturing at AI generally. Because labels attached at the platform level do not reliably survive screenshots, re-uploads, or clips, that residual risk is documented rather than presented as solved.
Design Assets
A Disclosure Kit organized by content type, across the deepfake limb and the public-interest text limb, plus a cross-cutting reduced-prominence variant for the artistic exemption. Prominence is the only thing the exemptions change; presence, accessibility, and specificity do not vary.
Deepfake limb (image, audio, video)
Image Disclosure Label
Identify a generated or manipulated still image at the moment it is seen.
Placement · On the asset itself, positioned to survive standard cropping and resizing. Not a caption, which detaches on re-upload.
- Short statement (“AI-generated” / “Digitally altered”), legible at thumbnail and preview scale.
Accessibility · The statement must also exist in alt text and caption. An on-image overlay alone fails the 50(5) accessibility limb.
Audio Disclosure Statement
Carry the disclosure into a medium with no persistent visual surface.
Placement · At the start of the clip, before the substantive content, so a listener who stops early has still been informed.
- Spoken statement (“This audio was generated or altered using AI”), plus a leading textual disclosure wherever a visual surface exists.
Accessibility · Include the statement in transcripts and captions. A spoken-only disclosure is not accessible.
Video Disclosure Overlay
Keep the disclosure attached to moving content that is routinely clipped and re-encoded.
Placement · Burned into the frame rather than a caption track, positioned to survive resolution variants, present for the duration.
- Persistent on-frame statement, and the same statement in the caption track and description.
Accessibility · Present in captions, description, and any accompanying text. A burned-in overlay is unreadable to assistive tech.
Public-interest text limb
Text Disclosure Banner
Tell a reader the text is AI-generated before they invest in reading it.
Placement · Above or immediately alongside the content, ahead of the body. Not a byline tooltip, not below the fold, not a footer.
- Plain statement (“This content was generated using AI”); where later edited, an accurate description rather than a blanket claim.
Accessibility · Real text in the reading order, ahead of the article body. Never an image of text, never distinguished by colour alone.
Cross-cutting
Reduced-Prominence Variant
Satisfy the artistic/creative/satirical/fictional exemption without eliminating disclosure.
Placement · Applies to any component above (deepfake limb) where the exemption is documented.
- A smaller or less intrusive treatment of the same statement.
- Still present, still legible at the size the audience actually views the content, still in alt text or captions.
- What it is not: absence. The exemption narrows prominence, it does not remove the obligation.
When to use
Risks & Guidelines
- The vanishing label: a disclosure that does not survive cropping, screenshotting, or re-upload, giving false confidence.
- The publish-click exemption: claiming editorial review because a person pressed publish. This is the exemption least likely to survive audit.
- Exemption cross-contamination: applying the editorial exemption to deepfake media, or stretching the artistic exemption to public-interest text.
- Exception overreach: stretching 'evidently satirical' to cover content a general audience could take at face value.
- Overlay without alt text, or inversely, disclosure only in metadata most viewers never see.
- Inconsistent format coverage, and confusing an embedded Article 50(2) mark with the visible labeling this pattern requires.
- Design for your actual redistribution paths, and document what cannot be solved for.
- Make the editorial exemption evidenceable through a gate that records the review and the responsible person.
- Check the limb before the exemption. The right exemption on the wrong limb is a confident, invalid claim.
- Put the statement in text as well as pixels, every time (alt text, captions, transcripts, description).
- Apply the artistic exemption narrowly, and cover every content type in a piece, not just the obvious one.
- Assign a named owner for labeling standards, and log every exemption with a reviewable basis.
How to Measure Success
How will we know this pattern is strengthening trust?
North Star
Label Coverage Rate: the percentage of published in-scope assets with a qualifying label at publication (clearly stating AI generation or manipulation, positioned before or alongside the content, matched to the content type, and perceivable by assistive technology), measured against the published-content inventory and reported per content type. Target: 100% at publication, 0 unlabeled in-scope assets in post-publication audit.
Editorial Exemption Integrity
the share of text-limb exemption claims backed by a logged review and a named responsible person.
Label Persistence Rate
whether labels survive real redistribution, sampled rather than assumed, reported as residual risk, not a target.
Accessibility Conformance
the share of labels present in text as well as visually.
Exemption Health
whether every claimed exemption is documented, correctly categorised against its limb, and reviewed.
Label-to-Complaint Correlation
whether coverage gaps track with audience reports of being misled.