What kind of content is it?
Image, audio or video follows the deepfake test. Pure text follows a separate public-interest test.
Choose the right path01 · Start here
Three questions. In this order.
The company using the AI system under its own responsibility is normally accountable—not the individual employee pressing the button.
Does the full test pass?
Every condition in the relevant test must be met. Minor technical editing alone is not enough.
All conditions, not just oneWill people notice the label?
Make it human-readable, visible or audible at first exposure. Metadata alone does not count.
Clear at first exposure02 · Images, audio & video
The four-part deepfake test
Label the content when all four statements are true. Judge realism from the perspective of the actual likely audience—including people with lower digital literacy.
01 · Made with AIGenerated or manipulated using AI?
02 · Strong resemblanceSubstantially resembles something real or realistic?
03 · Real-world subjectA person, object, place, entity or event?
04 · Appears authenticCould wrongly appear genuine or truthful?
All four = label it. A photorealistic person can qualify even if that person is entirely fictional. Clearly impossible fantasy—such as a dragon or an elephant driving a car—normally fails the authenticity test.
Rule of thumb
- The more realistic the person or environment, the clearer the duty to label.
- Adding a believable person, location or scene can turn an ordinary asset into a deepfake.
- Changing meaning, style or impression is more likely to be substantive.
Still check other laws
- Marketing and hidden advertising
- Copyright and trademarks
- GDPR and identifiable people
- Image, voice and personality rights
Usually no label Standard editing
- Colour, light, sharpness or noise adjustment
- Crop, rotate or upscale
- Compression or format conversion
- Red-eye, dust or tiny background fixes
Label Substantive change
- Add or remove objects that change the impression
- Alter body shape or skin colour
- Insert realistic people, places or environments
- Create a new realistic scene or animate a still image
03 · AI-generated text
Public-interest text has its own test
Pure text is not assessed as a deepfake. It requires a label only when all three conditions below are met—and the editorial-review exception does not apply.
01 · PublishedAvailable to a large, undefined audience?
02 · InformativeIntended to convey facts, knowledge or views?
03 · Public interestAbout a matter relevant to society?
All three = potentially label it. Public-interest topics include government, public services, fundamental rights, health, environment, consumer safety, politics, economics, science and culture.
The editorial-review exception
No label is required under this text rule when both are true:
- A qualified person performs a real, deliberate content review—not just spelling or grammar.
- An identifiable person or organisation takes editorial responsibility for publication.
Usually outside this text rule
- Internal notes, meeting summaries and unpublished drafts
- Private chatbot summaries visible only to the user
- Fiction and pure entertainment
- Ordinary ad copy without public-interest claims
04 · How to label
Visible or audible. From the start.
Place the disclosure on or immediately next to the content so it is clear and distinguishable at first exposure. The disclosure may be embedded in the asset or provided through an equally clear interface label. Metadata or a hidden image layer alone is not sufficient.
Image & webshop
Put a clear label on or immediately beside the image. Protect it from cropping and mobile scaling.
Never metadata-onlyVideo
Show from the start. Keep it throughout clips of 15 seconds or less; repeat in longer videos when AI content appears.
15 seconds is guidance, not lawLivestream
Keep the label visible throughout, or disclose at the start and repeat regularly for people joining later.
Account for late arrivalsAudio
Use a short audible disclaimer at the beginning. Repeat it in longer podcasts or radio formats.
Add a visual label when a screen is usedSocial media
Use a high-contrast on-asset label plus a clear note in the post text, before any “read more” cut.
Do not bury the disclosureDigital banner
Keep the label readable at the actual ad size and fixed in place through animation and loops.
Persistent in small formatsEU placement checklist
- Article 50 requires a clear and distinguishable disclosure no later than first exposure.
- The Act does not require fixed wording, size, position or a label baked into the asset.
- The voluntary Code of Practice recommends direct embedding by default but permits an equivalent interface overlay. EU icon guidance ↗
- Keep the label free from banners, controls and other overlays.
- Make sure it survives cropping and scaling.
Make the disclosure accessible
- Pair the icon with a plain-language label where possible; testing shows this improves recognition.
- Avoid jargon and unclear abbreviations other than “AI”.
- Add alt text or an ARIA label stating that the content is AI-generated or manipulated.
- Keep temporary disclosures visible long enough to be read and understood.
- For interactive labels, signal that more information is available and make the second layer accessible.
GENERATED WITH AIExample placement: high contrast, on the asset, visible from first exposure.
Match the wording to the change
Say what actually happened.
Generated with AIParts of this image were generated with AIBackground manipulated with AIEU icons are optional. Using an icon alone does not automatically prove compliance.
EU AI icon set (PNG) ↗EU AI icon set (SVG) ↗05 · Common examples
What the guidance says in practice
Follow the same dog from an untouched photograph through increasingly creative AI use. The decisive question is not simply whether AI was used, but whether the result could wrongly appear authentic.
Everyday AI photo editing
Using Firefly or a similar tool to remove distracting people, cars, hands, shadows or plants—or to extend the edges of a photo—does not normally require a label. If the subject, setting and meaning remain essentially the same, treat it like ordinary Photoshop work.
Ask before publishing if the edit:
- Removes something important to the story.
- Changes how a person or product appears.
- Creates a believable new place, event or situation.
Quick check: Does the edited image still tell the same truthful story as the original? If yes, usually no label.
The office dog
Not a deepfake · No labelOriginal photograph
No AI generation or manipulation. This is the reference image.
Not a deepfake · No labelMinor AI enhancement
The subject, setting and meaning remain unchanged. The same rule normally covers incidental cleanup or modest background expansion.
Usually no labelOriginal dog, AI office
The original photograph of the dog is preserved and only the background changes. Reassess if the context claims or implies that the dog was genuinely photographed there.
Not a deepfake · No mandatory labelFictional action poster
Its stylised, impossible scene does not falsely appear authentic. Voluntary AI disclosure is still good practice.
EU reference: Article 3(60) requires that a deepfake “would falsely appear to a person to be authentic or truthful.” The final Article 50 Guidelines, paragraph 116, explain that removing passers-by, extending existing backgrounds and replacing backgrounds for clearly aesthetic purposes are likely to have only a minor impact, depending on context.
Remember: “Not a deepfake” does not clear other legal questions. Copyright, trademarks, advertising rules and personality rights must still be assessed separately.
The coffee mug
Cards A–C preserve the real packshot. Card D changes the mug itself as well as its setting.
Not a deepfake · No labelPhoto, AI-enhanced
Only light, colour, noise and sharpness change. The real mug and setting remain authentic.
Not a deepfake · No labelReal packshot, neutral AI background
The real mug is preserved accurately. Only the neutral background is synthetic.
Not a deepfake · No labelReal packshot, graphic AI setting
The real mug remains accurate and the obviously designed background does not pretend to document a real scene.
Deepfake · LabelAI-manipulated mug and location
AI changes the product as well as the setting, creating an authentic-looking advertisement that no longer represents the real mug accurately.
Advertising rule: In advertising, a real product shown against an AI-generated background is acceptable provided the advertisement does not mislead about the product’s actual representation, its characteristics or use, whereas an AI-generated image that makes the product appear different from, or better than, reality is a deepfake. Source: Bird & Bird ↗. See also the final Article 50 Guidelines, p. 36.
Cards B–C preserve the real packshot accurately. Card D does not, so it requires a label.
Hands and people
A realistic body part can be enough. The person does not need to exist or copy a named individual.
Deepfake · LabelAI hand and environment
A realistic hand and office create the impression of an authentic lifestyle photograph.
Deepfake · LabelAI hand and face
Adding a photorealistic face strengthens the false impression that a real person was photographed.
Deepfake · LabelEverything AI, including model
Use “GENERATED WITH AI” here. A fictional but lifelike person can qualify because she looks like someone who could exist.
All people in these examples are fictional adults and must not resemble an identifiable real or public person.
AI-enhanced product photo
Only light, colour and noise are adjusted. The real product and overall authenticity remain unchanged.
Real packshot on an AI background
The real product is preserved accurately and only the background is generated.
AI-manipulated product and location
The realistic advertisement changes the product itself, so it no longer represents the real item accurately.
Realistic AI model or hand
A photorealistic person, face or hand can qualify even when no real named person is being copied.
Still photo animated into realistic video
Movement such as pouring wine creates an authentic-looking event that did not happen.
Editor-reviewed AI-assisted article
A qualified human genuinely reviews it and an identifiable publisher accepts editorial responsibility, so the separate text-labelling exception normally applies.
06 · Client delivery
Before you deliver to the client
The party placing the asset normally implements the disclosure. Legal responsibility follows the deployer—the organisation under whose authority and responsibility the AI system was used—and is not automatically transferred when an asset is handed over. EU Article 50 guidance ↗
When we place it · Supplier
We normally implement the required disclosure when we add the asset to the final website, print file, video or other publication.
When they place it · Customer
The customer normally implements the required disclosure when they add or publish the asset themselves. If we are the deployer, we must still ensure that they receive and apply the correct disclosure.
Every handoff: Identify in writing whether AI was used, what was generated or manipulated, what disclosure is required and who will implement it. Deliver a labelled version when the publisher cannot guarantee an effective disclosure in the final context.