Direct answer: Since 2 August 2026, Article 50 of the EU AI Act has required AI companies to mark synthetic content in a machine-readable way, and anyone publishing a deepfake or AI-written public-interest text in the EU to label it clearly at first sight. The only grace period left ends on 2 December 2026.
EU AI Act labelling is the part of Europe’s AI law that most ordinary publishers, creators and businesses will actually feel. It is not about banning AI images or videos. It is about making sure that when a person in the EU sees a realistic AI-made photo, hears an AI-made voice, or reads AI-written text about a public issue, they are told so, clearly, and before they take the content at face value.
This guide explains the EU AI Act labelling rules in plain English: who has to label what, the difference between the invisible machine marking and the visible label you actually see, the free EU icons, the deadlines, and the exceptions. It is general information, not legal advice, for a specific case, speak to a qualified EU lawyer.
Provider or deployer? A provider is the company that builds an AI system and puts it on the market under its own name, the tool maker. A deployer is the person or organisation that uses an AI system under their own professional responsibility, the publisher, creator or business that puts the output in front of people. Article 50 gives each of them different duties, and knowing which one you are is the first step.
What is Article 50 of the EU AI Act?
Article 50 of Regulation (EU) 2024/1689, the EU AI Act, sets out the law’s transparency obligations. It applies even when an AI system is not classed as high-risk, which is why it reaches far more people than the Act’s better-known high-risk rules. The European Commission’s own guidelines on Article 50, published on 20 July 2026 as C(2026) 5054, explain how the duties should be applied in practice. The guidelines are non-binding (the Court of Justice of the European Union has the final word on interpretation) but they are the best official reading available, and market surveillance authorities are expected to follow them.
Article 50 contains four duties. Two fall on providers and two on deployers, and they work in pairs: one invisible, technical duty and one visible, human-facing duty on each side.
The first provider duty, in Article 50(1), is that AI systems designed to interact directly with people (chatbots, voice agents, virtual assistants) must be built so that the person knows they are talking to an AI system, unless that is obvious from the circumstances. The second, in Article 50(2), is that providers of systems that generate synthetic audio, image, video or text must ensure their outputs are marked in a machine-readable format and detectable as artificially generated or manipulated, using technical solutions that are effective, interoperable, robust and reliable as far as technically feasible.
The deployer duties are the ones publishers meet most often. Article 50(3) requires deployers to inform people exposed to an emotion-recognition or biometric-categorisation system. Article 50(4) requires deployers to disclose deepfakes (AI-generated or manipulated image, audio or video that resembles real people, objects, places, entities or events and would falsely appear to a person to be authentic) and to disclose AI-generated or manipulated text published to inform the public on matters of public interest, unless that text went through human review or editorial control and someone holds editorial responsibility for it.
Article 50(5) then sets the standard for every disclosure: it must be provided in a clear and distinguishable manner, at the latest at the time of the person’s first interaction or first exposure. A note buried in terms and conditions, a footnote, or a label that only appears after the content has been consumed does not meet that standard, according to the Commission’s guidelines.
Who has to label AI content, the provider or the deployer?
Most confusion about EU AI Act labelling comes from mixing up the two roles. The table below is this article’s ranking asset: it maps who you are, what content you are dealing with, what Article 50 requires, and when the duty applies.
| Who you are | Content type | What Article 50 requires | Deadline |
|---|---|---|---|
| Provider (tool maker) | AI system that talks to people | Design it so users know they are interacting with AI (Art. 50(1)) | 2 August 2026 |
| Provider (tool maker) | Synthetic audio, image, video or text output | Mark outputs in a machine-readable format, detectable as AI-generated (Art. 50(2)) | 2 August 2026; 2 December 2026 for systems already on the market before 2 August 2026 |
| Deployer (publisher/user) | Emotion recognition or biometric categorisation | Inform the people exposed to the system (Art. 50(3)) | 2 August 2026 |
| Deployer (publisher/user) | Deepfake image, audio or video | Disclose clearly that the content is artificially generated or manipulated (Art. 50(4)) | 2 August 2026 |
| Deployer (publisher/user) | AI-generated text on a matter of public interest | Disclose the AI origin, unless human-reviewed with editorial responsibility (Art. 50(4)) | 2 August 2026 |
The split matters in practice. If you run a website and use an AI image generator to create a hero image of a real-looking person, you are the deployer: the labelling duty is yours, and you cannot pass it to the tool maker by saying the generator did not watermark the file. If you build the image generator, you are the provider: the machine-readable marking duty is yours, and it applies even if every one of your users labels their own posts perfectly.
The duties also reach beyond the EU’s borders. The Commission’s guidance confirms that the obligations apply to providers and deployers outside the EU whose AI output is used in the EU. A US company publishing AI-generated content to an EU audience is in scope; a UK publisher is too, because the test is where the content is used, not where the company sits.
One more distinction that trips people up: the machine-readable marking in Article 50(2) is invisible to readers. It lives in metadata, watermarks or content credentials embedded in the file. The visible label in Article 50(4) is what a person actually sees: a caption, an on-image note, or an icon. A file can carry perfect machine marking and still breach the law if nobody labels it visibly, and a visible label does not excuse a provider from marking the file. The two duties are separate, and both must be met.
What counts as a deepfake under Article 50?
A deepfake, under the AI Act, is AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful. Three questions decide whether something you publish is one.
First, was the content generated or manipulated by AI? A photograph taken on a camera and lightly colour-corrected is not a deepfake. A face swapped with AI, a voice cloned with AI, or a scene generated wholesale by a video model is.
Second, does it resemble something real, a real person, place, object or event? A cartoon dragon or an obviously fantastical illustration is not a deepfake, because no one would mistake it for a real recording. A realistic video of a politician saying something they never said is.
Third, could a viewer perceive it as a real recording? The decisive factor is the potential to deceive, not the intent to deceive. An AI image of a product that makes the product look better than it really is can count as a deepfake under the Commission’s guidelines, even if the seller only meant to make the listing look attractive.
If the answer to all three questions is yes, the content must be labelled. If any answer is no, the Article 50(4) duty usually does not apply, but the content may still be caught by other rules, such as advertising standards or the platform’s own policies, so “no label required by Article 50” is not the same as “safe to publish without any disclosure”.
It is also worth noting what Article 50 does not cover. It does not require labelling of every image an AI has touched. AI-assisted editing that does not substantially alter the input (for example, automatic noise reduction) is outside the marking duty, according to the Commission’s guidelines. And a purely private use, such as generating an image for yourself and never publishing it, triggers no deployer duty at all, because no member of the public is exposed to it.
What are the EU AI icons and how do you use them?
The European Commission has published a free set of EU icons that deployers can use to label AI-generated content. There are three of them, and each has a specific meaning. The icons are part of Section 2 of the Code of Practice on marking and labelling AI-generated content, which the Commission published on 10 June 2026.
| Icon | When to use it | Example |
|---|---|---|
| Basic icon | AI was involved in creating deepfake content or published text, or when you add your own text label or an interactive second layer | A deepfake video with the text label “voices generated with” followed by the basic icon |
| Fully AI-Generated | The entire content is AI-generated, with no human-created elements and no human editorial control beyond prompting | A fully AI-generated video of a fictional event, or an AI-composed piece of music |
| Partially AI-Modified | Pre-existing, human-made content was partly modified with AI, turning it into a deepfake or into public-interest text | A real photograph of an empty flat, furnished with AI; a real face swapped for a politician’s face |
Each icon comes in four variations (black, white, black at 50% transparency and white at 50% transparency) and can be downloaded as SVG or PNG files from the Commission’s EU icons for labelling AI-generated content page. They are free to use without attribution to the Commission or the AI Office, and the Commission’s user-testing found that the icons worked better when the basic icon appeared alongside a short text label, such as “AI-generated”.
Using the icons is optional. The labelling duty itself is not. The Commission states plainly that using an icon does not by itself establish legal compliance, you remain responsible for making sure the disclosure actually meets Article 50. You may design your own label instead, as long as it is clear, distinguishable and visible at first exposure.
Placement matters as much as wording. The Commission’s placement rules require the label to be clearly perceivable at the latest at first exposure, with no intervening overlay hiding it, and directly embedded in the content (not tucked in a footer) so that it stays visible when the content is reshared or downloaded. For images, a high-contrast in-image note or a clearly visible caption meets the requirement; for video, a notice at the start and in the description; for audio, a note in the title and episode description; for text, a notice at the top, near the headline. A label hidden in alt text, a website footer, or a terms page does not count, and metadata alone does not satisfy the visible-disclosure duty.
The icons also need to work for everyone. The Commission encourages accessibility measures: a clearly visible size, plain language without jargon (other than “AI”), alt text or ARIA labels so assistive technologies can read the disclosure, and (where a label appears only briefly) keeping it on screen long enough to be read and understood.
What is machine-readable marking, and how is it different from a label?
Machine-readable marking is the invisible half of Article 50. Under Article 50(2), providers of generative AI systems must mark their outputs (synthetic audio, image, video and text) in a machine-readable format so that the output can be detected as artificially generated or manipulated. The law is deliberately technology-neutral: it does not name a watermark, a metadata standard or a file format. What it asks for is a technical solution that is effective, interoperable, robust and reliable, as far as technically feasible.
In practice, the marking usually takes one or more of three forms. The first is embedded metadata, such as Content Credentials (C2PA), which record the tool that made the file and the edits applied to it. The second is a digital watermark, visible or invisible, baked into the pixels or the audio. The third is a fingerprint or hash that a detection service can recognise. The Code of Practice asks signatories to use at least two layers of machine-readable marking for audio, image, video and containerised text, because any single layer can be stripped: a screenshot removes metadata, a crop can defeat a watermark, and a re-encode can break a fingerprint.
That fragility is exactly why the law has two halves. A visible label tells the person looking at the content; a machine mark tells the software checking it. Neither survives every journey (a plain-text copy and paste carries no metadata, and a screen recording can lose a watermark) which is why the Commission’s guidance treats marking and labelling as complementary, not interchangeable.
As a viewer, you can check the machine-readable side yourself. Our guide to checking Content Credentials (C2PA) on a photo walks through reading the provenance record of an image file, and it explains the honest limit: a file with no credentials is not proof the image is real, only that no provenance record survived. For video, the same logic applies, a missing mark proves nothing, but a present one is strong evidence. Our guide to checking whether a viral video clip is AI-generated shows how to combine provenance checks with keyframe searches and platform labels, and it records a real finding from November 2026: OpenAI’s “This is Sora 2” promotional clip carried no AI label anywhere on its YouTube page.
The other half of the marking story is detection. Article 50(2) does not only ask providers to mark their outputs; the Commission’s guidelines give real weight to the second part of the duty, that means of detection must be available to the people exposed to the content. Marking without available detection does not achieve compliance, in the guidelines’ words. That is why the Code of Practice sets 2 February 2027 as the date by which signatories must have an interoperability solution for watermark detection, so that a mark made by one tool can be read by another tool’s detector.
What are the deadlines? 2 August 2026, 2 December 2026 and 2 February 2027
Three dates matter, and confusing them is the most common mistake in coverage of this law.
| Date | What happens | Who it affects |
|---|---|---|
| 2 August 2026 | Article 50 transparency obligations apply: chatbot disclosure, deepfake labels, public-interest text labels, emotion-recognition notices, and machine-readable marking for systems placed on the market from this date | All in-scope providers and deployers |
| 2 December 2026 | End of the four-month transition for machine-readable marking under Article 50(2) for generative systems already on the market before 2 August 2026; new prohibitions on AI systems generating non-consensual intimate imagery also apply from this date | Providers of pre-August generative systems |
| 2 February 2027 | Watermark-detection interoperability commitment for signatories of the Code of Practice | Code signatories only |
The 2 December 2026 date is the one most often misreported. It is not the day labelling starts. The labelling duties have applied since 2 August 2026, with no grace period at all. What ends on 2 December 2026 is a narrow, four-month transition added by the Digital Omnibus on AI, Regulation (EU) 2026/1744, which entered into force on 27 July 2026. Its new Article 111(4) says that providers of generative AI systems placed on the market before 2 August 2026 must take the necessary steps to comply with Article 50(2) by 2 December 2026. Systems placed on the market on or after 2 August 2026 never had a grace period and have had to mark their outputs from day one.
So the accurate way to describe 2 December 2026 is that it is the day the last lawful reason for unmarked synthetic content from an in-scope provider expires. From 3 December 2026, there is no cohort of generative AI systems on the EU market for which Article 50(2) does not bind. If a provider of a pre-August system has done nothing by then, it is not exercising a grace period, it is running one out.
Two other points about timing are worth knowing. First, the Digital Omnibus moved the AI Act’s high-risk deadlines (stand-alone high-risk systems to 2 December 2027 and AI embedded in regulated products to 2 August 2028) but it did not move Article 50. Anyone who tells you labelling was delayed is confusing the two. Second, the rules are generally not retroactive: content generated before 2 August 2026 does not need to be labelled retrospectively, according to the Commission’s guidelines, although AI-generated text published on or after 2 August 2026 must be labelled even if it was generated earlier, because for text the publication date is what matters.
What is exempt from labelling?
Article 50 is narrower than many headlines suggest, and its exceptions are part of the law, not loopholes in it.
| Exception | What it covers | What still applies |
|---|---|---|
| Evidently artistic, creative, satirical or fictional work | Deepfakes that form part of an evidently artistic, creative, satirical, fictional or analogous work | Disclosure in an appropriate manner that does not hamper the display or enjoyment of the work |
| Human review and editorial control (text only) | AI-generated text on matters of public interest that went through human review or editorial control, with a named person holding editorial responsibility | Nothing under Article 50(4) for that text, but the review must be genuine, not a rubber stamp |
| Authorised by law (crime detection) | Deepfakes or public-interest text used to detect, prevent, investigate or prosecute criminal offences, where authorised by law | The authorisation must be real; it is not a general press or research exception |
| Obvious AI interaction | AI systems where it is obvious to a reasonably well-informed person that they are interacting with AI | The Article 50(1) disclosure duty only; other duties still apply |
| Content generated before 2 August 2026 | AI content generated before the obligations applied | No retroactive labelling; but text published on or after 2 August 2026 must still be labelled |
The art and satire exception is the one creators ask about most. If a deepfake forms part of an evidently artistic, creative, satirical or fictional work, the transparency duty is reduced to a disclosure in an appropriate manner that does not hamper the display or enjoyment of the work, for example, a credit line rather than a banner across the image. The key word is “evidently”: a realistic fake presented as news does not become art because the publisher later calls it satire. The Commission’s guidelines treat the exception as applying where the artistic or satirical character is clear to the audience.
The human-review exception is equally specific, and it applies to text only, there is no equivalent exception for deepfake images or video. For AI-generated text on a matter of public interest, no disclosure is required where the text has undergone a process of human review or editorial control and a natural or legal person holds editorial responsibility for its publication. Both conditions must be met together, and the review must be real: the Commission’s guidelines warn that the mere possibility of a person reviewing outputs cannot be used to dodge the rule. A human who actually reads, checks and takes responsibility for the text is what counts.
There is also a practical boundary that is not an exception but a scope limit: Article 50(4) only reaches text published with the purpose of informing the public on matters of public interest: politics, public health, public security and similar topics. Ordinary marketing copy, a product description, or professional correspondence between people is not published to inform the public on a matter of public interest, so the text duty does not bite there. But the deepfake duty is different: it applies to image, audio and video regardless of topic, so an AI-generated product photo that makes the product look better than reality can still be a deepfake that must be labelled.
How do you check whether a photo or video is labelled as AI?
Article 50 is written for publishers, but it changes what viewers should look for too. When you open a photo or video that might be AI-generated, run through these checks in order.
- Look for the label first. Since 2 August 2026, a deepfake published in the EU should carry a visible disclosure at first exposure: an on-image note, a caption, or one of the EU icons. On social platforms, look for the platform’s own label as well: our guide to AI labels on YouTube, TikTok, Instagram and Facebook records the exact wording each platform uses, and shows that a missing label proves nothing, Instagram, for example, carried no label on a clearly computer-generated influencer post we checked in November 2026.
- Check the file’s provenance. If you have the original file, look for Content Credentials or other machine-readable marks. A present mark is strong evidence; an absent one is not evidence of anything, because marks are stripped by screenshots, downloads and re-uploads.
- Reverse-search a still frame. Take a clear frame from the video, or the photo itself, and run it through a reverse image search. Our how-to reverse image search guide walks through the full method. A frame that appears only on the publisher’s own pages is consistent with a fresh AI generation; a frame that appears on stock sites or older news pages points to a real, reused photo.
- Test the image itself. If the photo still looks suspicious after the first three steps, work through the visual checks in our guide to telling whether an image is AI-generated: hands, text, reflections and background detail are still the fastest tells, even as generators improve.
- Ask who benefits. A label that is present but vague (“AI-enhanced” on a photo that is entirely generated) is not a clear disclosure. Article 50 requires the disclosure to be clear and distinguishable, and a label that understates what the AI did fails that test.
None of these steps proves an image is real on its own. Together, they are the same layered approach the law itself takes: a visible label for people, a machine mark for software, and a search for the original when both are missing.
What happens if you do not label AI content?
Breaching Article 50 can result in fines of up to โฌ15 million or 3% of total worldwide annual turnover, whichever is higher, under Article 99 of the AI Act. That is the transparency tier of fines; the higher โฌ35 million ceiling applies to the prohibited-practices tier, not to labelling. Enforcement sits with national market surveillance authorities together with the Commission’s AI Office, and it began on 2 August 2026.
In practice, the first risk for most small publishers is not a fine. It is a complaint, a platform takedown, or a loss of trust when an audience discovers an undisclosed AI image. The Commission’s own reasoning for the labelling rules is framed in those terms: the icons are intended to support public trust, reduce the risk of misinformation, and help people make informed judgements about what they see. A label is cheap; a discovered deception is expensive.
There is also a practical reason to label even where you think you might be exempt. The exceptions (art, satire, human review) are fact-specific, and the burden of showing they apply will fall on the publisher. A short, honest label removes the argument entirely. For the same reason, keep a record of your labelling decisions: which icon or wording you used, where it was placed, and why you chose it. If a regulator ever asks, that record is your evidence.
Worked test: three real cases through the Article 50 test
To show how the rules work on real content rather than in the abstract, I ran three genuine cases through the Article 50(4) three-question test on 11 November 2026, and checked the Commission’s own labelling tools the same day.
Case 1: OpenAI’s “This is Sora 2” promotional clip on YouTube. This clip was already checked live in our earlier AI-video guide: on 10 November 2026 we opened the YouTube page for the 30 September 2025 video and found no AI label anywhere on the page, no badge under the title, no disclosure in the description. Apply the three questions: the clip was generated by AI, yes; it resembles real scenes, yes; but could a viewer perceive it as a real recording? The video is titled “This is Sora 2”, presented by OpenAI as a demonstration of its own generator, so a viewer who reads the title knows what it is. It is not passed off as genuine footage, which is why it sits outside the deepfake definition even though no formal label is shown. The lesson is that context can do the work of a label, but only when the context is unmissable.
Case 2: the viral “Pope Francis in a white puffer jacket” image. This Midjourney image from March 2023 is the best-known AI fake photograph ever published: a real person, rendered realistically, in a scene that never happened. Under the three-question test it is a textbook deepfake, AI-generated, resembling a real person, and capable of being mistaken for a genuine photograph. If a deployer published it in the EU today as if it were real, Article 50(4) would require a clear disclosure. But the image was generated before 2 August 2026, and the Commission’s guidelines say content generated before that date does not need to be labelled retrospectively. The date of generation matters, not just the date of publication, except for public-interest text, where the publication date governs.
Case 3: the EU icons themselves. On 11 November 2026 I opened the European Commission’s EU icons page and confirmed what the page promises: three icons (Basic, Fully AI-Generated and Partially AI-Modified) each in four variations (black, white, and both at 50% transparency), downloadable as SVG and PNG ZIP files, free to use without attribution. The page also states the two caveats that matter most: use of the icons is optional, while the labelling duty is not, and using an icon does not by itself establish legal compliance. I also confirmed on the Commission’s Code of Practice page that the code was published on 10 June 2026, has two sections (providers and deployers) and had about 190 signatory organisations by the end of July 2026. Signatories can rely on the code’s measures to demonstrate compliance; everyone else must show their own measures are adequate, assessed individually by different market surveillance authorities.
The pattern across all three cases is the same: the law asks for a visible, honest disclosure at the moment a person first sees the content, backed by an invisible machine mark in the file. Where either half is missing, the viewer is left to do the checking themselves, which is exactly what the labelling rules exist to prevent.
FAQs
Does the EU AI Act require me to label every AI-generated image I post?
No. Article 50(4) requires a visible label only for deepfakes (AI image, audio or video that resembles real people, objects, places or events and could be mistaken for genuine) and for AI-generated text on matters of public interest. An obviously fantastical illustration, an unpublished private image, or minor AI-assisted editing is outside the duty.
What is the EU AI Act labelling deadline in December 2026?
2 December 2026 is the end of a four-month transition for machine-readable marking under Article 50(2), and it applies only to generative AI systems already on the market before 2 August 2026. The visible labelling duties for deepfakes and public-interest text have applied since 2 August 2026 with no grace period, and systems launched after that date never had one.
Are the EU AI icons free, and do I have to use them?
Yes, the EU icons (Basic, Fully AI-Generated and Partially AI-Modified) are free to use without attribution, in SVG or PNG format, from the European Commission’s website. No, you do not have to use them: the icons are optional, but the labelling duty is mandatory. You may design your own label, and using an icon does not prove compliance.
Do the EU AI Act labelling rules apply outside the EU?
Yes, in effect. The duties apply to providers and deployers outside the EU whose AI output is used in the EU, so a US or UK publisher reaching an EU audience is in scope. The test is where the content is seen, not where the company is based, and fines can reach โฌ15 million or 3% of worldwide annual turnover.







