The obligation to label AI-generated content under Article 50 of the AI Act applies to two cases: deepfakes and AI-generated texts on matters of public interest published without editorial review. This obligation takes effect on August 2, 2026. Not every blog post or social media graphic requires an AI label, qualification depends on the type of content and how it’s published.
This guide walks through it step by step: who is subject to the regulation, which materials require labeling, what the technical labeling process looks like (EU icons, watermarks, metadata), and what penalties apply for failing to comply.
In short
- Since August 2, 2026, Article 50 of the AI Act has been in force, requiring the labeling of deepfakes and AI texts on matters of public interest published without editorial review.
- Failure to label AI content can result in a fine of up to €15 million or 3% of a company’s global annual turnover, whichever is higher.
- A text written with AI assistance but edited and signed off by a journalist does not require an AI label thanks to the editorial review exception.
- The provider (the model creator) is responsible for technically labeling the output, while the deployer (e.g., a publisher or marketer) must disclose to the recipient that the content is artificially generated.
- The European Commission has made free AI icons available in three variants: Basic, Fully AI-Generated, and Partially AI-Modified.
- Around 190 organizations have already signed the Code of Practice on AI content transparency, with its final version published on June 10, 2026.
- The AI Omnibus of July 27, 2026 pushed back the deadlines for high-risk systems (to December 2027 and August 2028), but did not change the August 2, 2026 deadline for Article 50.
Article 50 of the AI Act – the legal basis for labeling AI content
The obligation to label AI-generated content stems from a single provision: Article 50 of the EU AI Act (Regulation 2024/1689). Not all AI content falls under this obligation – the provision precisely identifies two scenarios where a label is required.
Article 50 consists of seven paragraphs, each addressed to a different group of entities. Paragraph 1 covers chatbots and interactive systems, where users must know they’re talking to a machine. Paragraph 2 requires model providers to technically label output in a machine-readable format for audio, image, video, and text. Paragraph 4 is the most practically relevant part for publishers, mandating disclosure of deepfakes and public-interest texts published without editorial oversight.
> “Deployers of an AI system that generates or manipulates image, audio or video content constituting a deep fake, shall disclose that the content has been artificially generated or manipulated.” – Article 50(4) AI Act
The transparency requirements under Article 50 apply starting August 2, 2026, based on the implementation timeline set out in Article 113 of the AI Act.
What Article 50 of the AI Act covers – scope of the provision
The provision covers four categories of obligations: disclosing interaction with AI, technically labeling output, disclosing emotion recognition and biometric categorization, and disclosing deepfakes and texts. Paragraph 5 specifies when this information must reach the recipient, at the latest at the time of first interaction or exposure.
Providers vs deployers – who’s who under the provision
A provider is a company that develops and places on the market a model that generates content, for example a language model or image generator vendor. Its obligation under paragraph 2 ends at the stage of technically labeling the output.
A deployer, meaning the entity using such a system in day-to-day operations – a publisher, marketing agency, or content creator – is responsible for paragraph 4. Its obligation is to disclose to the recipient that specific content has been generated or manipulated by AI.
Which AI content needs labeling, and which is exempt
Two types of material require labeling: deepfakes (image, audio, video) and AI-generated texts on matters of public interest published without editorial review. Content verified by a human, satire, and art are governed by different rules.
Deepfake – the AI Act’s definition
A deepfake under the AI Act is a broader concept than the common association with a fake video of a politician. The definition covers any AI-generated or manipulated image, audio, or video that resembles existing people, objects, places, entities, or events in a way that could mislead the viewer. The material doesn’t need to depict a real person, it’s enough that it looks credible.
Exception: editorial review and human verification
Text generated by AI does not require a label if it has undergone editorial review and a specific natural or legal person has taken responsibility for it. This distinction matters for publishers: an article written with AI assistance but edited and signed off by a journalist is not subject to the disclosure requirement. Automatically generated news published without any editorial oversight is.
Exception: artistic, satirical, and fictional content
Materials of an artistic, satirical, or fictional nature have limited requirements – a discreet label that doesn’t disrupt the experience of the work is sufficient. An example from the guidelines: a live video containing deepfake elements can display the AI icon for just the first 5 seconds of the material. Basic photo edits, like color grading or sharpening, don’t trigger a labeling obligation. An additional authorised by law exception exempts systems used for detecting and prosecuting crimes.
Who must label AI content – providers, deployers, and publishers
The obligation is split between two roles defined in the Code of Practice: providers (model creators) and deployers (publishers, marketers, editorial teams). The Code divides tasks into Section 1 for providers and Section 2 for deploying entities.
Provider – machine-readable labeling at the source
A provider is responsible for ensuring that the model’s output, whether audio, image, video, or text, is marked in a machine-readable format and detectable as artificially generated. The technical solutions must be effective, interoperable, and robust against manipulation, to the extent that the current state of the art allows.
Deployer – the obligation to disclose to the recipient
A deployer, meaning a publisher running an AI-generated article or a marketer creating an ad with a generative model, must disclose to the recipient that the content is artificially generated or manipulated. Example: a publisher using a chatbot to generate news summaries without journalist involvement becomes a deployer under paragraph 4 and must label such material.
When does AI content labeling take effect – the 2026 timeline
The transparency requirements under Article 50 take effect on August 2, 2026, and around 190 organizations have already signed the Code of Practice on AI content transparency. This is the final stage in a multi-step rollout of the AI Act, spread across four years.
| Date | What takes effect |
|---|---|
| August 1, 2024 | AI Act enters into force |
| February 2, 2025 | Prohibited AI practices (1-8) |
| August 2, 2025 | Obligations for GPAI models |
| August 2, 2026 | Transparency rules under Article 50 |
| December 2, 2027 | High-risk systems under Annex III |
| August 2, 2028 | High-risk systems under Annex I |
August 2, 2026 – transparency rules kick in
From this date, providers and deployers must comply with the labeling and disclosure requirements under Article 50. The Code of Practice, which makes it easier to demonstrate compliance, went through three versions: a first draft on December 17, 2025, a second draft in March 2026, and the final version published on June 10, 2026.
The AI Omnibus of July 27, 2026 – what changed
The AI Omnibus (Digital Omnibus on AI) entered into force on July 27, 2026, pushing back deadlines for high-risk systems. Instead of the previously planned August 2027, obligations for Annex III systems will now apply from December 2027, and for Annex I not until August 2028. The Article 50 deadline hasn’t changed, the transparency rules still take effect on August 2, 2026.
Source: AI Act (Regulation 2024/1689), implementation timeline under Article 113
How to label AI content – EU icons, watermarks, and metadata
The European Commission has made a free set of AI icons available in four variants: black, white, black with 50% transparency, and white with 50% transparency. A visual label alone isn’t enough, content must also be technically marked in a way that machines can read.
3 icon types: Basic, Fully AI-Generated, Partially AI-Modified
The icons come in three usage variants. Basic is used when AI was involved in creating a deepfake or text, but the material has an additional text description. Fully AI-Generated marks content entirely generated by AI, with no human involvement beyond entering a prompt, for example a fully AI-generated video of a politician. Partially AI-Modified applies to existing, human-made content that has been altered by AI, like a face swap on an authentic photo.
Using the icons themselves is optional, but the Article 50 labeling obligation applies regardless of the method chosen. The icon must be noticeable no later than the moment of the user’s first exposure, for long enough to be noticed, in line with paragraph 5 of Article 50. The requirement also applies to content that’s reshared or downloaded.
Source: European Commission – EU icon set for labeling AI content
Watermark, invisible watermark, and the C2PA standard
Companies rely on several technical mechanisms at once. A visible watermark placed on an image is the simplest form of labeling. An invisible watermark embedded in the file’s pixels makes it possible to detect AI origin even after editing or compression. Metadata saved within the file rounds out the technical safeguard.
The C2PA standard (Content Credentials) makes it possible to verify content origin regardless of the platform a file ends up on. TikTok has committed to adopting it as part of its policy on AI content.
Google’s SynthID and IPTC metadata
Google uses its own technique for labeling images and audio called SynthID, which embeds an invisible signal in the file structure. The IPTC standard lets you save information about AI generation in metadata, which photo libraries and editorial systems can read without opening the file itself.
How to recognize AI content – detection and available tools
Besides icons and watermarks, AI content can be identified using tools that analyze metadata, language patterns, and image artifacts, known as AI fingerprints, though none of these methods guarantee complete certainty. Providers of AI systems are required to offer publicly available verification tools, such as APIs, detectors, or user interfaces, that let people check the origin of content in accordance with paragraph 2 of Article 50.
AI content detection tools – how they work
No single labeling technique is sufficient on its own. The Code of Practice recommends a layered approach, combining watermarks, metadata, and system-level signals. A Stanford HAI study from July 2025 found that AI labels don’t always reduce the persuasiveness of content, even when the audience knows the material has an AI origin. Similar conclusions come from MIT research, which questions the effectiveness of labeling alone as a protective mechanism. The ITIF organization criticizes mandatory labeling because markers are easy to remove and there’s no scalable detection method.
Can AI content be blocked on a website?
The law doesn’t provide for a technical mechanism to block AI content at the regulatory level. The AI Act imposes a labeling and detectability obligation, not a publication ban. Decisions about removing or limiting the visibility of such material fall under a given platform’s moderation policy, not EU regulation.
AI labeling checklist for publishers – 3 mistakes I see companies make
The biggest mistake publishers make isn’t skipping the icon, it’s confusing AI-assisted content with content generated without editorial review. That distinction determines whether labeling is required. Looking at rollouts across a dozen or so publishers, three recurring mistakes stand out.
The first mistake is not having a written editorial policy that specifies who reviews AI-assisted text and when. Without such a document, it’s hard to prove to a supervisory authority that the material qualifies for the human review exception. The second mistake is using an icon without an accompanying text description – user testing shows that an icon alone, without a written label, is less recognizable than an icon paired with a caption. The third mistake is failing to audit historical content published before the regulation took effect.
There’s no obligation to retroactively label material published before the rules came into force, the requirement only applies to content that’s significantly updated or republished. That said, this doesn’t remove the need for an audit: publishers should check which articles have gone back into circulation, for example via social media, after August 2, 2026.
Three steps for the week ahead:
- Write down an editorial AI policy – who reviews content, who signs off on responsibility.
- Choose a set of EU icons and pair them with a text description, not just the graphic.
- Compile a list of content from the past 12 months that may need to be republished with a label.
This approach won’t eliminate the risk of a fine of up to €15 million or 3% of global turnover, but it reduces that risk to a minimum with the least amount of effort.
Questions we get asked most often
Does all AI content need to be labeled?
No. The obligation under Article 50 of the AI Act only covers deepfakes and texts on matters of public interest published without editorial review. A graphic on a company blog or a personal social media post doesn’t require a label if it doesn’t resemble an existing person or event and isn’t about informing the public on public matters.
How can you recognize AI-generated content?
You can spot AI content through file metadata, invisible watermarks, and patterns typical of generative models, though no method offers complete certainty. AI system providers must also make public verification tools available, such as APIs or detectors, in line with paragraph 2 of Article 50.
How can you block AI content on a website or in search?
EU law doesn’t provide a mechanism for blocking AI content at the regulatory level. The AI Act requires labeling and detectability, not a publication ban. The decision to remove or limit the visibility of such material is made by the platform administrator as part of its own moderation policy, not by EU lawmakers.
What penalties apply for failing to label AI content?
Violating the obligations under Article 50 can result in a fine of up to €15 million or 3% of a company’s total worldwide annual turnover, whichever is higher. The fine amount is set by the market surveillance authority in the relevant member state.
Where should the AI label go – on the image or in the caption?
The AI icon should be placed directly on the content itself, on the image, at the start of the video, or in the interface layer accompanying the material, not just in the caption below a post. The requirement covers visibility from the user’s first exposure, and the icon must remain visible even after the content is downloaded or shared.
Sources
- sip.lex.pl/akty-prawne/dzienniki-UE/rozporzadzenie-2024-1689-w-sprawie-ustanowienia-zharmonizowanych-72359064/art-50 – Lex/Sip: full text of Article 50 of the AI Act (Regulation 2024/1689).
- digital-strategy.ec.europa.eu/pl/policies/code-practice-ai-generated-content – European Commission: Code of Practice on AI content transparency, Section 1/Section 2 breakdown, number of signatories.
- gov.pl/web/ai/zakazane-systemy-ai – Gov.pl: list of prohibited AI practices in force since February 2, 2025.
- digital-strategy.ec.europa.eu/pl/news/ai-omnibus-enters-force – European Commission: entry into force of the AI Omnibus on July 27, 2026 and new deadlines for high-risk systems.
- ec.europa.eu/commission/presscorner/detail/de/ip_26_1714 – European Commission (Presscorner): confirmation of the December 2, 2027 and August 2, 2028 dates for high-risk systems.
- digital-strategy.ec.europa.eu/pl/policies/eu-icons-labelling-ai-generated-content – European Commission: EU icon set for labeling AI content, variants and visibility requirements.
- hai.stanford.edu/policy/labeling-ai-generated-content-may-not-change-its-persuasiveness – Stanford HAI (EN): study on the limited effectiveness of AI labels in reducing content persuasiveness.
- mit-genai.pubpub.org/pub/hu71se89 – MIT GenAI (EN): analysis of the effectiveness of AI content labeling.
- itif.org/publications/2024/12/16/why-ai-generated-content-labeling-mandates-fall-short – ITIF (EN): critique of mandatory AI content labeling and detection limitations.
