One clear label, five levels — from human-authored to fully generated. A voluntary way to show how content was made, with a separate plain-language check for when the EU AI Act requires disclosure.
AI INVOLVEMENT
Voluntary label
Human-authored — A person created it. AI appeared only as a checker, a search tool or a sounding board for ideas — nothing AI-generated made it into the finished piece.
Pick the closest match for how the content was made. That is the only thing a grade describes: it is not an AI Act risk class, it does not select an EU icon, and it never decides on its own whether the law requires you to disclose anything.
AI-drafted, substantively reworked by a person
AI produced a draft or substantial components. A person with relevant knowledge substantively transformed the material — restructuring, correcting, rewriting — and approved what went out.
Typical examples
Human review at this level
Deliberate examination of the substance by someone with relevant knowledge and the authority to change or reject it.
Article 50 context
AI Act · bindingHuman review with editorial responsibility matters under Article 50(4): the duty to flag AI-generated public-interest text does not apply where the text underwent human review or editorial control and a person or organisation holds editorial responsibility. It never lifts the separate deepfake duty.
One answer suggests a voluntary grade for how the content was made. The other, kept deliberately apart from it, flags whether Article 50 might require you to disclose anything. They are different questions, and a grade never settles the legal one. Nothing is sent anywhere.
Your two answers appear here
Start with the first question on the left.
The AI Office publishes these as part of the voluntary Code of Practice on Transparency of AI-Generated Content — free for anyone to use, no attribution required. Article 50 requires clear disclosure in defined situations, but it does not mandate these particular icons, and showing one does not by itself establish compliance. Our A–E scale is a separate, voluntary description of how content was made.
AI was involved, and a custom text label or an interactive second layer carries the detail.
Pre-existing, human-made content was partly changed by AI.
The capitalised acronym AI is the main element, and its letters share one vertical dimension. Resize it however you like, as long as the proportions hold. Colour, contrast and typography are yours to choose — the mark only has to stay readable and recognisable.
Somewhere no other overlay reaches, with enough space around it, readable against whatever is behind it. The statutory standard behind this is that disclosure registers at first exposure, without the reader clicking, hovering or paying close attention.
User testing across several member states found every design was noticeable, but the variants carrying a plain word — modified, generated — were significantly clearer. Explicit text removes the ambiguity a bare symbol leaves behind.
An A–E grade cannot pick an icon for you. The grade says how a piece was made; the icons answer a narrower question about what the content depicts and how it is published. Work through it in the order the law does.
A possible deepfake, or public-interest text without review and named responsibility? Then disclosure is required — clear, distinguishable, accessible, and there by first exposure. No trigger, and any labelling you do is voluntary.
Nothing authentic ever captured points to AI GENERATED. Authentic material altered, or combined with synthetic parts, points to AI MODIFIED. This is the source axis in the caption builder, and it is a question about provenance — not about how much of the work a machine did.
The icons are one good way to disclose, published with the voluntary Code of Practice. A clear plain-language notice can satisfy Article 50 just as well, and an icon on its own does not establish compliance.
The label finder runs the same check and reports it separately from the grade.
Use the voluntary badge anywhere: hot-link it on the web, or download it for slides, documents, and PDFs. Every badge links back to a plain-language explanation of what the grade means, and carries the scheme version it was declared under.
1 · Pick your grade
2 · Pick a format
3 · Pick a size
4 · Pick a tone
5 · Copy the snippet
<a href="https://www.aiflags.eu/label/B" class="u-url" rel="ai-label"
data-ai-label="B" data-ai-scheme="aiflags-v1" data-ai-declared="2026-08-04">
<img src="https://www.aiflags.eu/badge/v1/B.svg" alt="AI involvement: B — AI-assisted" height="28" loading="lazy" />
</a>Live preview
Download for documents
The snippets point at /badge/v1/, which is pinned: those bytes never change, so a disclosure you publish today still says the same thing years from now. Want a badge that follows our latest design instead? Drop the v1/ from the URL. The PNG is double-resolution — place it at 28 px for crisp retina rendering in Word, Google Docs, Keynote, and PDF exports.
Putting it on a photo or a video frame instead? The voluntary Code recommends the top-right corner, clear of other overlays, readable at first glance and still there after a reshare — placement and captions covers each medium.
The EU icon answers one question — generated or modified. Underneath it, three axes describe where the material came from, what the model did to it, and which system did it. Together they give you alt text a screen reader can read aloud and metadata a machine can parse.
Was anything authentic ever captured? If you decide to show an EU icon, this axis says which one fits — AI GENERATED where nothing authentic was captured, AI MODIFIED where authentic material was altered. Whether you have to disclose at all is the separate Article 50 question.
Generating a whole frame, filling in one region, restyling, or just denoising are very different acts. The icon cannot tell them apart, so the caption has to.
The system name, the provider, and a timestamp. Section 1 of the Code asks signatory providers to record exactly this in the file; naming it in the caption makes it legible to people too.
1 · What kind of media
2 · Source — where the material came from
3 · Mode — what the model did
4 · Describe it, and name the tool
5 · Which voluntary grade goes with it
Which icon fits follows from the source axis. The A–E grade is a separate, voluntary judgement, and whether disclosure is required at all is a third question — the finder answers the last two side by side.
Matching EU icon — if you show one
Fully AI-generated — The whole thing is AI-made, with no human-created elements and no editorial control beyond prompting.
screen readers, the DOM
An overhead view of a solarpunk library with plants growing over the bookshelves. Fully AI-generated: made from a text prompt, using Midjourney v7.Description first, provenance last. The Code asks for plain language and no abbreviations other than “AI”, so the axis codes stay out of the spoken sentence — they belong in the caption and the metadata.
Article 50 requires disclosure to be clear, distinguishable, accessible and perceivable by first exposure or first interaction, without any special tools. The placements below are recommended implementations, and follow the Code of Practice specifications that Code signatories commit to.
Under all four, the same statutory test: it has to be recognised without the reader doing anything, stay up long enough to notice, and survive being reshared or downloaded.
| Media | EU icon | Source · mode | Caption |
|---|---|---|---|
| Image | AI GENERATED | Pure synthetic · Gen-create | Fully AI-generated (text to image) · Stable Diffusion XL |
| Image | AI MODIFIED | Authentic, modified · Gen-edit | Partially AI-modified (generative inpainting) · Adobe Firefly |
| Video | AI GENERATED | Pure synthetic · Gen-create | Fully AI-generated video (text to video) · Sora |
| Video | AI MODIFIED | Composite · Style transfer | Partially AI-modified video (video-to-video restyle) · Runway Gen-3 |
| Audio | AI MODIFIED | Authentic, modified · Style transfer | Partially AI-modified audio (voice conversion) · ElevenLabs |
| Audio | AI | Authentic, assisted · Enhance only | Optional: AI-assisted (noise reduction and mastering only) · iZotope RX |
The same model can be harmless in a video game and prohibited in a police van. Europe regulates the use case, on four levels. This is a simplified map: the Act's precise definitions, its scope provisions and its exceptions are what decide real cases, and this summary compresses all three.
Prohibited
Practices the Act prohibits where it applies — subject to its precise definitions, its scope provisions, and a few narrow exceptions.
Social scoring, manipulative or exploitative techniques, untargeted scraping of facial images, real-time remote biometric identification by law enforcement in public spaces (narrowly defined exceptions apply), and — from 2 December 2026 — AI making non-consensual intimate imagery or child sexual abuse material
Strictly regulated
Allowed, but only with testing, documentation, human oversight, and registration.
Hiring and CV screening, exam grading, credit scoring, medical devices, critical infrastructure
Transparency required
Specific uses carry transparency duties so people know when AI is involved.
Chatbots, deepfakes, certain public-interest AI text, and telling people when emotion recognition or biometric categorisation is used on them
No risk-tier rules
The Act does not add use-case rules for most low-risk applications. Voluntary codes are encouraged.
Spam filters, recommendation feeds, video game AI, inventory forecasting
Some uses cut across these boxes. Two of these cut across tiers. Emotion recognition is prohibited in workplaces and schools, with narrow exceptions for medical or safety reasons, and elsewhere can be high-risk or merely subject to disclosure. Biometric categorisation is prohibited where it infers sensitive attributes, and otherwise can be high-risk or subject to disclosure. Where a use lands depends on the context and on the Act's precise definitions.
The two scales answer different questions.A–E says how a piece of content was made. The legal risk tiers classify an AI system's use case. A content grade never determines a legal risk tier by itself.
Article 50 applies from 2 August 2026. These duties attach to providers or deployers in specific situations — not to every person who uses AI. Plain-language summaries of binding obligations; the Act itself is what governs.
Providers must build chatbots and voice agents so people are told they are dealing with AI — unless that is obvious to a reasonably well-informed person in the circumstances.
Providers of systems that generate synthetic text, images, audio or video must mark outputs so machines can detect they are AI-made. Tools that only assist standard editing, without substantially altering the input, fall outside this duty.
Deployers must tell people when an emotion recognition or biometric categorisation system is used on them, by the time they are first exposed to it. Data protection law still applies on top.
Deployers must disclose AI-made or manipulated image, audio or video that constitutes a deepfake.
Deployers must disclose AI-generated or manipulated public-interest text unless it had human review or editorial control and someone holds editorial responsibility.
Required notices must be clear, distinguishable, accessible, and given by the first interaction or exposure.
The AI Act arrives in waves. Article 50 covers specific transparency cases from August 2026; high-risk system rules follow later.
1 Aug 2024
Regulation (EU) 2024/1689 becomes law across all 27 member states.
2 Feb 2025
Prohibited practices become illegal, and the AI literacy duty starts to apply.
2 Aug 2025
Obligations for general-purpose AI models, governance bodies, and penalties start to apply.
27 Jul 2026
Regulation (EU) 2026/1744 enters into force: it pushes back the high-risk deadlines, softens the AI literacy duty, and adds new prohibitions.
2 Aug 2026
Defined transparency duties now apply, including chatbot notices and deployer disclosures for deepfakes and certain public-interest AI text.
2 Dec 2026
AI systems that make intimate imagery of a real person without consent, or child sexual abuse material, become prohibited. Generative AI systems already on the market before 2 August 2026 must also meet Article 50(2) machine-marking duties.
Upcoming2 Dec 2027
Requirements apply to high-risk uses in areas such as biometrics, education, employment, essential services, law enforcement and migration.
Upcoming2 Aug 2028
Requirements apply to high-risk AI used in products covered by the safety laws listed in Annex I.
UpcomingTwo reference sheets you can hand to a newsroom, a classroom, or a marketing team. Both are generated from the same data as this site, so they cannot drift out of step with it.
Artwork for this independent project. Not European Union publications, and not a legal decision aid.
No. aiflags.eu is an independent project, not affiliated with the European Union. The A–E scale is our own voluntary proposal for describing how content was made, modelled on the energy label everyone already understands. It is not an EU scheme, not a legal risk class, and carries no legal effect. The plain-language Article 50 material on this site is an introduction to the law, kept clearly separate from our scale.
