EU AI Transparency Navigator

A guided check on the labelling duty under Art. 50 EU AI Act - for image, video, text, audio and chatbot. Including the detailed deepfake check, exceptions, deadlines and UWG notes for CH, DE and AT.

Step by step to your label.

Answer a few sharp questions about your content. You get a clear labelling verdict with duties, deadlines and concrete recommendations. Little effort. Full clarity.

  • Role: provider or deployer: The Art. 50 duties differ by role - we classify you first.
  • Deepfake & exceptions: Deepfake definition, art/satire and editorial exception - checked properly, not guessed.
  • Labelling per medium: Concrete guidance for image, video, text, audio and chatbot - plus photo editor and templates.
  • Deadlines & UWG: What applies from when (2 August 2026 / 2 December 2026) - plus unfair-competition notes for CH, DE, AT.

The questionnaire at a glance

Location: Which market is the content primarily meant for?

Relevant for additional notes on the respective unfair-competition law (UWG).

  • Switzerland - Audience primarily in Switzerland
  • Germany - Audience primarily in Germany
  • Austria - Audience primarily in Austria
  • Several / DACH / EU - CH, DE, AT or EU at the same time

Role: In which role do you use AI?

The Art. 50 duties differ: deployers label visibly (para. 4), providers additionally mark machine-readable (para. 2). Most common case: deployer.

  • Provider - You build/offer an AI system under your own name
  • Deployer - You use an existing AI tool professionally
  • Unsure - When in doubt we treat you as a deployer

Type of output: What do you want to check?

One content type per run - so we can cleanly check deepfake, art exception and editorial exception. You can check further content afterwards.

  • Text - Article, post, ad copy, blog, product description
  • Image - Photo, illustration, graphic, visual
  • Video - Clip, reel, ad, explainer video
  • Audio - Podcast, jingle, voice-over, voice recording
  • Chatbot / AI assistant - Dialogue system with direct user contact
  • Emotion recognition / biometrics - You run a system that recognises emotions or categorises people biometrically - special case under Art. 50 (3)
  • None of these / internal only - No AI content with an audience - you still get a short assessment

Type of processing: Was the meaning, structure or message substantially created or changed by AI?

Routine tools (sharpening, denoising, colour correction) do not count. This is about a substantial change. Very short or purely decorative elements may not fall under the marking duty at all.

  • No - Human-made or only routine editing
  • Yes - AI created or substantially changed the content
  • Unsure - Not clear - check further when in doubt

Deepfake check: Does the content look realistic - so that a person could take it as real or authentic?

Deepfake under Art. 3 (60) AI Act: a realistic-looking AI depiction. Intent to deceive does not matter - when in doubt, treat it as a deepfake.

  • Yes, looks realistic - Realistic depiction - including invented but real-looking people/scenes
  • Unsure / grey zone - Not clear - when in doubt, towards labelling
  • No, clearly impossible / fantastical - Evidently artificial, physically impossible (dragons, flying people)

Distribution: Is this deepfake content published or distributed?

  • Yes, distributed publicly - Social media, website, ads, newsletter, TV
  • No, internal/professional only - but people see it - e.g. internal training video, employees as audience
  • No, purely private / non-professional - Personal, non-professional use (Art. 2 (10))

Art/satire check: Is the content evidently part of an artistic, creative, satirical or fictional work?

Exception under Art. 50 (4) subpara. 1 sent. 2: for evidently artistic works, disclosure must take a form that does not spoil the work.

  • Yes, evidently art/satire/fiction - Evident to the audience as a work/film/parody
  • No or unsure - No evident artistic classification - when in doubt, here

Distribution: Is the AI-generated content distributed publicly?

  • Yes, public - Social media, website, ads, print, TV
  • No, purely private / internal - Internal use only, no audience

Type of processing: Was the meaning or message of the text substantially created or changed by AI?

Mere spelling or grammar assistance does not count. Very short sequences (single words, captions, alt texts, UI labels) as well as source code/JSON and pure machine-to-machine communication do not fall under it.

  • No - Written by a human, only correction tools
  • Yes - AI created or rewrote substantial parts
  • Unsure - Not clear - check further when in doubt

Topic: Is it about matters of public interest?

This includes e.g. politics, health, science, consumer protection, societal debates.

  • Yes, public interest
  • Unsure - When in doubt, treat as public interest
  • No, e.g. pure product/ad/service text - Careful: ad copy WITH claims on health, safety or sustainability can still be public interest

Editorial review: Before publication, was the text substantively reviewed by a human who carries editorial responsibility?

The exception only applies with a genuine substantive review by a competent person with clearly attributable responsibility. Mere spelling/grammar checks are not enough.

  • Yes, substantively editorially responsible - Expert content review with clearly attributable responsibility
  • Unsure - Conditions not surely met - label when in doubt
  • No, published unedited

Distribution: Is the text published publicly?

  • Yes, public
  • No, internal only

Use context: In what context is the content used?

Drives the extra notes on fair-competition law (UWG).

  • Commercial / advertising
  • Editorial / journalistic
  • Political / opinion-forming
  • Private / non-commercial

Labelling duties under Art. 50

Disclose the AI interaction

Users must learn, at the latest on first contact and in plain language, that they interact with an AI. Not enough: terms, metadata or vague labels like "assistant". Agentic AI must also disclose itself to third parties it contacts.

  • Users must learn on first contact that they talk to an AI - terms, metadata or a vague "assistant" are not enough.
  • Agentic AI must also disclose itself to the third parties it contacts, not only to the original user.

Most tools set the notice automatically - check it once actively, otherwise add a line before the first bot turn.

Formally Art. 50 (1) addresses the provider - as a deployer you should still ensure the notice in practice (UWG + user expectation).

Legal basis: Art. 50 (1) EU AI Act - in force from 2 August 2026

Mark synthetic content machine-readable

As the provider of a generative AI system you must mark the outputs (image, audio, video, text) machine-readable as AI-generated/manipulated (e.g. C2PA/watermark). The Digital Omnibus sets the deadline to 2 December 2026.

  • The machine-readable marking hits the provider of the AI system - not you as a deployer who only uses the tool.
  • If you provide an AI system under your own name, you are the provider - then the duty hits you directly, for image, audio, video AND text.

As a pure deployer the visible label under (4) is enough - the technical marking comes from your tool provider.

The Digital Omnibus shortens the deadline to 2 December 2026.

Legal basis: Art. 50 (2) EU AI Act - in force from 2 December 2026

Disclose emotion recognition / biometrics

Deployers of emotion-recognition or biometric categorisation systems must inform affected people before use. In parallel the GDPR rules on special categories of personal data apply (Art. 9).

Legal basis: Art. 50 (3) EU AI Act + Art. 9 GDPR

Label deepfakes

Whoever publishes deepfakes (realistic-looking AI image/audio/video) must disclose the AI origin - regardless of intent to deceive. For evidently artistic/satirical works a form that does not spoil the work is enough.

  • The duty hits the deployer - you, not the tool maker (OpenAI, Midjourney).
  • It applies regardless of intent: even a harmless-looking, realistic AI image is a deepfake under Art. 50 (4).

For evidently artistic or satirical works a label that does not spoil the work is enough - a discreet note suffices.

Without disclosure, fines of up to 15 million euros or 3% of annual turnover apply (for SMEs the lower amount).

Legal basis: Art. 50 (4) subpara. 1 EU AI Act - in force from 2 August 2026

Label AI text on public-interest matters

If AI-generated text on matters of public interest is published, the AI origin must be disclosed - unless there is substantive editorial review under attributable human responsibility.

Legal basis: Art. 50 (4) subpara. 2 EU AI Act

Voluntary transparency as best practice

Even without a legal duty, a clear voluntary AI note builds trust and guards against accusations of deception (UWG).

Legal basis: Best practice

Why does Art. 50 exist?

Your system creates content or interacts with people but does not decide about them - so the transparency duties under Art. 50 apply.

What matters is that nobody is deceived: AI content and AI interaction must be recognisable as such.

This applies for example when:

  • you publish AI-generated images or videos
  • a chatbot answers customer questions
  • AI texts on topics of public interest appear

Missing labelling can cost up to 15 million euros or 3% of annual turnover (for SMEs the lower amount).

Labelling per medium

  • Image: Place the label clearly and unobstructed; also mark it machine-readable (C2PA/metadata).
  • Video: Show the AI note at the start; keep it visible or repeat it for longer clips.
  • Audio: Put a clear spoken disclaimer at the start of the track; repeat for longer formats.
  • Text: Place a clear "Created using AI" note in the primary viewing area (headline, start, byline).
  • Chatbot: Disclose at the latest on first contact in plain language; not just terms/metadata.

How to switch on the AI label per platform

Instagram (AI info)

  1. Create your post, reel or story in the Instagram app.
  2. Open the advanced settings of the post before sharing.
  3. Switch on the "AI info" option to label the content as AI-generated.
  4. Publish the post - the label shows directly on it.

Instagram labels automatically when your image carries C2PA/IPTC metadata - exactly the marking our photo editor embeds.

The label can currently only be set in the smartphone app, not on desktop.

Meta: approach to labeling AI-generated content

Facebook (AI info)

  1. Create your post, story or reel in the Facebook app.
  2. Open the advanced settings of the post before sharing.
  3. Switch on the "AI info" option to label the content as AI-generated.
  4. Publish the post - the label shows directly on it.

Facebook labels automatically when the file carries C2PA/IPTC metadata or the content comes from Meta AI tools.

Meta: approach to labeling AI-generated content

TikTok (AI-generated content)

  1. Tap the plus button and create or upload your content.
  2. Open the settings on the posting page ("More options").
  3. Switch on the "AI-generated content" toggle.
  4. Publish the post - TikTok shows "Disclosed by creator as AI-generated" on the content.

The platform labels content with C2PA metadata and TikTok AI effects automatically.

TikTok Support: AI-generated content

YouTube (Altered content)

  1. Upload your video in YouTube Studio (computer or app).
  2. In the "Details" step, answer the "Altered content" section.
  3. Pick "Yes" when realistic content is AI-generated or meaningfully altered.
  4. Publish the video - the "Altered or synthetic content" label shows in the description, for sensitive topics right on the player.

When you use YouTube's own AI tools, YouTube sets the disclosure automatically.

YouTube Help: disclosing altered or synthetic content

LinkedIn (C2PA "CR" badge)

  1. LinkedIn has no manual toggle - the platform reads the C2PA metadata of your file.
  2. Upload the image with its metadata intact (do not compress or strip it first).
  3. LinkedIn shows the "CR" badge on the image automatically; a click reveals the tool and AI involvement.
  4. Add a visible AI note in the post text as well - the badge alone is easy to miss.

The "CR" badge only shows for signed C2PA credentials (e.g. from Photoshop, Firefly or DALL-E). An unsigned XMP/IPTC marking - the kind our photo editor embeds - covers Art. 50 machine readability but does not trigger the badge.

Content Credentials (C2PA)

X (No dedicated AI label)

  1. X currently offers no labelling toggle for AI content.
  2. Label directly in the post text, e.g. "AI-generated" or #AIgenerated.
  3. Use the copy-paste blocks from the template generator further down.
  4. Place the note so it is visible at first glance - not hidden behind "Show more".

X's synthetic media policy bans misleadingly manipulated media regardless of any label.

X: Synthetic and manipulated media policy

Relevant deadlines - Digital Omnibus roadmap

  • 2 August 2026: Transparency duties (Art. 50) - The transparency duties under Art. 50 apply - unchanged from the original date despite the Digital Omnibus (chatbot disclosure, deepfake and AI-text labelling, emotion/biometrics).
  • 2 December 2026 (deferred): Provider: machine-readable marking - The Digital Omnibus cuts the provider grace period for implementing the machine-readable marking of synthetic content (Art. 50 (2)) from 6 to 3 months - new date 2 December 2026.
  • End of 2026: Switzerland: consultation draft - The Federal Office of Justice (EJPD) prepares a consultation draft on AI rules by the end of 2026 (transparency, data protection, non-discrimination, supervision) as a basis for ratifying the Council of Europe AI Convention. A dedicated Swiss AI act does not yet exist.

Glossary: roles & verdicts

Roles

  • Provider
  • Deployer

The possible verdicts

Labelling required: YES - Deepfake under Art. 50 (4) EU AI Act

  • The content is a deepfake under Art. 3 (60) EU AI Act - a realistic-looking AI depiction. Intent to deceive does not matter.
  • Public distribution requires disclosure by the deployer under Art. 50 (4).
  • What matters is the actual audience: if the content also reaches children, older or less AI-literate people, you cannot rely on the media literacy of your core audience.
  • The art/satire exception does not apply here. The duty applies from 2 August 2026.

Legal basis: Art. 50 (4) subpara. 1 Regulation (EU) 2024/1689 (EU AI Act)

Modified duty - Art/satire exception under Art. 50 (4) subpara. 1 sent. 2

  • The content is a deepfake - but evidently part of an artistic, creative, satirical or fictional work.
  • The disclosure duty still applies, but in a form that does not spoil the work (e.g. credits, metadata, description).

Legal basis: Art. 50 (4) subpara. 1 sent. 2 Regulation (EU) 2024/1689

Labelling required: YES - AI text on public-interest matters - Art. 50 (4) subpara. 2

  • The text is AI-generated, published and concerns matters of public interest (e.g. health, environment, consumer protection, politics, science).
  • There is no substantive editorial review under attributable human responsibility.
  • So Art. 50 (4) subpara. 2 applies: the AI origin must be disclosed by the deployer. The duty applies from 2 August 2026.

Legal basis: Art. 50 (4) subpara. 2 Regulation (EU) 2024/1689

Disclosure required: YES - AI interaction under Art. 50 (1) EU AI Act

  • Users interact directly with an AI system (chatbot / assistant).
  • Disclosure must happen at the latest on first contact in plain language. A note only in terms, metadata or a vague "assistant" is not enough.
  • Formally Art. 50 (1) addresses the provider - as a deployer of a bought-in chatbot you should still ensure the notice in practice. The duty applies from 2 August 2026.

Legal basis: Art. 50 (1) Regulation (EU) 2024/1689

Disclosure required: YES - Emotion recognition / biometric categorisation - Art. 50 (3)

  • The system recognises emotions or categorises people biometrically. Art. 50 (3) requires the deployer to inform affected people before use.
  • In parallel the GDPR rules on special categories of personal data apply (Art. 9 GDPR) - incl. consent and a data protection impact assessment.
  • Note: certain emotion recognition at work or in education is even prohibited under Art. 5.

Legal basis: Art. 50 (3) Regulation (EU) 2024/1689 + Art. 9 GDPR

No duty for you as a deployer - The marking duty hits the tool provider (Art. 50 (2))

  • The content is AI-generated but does not fall under a deployer labelling duty in Art. 50 (4) (no deepfake, no published public-interest text).
  • The machine-readable marking (Art. 50 (2)) hits the provider of the AI system (OpenAI, Midjourney, Google) - not you as a deployer who only uses the tool.
  • If you use the content commercially, unfair-competition transparency duties (UWG) may still apply. If you provide an AI system under your own name, you are the provider - then Art. 50 (2) hits you directly.

Legal basis: Art. 50 (2) Regulation (EU) 2024/1689

No duty (exception) - Editorial exception - Art. 50 (4) subpara. 2 sent. 2

  • The AI-generated text was substantively reviewed editorially and is published under attributable human responsibility.
  • So the exception in Art. 50 (4) subpara. 2 sent. 2 applies: the disclosure duty falls away. Document the review process (who checks what?).

Legal basis: Art. 50 (4) subpara. 2 sent. 2 Regulation (EU) 2024/1689

Labelling recommended - Grey zone - when in doubt, label transparently

  • Your content is in a grey zone: it could be a deepfake under Art. 3 (60), but a clear classification is not possible.
  • Intent to deceive does not matter - even a realistic-looking, invented depiction can be covered.
  • Important: unlike the text duty, the deepfake duty has no "published" threshold. Even purely internal but professional use very likely triggers it. Only purely private use (Art. 2 (10)) is exempt.
  • The legally safe and trust-building choice is therefore: when in doubt, label it.

Legal basis: Art. 50 (4) Regulation (EU) 2024/1689 with the EU guidelines (20 July 2026)

No labelling duty - Your content does not fall under Art. 50 EU AI Act

  • Either no relevant AI use, no public distribution, or no Art. 50 category applies.
  • A voluntary transparency note ("AI-assisted") stays best practice - it builds trust and protects against UWG accusations.

Legal basis: Art. 50 Regulation (EU) 2024/1689 (negative scope)

Sources

No warranty: AIghty20 Elevate AG delivers a structured assessment, not legal advice. We check the logic, you carry the responsibility.