Do I have to label this?

AIDEAID Knowledge · The AI labelling compass · v1.2 · as of 2026-07-15 · a snapshot; interpretation is still evolving
Aug 2, 2026 Since this day, the transparency obligations under Article 50 of the EU AI Act (Regulation (EU) 2024/1689) have applied. Since then, some AI content in marketing must be labelled as such.

Not every AI image and AI text needs a label, but some do. For the typical marketing assets, this compass tells you in under two minutes: must label, label when in doubt, or no obligation. Every classification is backed by the regulation text (EUR-Lex) and cross-checked by a second, independent review step.

Important: This is guidance, not legal advice. The compass covers the common cases; for edge cases (especially photorealistic images and the question of “public interest”), have it checked by a lawyer when in doubt.

1One ground rule first

Before you check individual assets, one distinction settles 90% of the cases: are you a provider or a deployer?

As an SME using off-the-shelf AI tools, you are a “deployer”.

If you use Midjourney, Sora, ElevenLabs, ChatGPT and the like professionally, you legally count as a deployer (not a provider). That relieves you of the most demanding obligation and leaves only two:

Tool providerThe technical, machine-readable marking (invisible watermark, metadata) is owed by the maker of the tool, not by you (Article 50(2)). Your only duty here: do not remove or obscure this marking.
YouOnly two visible disclosure obligations apply to you: the chatbot notice (Article 50(1)) and the disclosure for deep fakes and for AI text on matters of public interest (Article 50(4)). If an asset falls under neither, no labelling obligation of your own under Article 50 applies to you.

2The compass

Tap a tile for the reasoning, the practice (how exactly to label), and the legal basis in the regulation text. The filter above the tiles narrows the view to one status.

Filter:
Must label (obligation) Label when in doubt No labelling obligation

3If you must label: how, exactly?

The core rule is in Article 50(5): clear, distinguishable, and recognizable at first sight or hearing, accessible. The EU published official icons for this in June 2026.

✦ AI GENERATED

Fully AI-generated

For content created entirely by AI (image, video, audio).

✦ AI MODIFIED

AI-modified

For real recordings that were substantially modified with AI.

✦ AI

Base mark

The compact base sign for when space is tight (e.g. small formats).

Since June 10, 2026, the three icons have been available for download from the European Commission, free and without attribution, as SVG and PNG (four light/dark variants each). Using them is voluntary; the labelling obligation itself is not. A text label in your national language is equally permissible. A clear notice is all it takes:

Created with AIAI-generatedYou’re chatting with an AI assistantAI-generated voice

Placement: for images and video in a corner or the caption, for video a disclaimer at the start (continuous for live video), for chatbots in the greeting or as a label in the window, for text visibly on the piece.

4What this compass does not cover

Article 50 only governs AI transparency. Further obligations apply in parallel and make a missing or wrong label additionally contestable.

Unfair competition law (UWG)If AI advertising that looks like a genuine recording lacks the notice, that can already qualify as misleading today (Sections 5, 5a UWG) and can be pursued with a cease-and-desist letter via Section 3a UWG. A double attack surface: the supervisory authority and your competitors.
Personality and image rightsAn AI avatar or a cloned voice of a real person needs that person’s consent. But the consent does not replace the AI label: these are two separate obligations.
Fines, put honestlyThe statutory range for Article 50 violations goes up to €15 million or 3% of annual turnover (Article 99). For SMEs, however, the lower of the two applies, and fines must be proportionate: a minor labelling violation does not trigger a million-euro penalty. The range is the ceiling, not the norm.
Competent authorities (being set up)In Germany, the Bundesnetzagentur (the federal network agency) will become the central market surveillance authority; in Austria, the AI service desk at RTR (the telecom and media regulator). As of mid-2026, the national implementing laws are not yet in force everywhere; the EU obligation itself has applied since August 2, 2026 regardless.

5And if you set it up right?

Labelling is not a drag. It is a habit you set up once and that then runs along: the right label in the right place, a fixed sign-off step for AI texts, the chatbot notice in your standard setup. Exactly these recurring moves are what we automate with you, so compliance happens on the side instead of causing a headache every time.

Talk it through for your business

Methodology: Researched and cross-checked on July 14, 2026 against the official regulation text (EUR-Lex, CELEX:32024R1689), Recitals 132–134, the “Code of Practice on Transparency of AI-Generated Content” (European Commission, June 10, 2026), and the guidance published by the Commission, DIHK, IHK, Wettbewerbszentrale, RTR, and WKO. Every classification was checked against the primary text by a second, independent legal review step. This page is a snapshot: the Commission’s final Article 50 guidelines and the national implementing laws are partly still in progress.

Unsure about your specific case, or newer findings? Write to us: axel@aideaid.ai