The Code of Practice on Transparency of AI-Generated Content is a voluntary code under Article 50(7) of the EU AI Act, final version of 10 June 2026. The Commission considers it adequate (C(2026) 4839), but it creates no presumption of conformity (recital 41 of Regulation (EU) 2026/1744).
For images, Commitments 1 (label design and placement), 2 (internal processes) and 3 (artistic works) apply. Non-signatories should demonstrate compliance by other means, such as a gap analysis (guidelines para. 148).
What the Code is
Article 50(7) of the AI Act tasks the Commission with encouraging codes of practice at Union level that make it easier to detect, mark and label artificially generated or manipulated content. Taking utmost account of the AI Board's opinion, the Commission assesses whether following a code is adequate to meet the duties under Article 50(2) and (4). If not, it may adopt common rules by implementing act (Article 50(7) as amended by the Omnibus).
The Code of Practice on Transparency of AI-Generated Content is such a code. It has two sections:
- Section 1 addresses providers of generative AI systems and covers machine-readable marking under Article 50(2), for example with signed metadata (see C2PA, IPTC and metadata).
- Section 2 addresses deployers who must disclose deep fakes and certain texts under Article 50(4), which includes online retailers using AI images.
How the Commission reads the provision is set out in the Article 50 guidelines.
How the Code came about
- 17 December 2025: first draft by the Commission.
- 3 March 2026: revised second draft.
- 10 June 2026: final version, with the three EU icons in Annex 1.
- 8 July 2026: Commission Opinion C(2026) 4839, published on 9 July. The Commission and the AI Board consider the Code an “adequate voluntary tool”.
- End of July 2026: around 190 signatories. The Commission's list now shows 192 for Section 2 (as of 24 September 2026).
- 2 August 2026: Article 50 applies (Article 113 of the AI Act), see deadlines.
Legal effect: voluntary, no presumption of conformity
The Code is not law. Recital 41 of the Digital Omnibus, Regulation (EU) 2026/1744, states that the codes of practice under Article 50(7) and Article 56(6) have only limited legal effect and, in particular, do not create a presumption of conformity. For that reason they do not strictly need to be approved by implementing act.
The Commission, too, writes that providers and deployers “may rely on the Code as the EU-wide recognised instrument that is considered adequate”, but that “adherence to the Code does not constitute conclusive evidence of compliance with these obligations” (C(2026) 4839, para. 52).
It still carries practical weight:
- For signatories, the Commission and market surveillance authorities focus on whether they adhere to the Code (guidelines para. 147). Signatories who opt out of sections lose that benefit.
- Non-signatories are measured against the Code (para. 148).
- When a fine is set, the deployer's technical and organisational measures count (Article 99(7) of the AI Act). The guidelines name measures under the Code as a mitigating factor (para. 149, see fines and enforcement).
The measures for images
In Section 2, Commitments 1 to 3 concern images and Commitment 4 covers published text. The Code calls measures phrased with “will” “mandatory measures under the Code”. Signatories must meet them, and market surveillance authorities monitor them. Measures with “encouraged” or “may” are voluntary recommendations (Code, p. 28). The table lists every “will” measure for images in an online store, with an obvious way to implement it.
| Reference | Content of the “will” measure | In your store |
|---|---|---|
| Commitment 1, introduction | EU icon or an equivalent icon or label that follows the design and placement specifications | EU basic icon “AI” or your own label with “AI” |
| Measure 1.1(a) | “AI” as the main element, capitalised, in English; both letters the same height; proportions preserved when resized | Do not distort the icon or replace “AI” with a translation |
| Measure 1.1, audience | Take the audience and sensitive contexts into account | Add text in your customers' language |
| Measure 1.1, accessibility | Alternatives for visual elements, high contrast, screen readers, detectability by assistive technology | Accessible name for the label, sufficient contrast |
| Sub-measure 1.2.1(a) | Immediately recognisable, without user interaction or sustained attention | Label always visible, not only on hover |
| Sub-measure 1.2.1(b) | Visible for a sufficient duration | Visible for as long as the image is visible |
| Sub-measure 1.2.1(c) | Embedded in the content or an equivalent overlay; best efforts along the distribution chain | Overlay in the store; a version with the label in the image for other channels |
| Sub-measure 1.2.1(d) | At the latest at first exposure; spacing to other overlays; visible against any background | Label appears with the image; spacing to sale badges and wishlist icons |
| Sub-measure 1.2.2(a) | A place free of overlapping elements, for example the top right corner | A free corner of the image |
| Measure 2.1 | Internal processes and documentation; a verification process where AI is used regularly | Written procedure, record per image, spot checks in the live store |
| Measure 2.2 | Awareness among staff and contractors involved | Short briefing for everyone involved |
| Measure 2.3 | Internal review and external feedback; review reported cases and remedy them without undue delay; cooperate with authorities | Correction process, reporting link, contact for authorities |
| Commitment 3(a) to (c) | Artistic and similar works: icon under Measure 1.1, clear, accessible, at the latest at first exposure, for a sufficient duration; may sit next to the image | Only where the image itself is the work, such as an art print you sell |
Not relevant for images are the measures on video (Sub-measure 1.2.2(b)), closed internal contexts (1.2.2(c)), audio and text (Commitment 4). Label details are in AI label design, the icons in EU AI icons. The guidelines read the category of artistic works narrowly (para. 122). In our reading, Commitment 3 does not apply to ordinary product and advertising images.
What to do if you do not sign
Most retailers will not sign the Code. The Commission's FAQ on signing states: “Not signing it does not constitute non-compliance with the AI Act.” Your duty under Article 50(4) stays the same, and the guidelines expect evidence:
“For instance, they should carry out a gap analysis that compares the measures they have implemented with the measures set out by a code of practice that is assessed as adequate.”
Guidelines C(2026) 5054, para. 148
Under para. 148, deployers may also be asked about their labelling practices. A gap analysis has five steps:
- List the measures: the “will” measures from the table.
- Record your status: implemented, partly, not implemented or not applicable.
- Link your evidence: screenshot, setting, record entry or written procedure.
- Explain differences: why your solution serves the same purpose.
- Keep it current: review it when your store, AI tools or the Code change.
A checklist for your internal procedure under Measures 2.1 to 2.3 and a template text for customers are in the AI disclosure statement guide.
Sign or not?
Any provider and deployer of a generative AI system may sign the Code using the Commission's signatory form, section by section but not for individual commitments (Commission FAQ). Signing gives you a clear benchmark and binds you to every “will” measure of the section.
You may use the EU icons without signing. According to the Commission, using them does not signal adherence to the Code. ai-label.app has not signed the Code and has not been reviewed by the Commission. Its tools follow the Code's design rules.
Frequently asked questions
Is the Code of Practice binding?
It is voluntary and creates no presumption of conformity (recital 41 of Regulation (EU) 2026/1744). Signatories are measured against its “will” measures (guidelines para. 147).
Do I need to sign the Code to label AI images?
No. According to the Commission's FAQ, not signing does not constitute non-compliance with the AI Act. The duty to label deep fakes under Article 50(4) applies either way.
How many signatories does the Code have?
Around 190 by the end of July 2026. For Section 2, labelling by deployers, the Commission's list shows 192 (as of 24 September 2026).
What is a gap analysis?
A comparison of the measures you have implemented with those in the Code. The guidelines give it as an example of how non-signatories can demonstrate compliance (para. 148).
Can I mention the Code in my marketing?
Only factually, for example by saying that you follow its design rules. Claiming a signature you do not have, or approval by the Commission, can be misleading.
Sources
Para. refers to the paragraph number in the European Commission's guidelines on Article 50, C(2026) 5054.
- Code of Practice on Transparency of AI-Generated Content (final version)Europäische Kommission · Code of practice · 10 Jun 2026
- Commission Opinion on the adequacy of the Code, C(2026) 4839 finalEuropäische Kommission · Commission · 8 Jul 2026
- Signatories of the Code of PracticeEuropäische Kommission · Commission · 24 Sept 2026
- Regulation (EU) 2026/1744 (Digital Omnibus on AI)EUR-Lex · Law · 24 Jul 2026
- Regulation (EU) 2024/1689 (AI Act)EUR-Lex · Law · 12 Jul 2024
- Commission Guidelines on the transparency obligations under Article 50 AI Act, C(2026) 5054 finalEuropäische Kommission · Commission · 20 Jul 2026
- EU icons for labelling AI-generated contentEuropäische Kommission · Commission · 24 Sept 2026
