Making AI transparency work: four implementation questions for the Article 50 AI Code of Practice

Introduction

The European Commission’s (the Commission) voluntary Code of Practice on Transparency of AI-generated Content (the Code) aims to facilitate the effective implementation of the transparency obligations under Article 50 of the EU Artificial Intelligence Act (the AI Act).

Following the publication of an initial draft and stakeholder consultations, on 3 March 2026, the Commission released a second draft of the Code (the Draft Code).

The Code is intended to provide practical guidance on how organisations can demonstrate compliance with requirements to apply machine-readable marks to AI-generated and AI-manipulated content, and to apply visible, human-readable labels to content that falls within the AI Act’s definition of “deep fakes” and certain AI-generated text, ahead of those obligations generally becoming applicable on 2 August 2026 (subject to the EU’s forthcoming Digital Omnibus on AI establishing a transitional period for certain Article 50(2) AI Act marking obligations).

We have been tracking the development of the Code closely. Building on the substantial work undertaken by the Commission, the AI Office and the AI Board, we recently convened a roundtable discussion with industry, supported by Amazon, to debate and discuss the Draft Code. The discussion brought together a diverse range of stakeholders from across the AI value chain, spanning the tech, education and training, creative, industrial and entertainment sectors. Participants included European-headquartered companies, spanning large operators, scale‑ups and SMEs, and providers of AI tools, digital content generation technologies and online services, including Black Forest Labs, Colossyan, Photoroom and Synthesia.

This briefing highlights the key implementation issues identified by participants, with the intention of informing the ongoing debate and supporting the successful drafting of a Code that is workable, proportionate and fit for the future.

1. Feasibility and proportionality of imperceptible watermarking

Sub‑Measure 1.1.2 of the Draft Code requires imperceptible watermarking as a default element of compliance and Measure 2.3 of the Draft Code sets out detailed expectations around verification and detection, including how results should be communicated to users, the techniques used, and the provenance of the detected content.

Participants raised practical concerns regarding the feasibility and proportionality of such proposed requirements across content modalities and deployment contexts, given the limitations in the reliability and maturity of available watermarking and detection approaches, the integration of third‑party and upstream models, and the compute, infrastructure and operational burdens associated with implementing and maintaining multi‑layer marking approaches, particularly for scale‑ups and SMEs. For companies operating globally, there is also a concern that the Draft Code's mandatory marking requirements go further than existing approaches in other major jurisdictions, such as California, South Korea, China and India.

Key considerations and views from the roundtable: The Code should emphasise outcomes based and standards driven transparency approaches rather than prescribed watermarking techniques.

2. Prescriptiveness of testing, verification and monitoring obligations

Measure 4.2 of the Draft Code requires providers to regularly test and monitor marking and detection solutions in real-world conditions, using adaptive and use-case-specific threat modelling. It further requires providers to update benchmarks and other measurement and testing methodologies, periodically re-evaluate detection thresholds, and to document and address compliance shortcomings or adversarial attacks reported by third parties

Participants generally recognised the importance of these measures in supporting transparency objectives. However, many expressed concern that the Draft Code goes beyond the outcomes‑based requirements set out in Article 50 AI Act and could give rise to ongoing compliance expectations that may be difficult to scale, particularly for SME providers.

Key considerations and views from the roundtable: The Code should frame testing, verification and monitoring measures as voluntary best practice guidance, rather than compliance requirements.

3. Scope of personnel training requirements

Measure 4.3 of the Draft Code requires signatories to make proportionate efforts to provide appropriate training to personnel responsible for ensuring that the marking and detection measures under the Draft Code are effectively implemented.

Participants generally acknowledged the importance of training and internal awareness of transparency measures as part of responsible AI governance and adoption, and many organisations already invest in internal education and cross functional coordination. However, some noted that Article 4 AI Act already addresses AI literacy and is being operationalised by providers in scope of the AI Act – a separate mandatory training requirement for transparency was therefore considered beyond the scope of Article 50 AI Act.

Key considerations and views from the roundtable: The Code should recognise and encourage personnel training as a part of wider AI literacy efforts, rather than framing it as a standalone compliance obligation.

4. Breadth of the proposed labelling obligation for deep fakes and certain AI generated or AI manipulated content

The Draft Code also includes a labelling framework to operationalise the Article 50(4) AI Act requirement for deployers to disclose when image, audio or video content has been generated or manipulated by AI and constitutes a deep fake, and similarly where AI-generated or AI-manipulated text is published in order to inform the public on matters of public interest.

Participants highlighted the need for clearer guidance on when synthetic content meets the "deep fake" threshold under the AI Act. Without a clear distinction between benign or non-deceptive uses of synthetic content and content that is genuinely false or misleading, there is a risk that broad and undifferentiated labelling could lose explanatory value over time and, in turn, weaken rather than reinforce consumer trust.

Key considerations and views from the roundtable: The Code and final Commission guidelines should clarify the scope of deep fakes by reference to a context-based threshold, such as harm or deception.