AI content generators have made it effortless to produce convincing text, voice, images, and video at scale. That convenience has also amplified familiar online harms, misinformation, fraud, impersonation, and consumer deception, because synthetic media can be created and shared faster than people can verify it.
In response, the European Commission is moving toward a practical labeling and marking framework under the EU AI Act. A new voluntary “code of practice” is being developed to help organizations meet the AI Act’s transparency duties, especially for deepfakes and certain AI-generated communications, before the rules become applicable on 2 August 2026.
Why the EU is pushing labeling for AI-generated content now
The Commission has framed the upcoming code as a direct response to real-world risks: misinformation campaigns, fraud schemes, impersonation, and consumer deception. When synthetic content is hard to distinguish from authentic material, trust in the wider information ecosystem erodes, particularly during elections, crises, or market-moving events.
At the center of the EU’s approach is transparency: making it clearer when people are seeing, hearing, or reading AI-generated or AI-manipulated content. The goal is not to ban synthetic media, but to reduce the likelihood that it is used to mislead audiences in ways that cause harm.
The Commission also emphasizes practical implementation. A labeling regime only works if it can be applied consistently across platforms and tools, and if it can be audited when regulators ask how an organization complied. That is why the code is positioned as technical and operational guidance rather than a purely policy-level statement.
How Article 50 of the AI Act sets the transparency baseline
The legal anchor for the initiative is Article 50 of the EU AI Act, which establishes transparency obligations for certain AI systems and outputs. In the Commission’s wording, Article 50 “includes obligations… to mark AI-generated or manipulated content in a machine-readable format,” alongside user-facing disclosure in defined cases.
Machine-readable marking matters because it enables automated detection, by platforms, fact-checkers, and other tools, at internet scale. Human-readable labels alone can be stripped, cropped, re-encoded, or simply ignored, whereas robust marking schemes aim to travel with the content (or at least remain detectable) across common transformations.
Crucially, Article 50 is paired with application timing. Across Commission materials and subsequent commentary, a key compliance date is repeated: 2 August 2026, when these transparency obligations become applicable. The code of practice is meant to provide a clearer path to compliance a of that deadline.
From consultation to drafting: the EU timeline you should track
The Commission’s work did not begin with the first draft. On 4 September 2025, it launched a consultation and a call for expressions of interest to develop AI transparency guidelines and a code of practice, inviting stakeholders to share their views by 9 October 2025. This was the precursor step that set the process in motion.
On 5 November 2025, the Commission formally kicked off work on a new voluntary code of practice for marking and labeling AI-generated content, covering deepfakes and synthetic text/audio/image/video. The Commission linked the effort to the AI Act’s transparency duties, highlighting machine-readable marking and clear disclosure, especially in public-interest communications.
Then, on 17 December 2025, the Commission published the first draft of the Code of Practice and opened a public feedback window ending 23 January 2026. The Commission’s stated timeline is specific: a second draft is due mid-March 2026, the final code is targeted for June 2026, and the underlying transparency rules apply from 2 August 2026.
What the first draft covers: providers vs deployers
The Commission describes the draft as structured in two sections. The first section focuses on duties for providers related to marking and detection, how AI-generated or manipulated content should be marked “in machine-readable formats to enable detection.” This is aimed at making compliance technically feasible and demonstrable.
The second section focuses on deployers, organizations that use AI systems in real contexts, and their obligations to label deepfakes and certain AI-generated text in public-interest scenarios. In other words, the code tries to clarify both the upstream technical responsibilities and the downstream disclosure responsibilities.
The Commission’s policy materials also reflect this split operationally: two working groups aligned to Article 50’s structure, one for Providers and one for Deployers. For providers, the ambition is to develop solutions that are “effective, interoperable, robust, and reliable,” as far as technically feasible.
Machine-readable marking: what it is and why it’s hard
Machine-readable marking typically means embedding signals that detection systems can read, often via watermarking, metadata, cryptographic provenance techniques, or hybrid approaches. The Commission’s emphasis on machine-readable formats signals a preference for scalable detection over purely visual labels that can be easily removed.
But robust marking is technically difficult. Content gets compressed, resized, translated, re-recorded, and remixed; attackers can also attempt deliberate “washing” to remove detectable patterns. That is why interoperability and reliability are highlighted as design goals: a fragmented labeling ecosystem could fail in practice even if every individual tool claims compliance.
From a compliance perspective, the technical details will matter: what “marking” counts, how detection should be tested, and what evidence a provider can retain to show it applied marking appropriately. The code is explicitly framed as practical guidance intended to be usable in compliance demonstrations, not merely aspirational principles.
Disclosure and labeling: deepfakes and public-interest AI text
Alongside machine-readable marking, the code targets “clear disclosure” obligations, especially for deepfakes. The basic idea is straightforward: if content is manipulated or generated in ways that could mislead, audiences should be informed in a clear and timely manner.
The Commission also highlights disclosure expectations for certain AI-generated text in public-interest communications. While the details will be refined through drafts, the direction is clear: when AI outputs could shape public understanding of important issues, transparency is expected so that recipients can calibrate trust and seek verification when needed.
Operationally, this will likely translate into product and communications changes for deployers: UI labels, content notices, contextual explanations, and internal policies for when disclosures must appear. It also raises practical questions for organizations that distribute content across channels where labeling formats differ (web, social platforms, audio-only streams, broadcast video, and printed derivatives).
Governance: a stakeholder-driven process led by the European AI Office
The Commission characterizes the drafting as an “inclusive, seven-month, stakeholder-driven process” led by independent experts appointed by the European AI Office. The aim is to incorporate technical reality from industry, civil society concerns about harms, and public-sector needs for enforceability.
According to the Commission’s library entry for the first draft, it is “the result of a collaborative effort involving hundreds of participants” and includes contributions from Member States. That breadth is intended to strengthen legitimacy and increase the chance that the final code is workable across different markets and sectors.
This governance design also signals how the EU expects the code to function: not as a niche technical annex, but as a shared reference point for what “good” compliance looks like. Even though the code is voluntary, wide stakeholder participation can make it influential in practice.
Why a voluntary code may still become the de facto standard
Legal commentary in January 2026 argued that the draft labeling code is likely to become a “benchmark” for regulatory expectations. That is a common dynamic in EU tech regulation: voluntary instruments can effectively harden into standards when they are repeatedly cited in audits, procurement requirements, and supervisory dialogues.
The Commission’s own timeline supports that interpretation. With a first draft in December 2025, feedback until 23 January 2026, more drafts in the first half of 2026, and a final targeted for June 2026, organizations will be pressured to align early so they can meet the 2 August 2026 application date for Article 50 transparency obligations.
Secondary reporting on the 5 November 2025 launch also reinforced the same seven-month drafting window and August 2026 applicability. The practical implication is that teams responsible for product, trust and safety, legal, and security may treat the code as a near-term roadmap rather than optional reading.
Don’t confuse it with the separate GPAI code of practice
One reason for confusion is that the EU has been active on multiple “codes of practice” tied to AI Act implementation. On 10 July 2025, the EU released a separate voluntary General-Purpose AI (GPAI) code of practice, focused on areas such as transparency, copyright, and safety for GPAI models.
That GPAI code is not the same as the marking and labeling code for AI-generated content. The labeling code is explicitly oriented around Article 50 transparency duties, machine-readable marking and disclosure for deepfakes and certain AI-generated communications, rather than broad provider responsibilities for general-purpose models.
For organizations, the distinction matters because different teams may own each topic. Model governance and training data compliance might map to GPAI-related work, while watermarking, provenance tooling, and content disclosure workflows map to the Article 50 labeling and marking effort.
The EU’s push for an AI content labeling code reflects a pragmatic bet: transparency can reduce harm without halting innovation. By pairing machine-readable marking with clear disclosure duties, the Commission is aiming to make synthetic media more accountable and easier to detect, especially in high-stakes contexts.
With the first draft already published (17 December 2025), feedback closed on 23 January 2026, a second draft due mid-March 2026, and a final targeted for June 2026, the runway is short. Even as a voluntary instrument, the code is poised to shape how providers and deployers prepare for Article 50 transparency obligations that apply from 2 August 2026.
