Last updated: July 2026
Reading Time: 6 min read
If your compliance team was waiting for the regulatory ink to dry before locking down your generative AI disclosure workflows, your time has arrived.
The European Commission and the European AI Board have formally finalized two critical compliance instruments governing Article 50 transparency obligations under the EU AI Act:
- The Code of Practice (CoP) on Transparency has been formally assessed as adequate to demonstrate compliance with Articles 50(2), 50(4), and 50(5).
- The EU Commission’s Article 50 Guidelines have officially transitioned from draft status to their final, binding interpretive form.
Together, these developments provide clear, actionable rules for AI providers and deployers ahead of the initial August 2, 2026 enforcement milestone. Here is what changed, why the Code of Practice matters, and how to calibrate your operational strategy.
1. The Code of Practice (CoP): Your Compliance Safe Harbor
Developed via a multi-stakeholder process facilitated by the EU AI Office, the Code of Practice provides a practical blueprint for operationalizing synthetic content marking and deepfake disclosures. With the adequacy conclusions delivered by the European Commission and AI Board, the CoP is currently the only EU-wide recognized compliance framework.
Why Becoming a CoP Signatory Matters
While adhering to the Code does not grant immunity from national surveillance authorities, it substantially shifts your enforcement posture:
- Enforcement Predictability: Market surveillance authorities will focus primary oversight on checking adherence to the Code rather than conducting ad-hoc, unpredictable investigations.
- Penalty Mitigation: In enforcement actions, regulatory authorities may consider CoP adherence as a mitigating factor when calculating administrative fines.
- Reduced Audit Friction: Non-signatories must independently prove that their alternative compliance mechanisms satisfy EU requirements on a country-by-country basis.
eyreACT Takeaway: For organizations deploying generative AI across multiple EU jurisdictions, signing the CoP establishes a harmonized compliance baseline across all 27 Member States.
2. Final Guidelines: Key Interpretive Updates
While the final Guidelines retain the general structure of the draft version, the Commission introduced critical refinements that directly impact engineering, product, and legal teams.
AI Agents & System Architecture
AI agents are now explicitly addressed. Where an AI agent interacts directly with natural persons—or where such interaction is reasonably foreseeable—transparency disclosures must be built directly into the system’s architecture. The agent must disclose both its artificial nature and the identity of the entity on whose behalf it acts at key checkpoints (e.g., authorization, reporting, and validation workflows).
Expanded Content Exclusions
The final Guidelines expand the list of AI outputs exempt from Article 50(2) marking obligations:
- Exempt Categories: AI-generated source code, AI translations, very short outputs, machine-to-machine communication, and intermediate outputs in closed-loop production pipelines (provided final user-facing outputs are marked).
- B2B / Industrial Carve-Out: In-house and internal B2B tools remain exempt only if outputs never leave the organization, operates in controlled environments, and robust technical safeguards against external dissemination are in place.
Deepfakes, Advertising, & Creative Works
- Broad “Existing” Standard: The four-element deepfake test under Article 3(60) covers photorealistic portraits of non-existent individuals if they could plausibly exist.
- Strict Creative Exception: Reduced labeling rules for artistic, fictional, or satirical deepfakes will be strictly interpreted. Commercial content and standard advertising cannot rely on light-touch exceptions; default labeling applies.
Legacy Content & Retroactivity
- No Retroactive Labeling: Media generated prior to August 2, 2026 does not require retroactive watermarking or labeling.
- Public Interest Text Exception: AI-generated text regarding matters of public interest generated before August 2 must be labeled if it is published on or after August 2, 2026 (unless human editorial responsibility is established).
3. Transparency Compliance Timeline
| Target Date | Obligation | Scope & Requirements |
|---|---|---|
| August 2, 2026 | Article 50(4) | Labeling obligations take effect for deepfakes and AI-generated text informing the public on matters of public interest. |
| December 2, 2026 | Article 50(2) | Machine-readable marking and detection deadlines apply for generative AI systems placed on the market (under the AI Omnibus extension). |
| February 2, 2027 | Interoperability | Implementation deadline for providers to deploy interoperable watermark detection solutions. |
4. Immediate Action Items for Providers & Deployers
- Maintain Existing Roadmap: The final Guidelines affirm earlier draft frameworks. If your system architecture was designed around draft guidance, no major pivots are required.
- Evaluate CoP Participation: Review whether becoming a formal CoP signatory aligns with your product strategy to secure enforcement predictability.
- Audit Multi-Party Distribution Chains: Deployers distributing AI-generated content through third parties should update vendor agreements to ensure downstream tools preserve transparency tags.
- Update AI Agent Workflows: Ensure product development teams embed clear AI identification prompts at key interaction milestones within conversational agent frameworks.
