The deadline your compliance plan is probably missing
Most EU AI Act compliance roadmaps have one milestone: December 2, 2027 — the high-risk Annex III deadline established by the May 7 Digital Omnibus provisional agreement. That date is correct. But a separate cluster of obligations, covering a far broader set of products, has two earlier deadlines: August 2, 2026 and December 2, 2026.
Article 50 of the EU AI Act. Four transparency provisions. Neither a conformity assessment nor EU AI database registration is required. They apply based on what your AI system does — regardless of whether it has been classified as high-risk.
Most enterprise products with AI features fall inside at least one Article 50 obligation.
What Article 50 covers — four distinct obligations
Article 50(1) — AI interaction disclosure (applies August 2, 2026)
When an AI system is designed to interact directly with natural persons, the provider must ensure users are informed they are interacting with an AI — unless it is "obvious from the context." The provision covers customer service chatbots, AI assistants, guided workflows, and any conversational interface where the AI-nature is not explicitly stated.
The "obvious from context" exception is narrow. An interface labelled with a product name rather than an explicit AI disclosure does not automatically qualify. Using a synthetic voice or a fast response time as the only signal does not satisfy the obligation. The Commission's draft implementation guidelines (published May 8, 2026, consultation closed June 3) address the scope and interpretation of this exception.
Article 50(3) — Synthetic media and deepfake disclosure (applies August 2, 2026)
AI systems that generate or manipulate images, audio, or video depicting real persons or real events in a way that appears authentic must ensure the output is disclosed as AI-generated or manipulated. This provision applies to deployers — not only providers — when they present such content to users.
Article 50(4) — AI-generated text in public interest contexts (applies August 2, 2026)
AI systems generating text on matters of public interest — news, public affairs, political content — must label outputs as AI-generated. The provision includes a specific exception for human editorial review, but that exception requires genuine editorial involvement, not nominal oversight.
Article 50(2) — Machine-readable watermarking (applies December 2, 2026)
This is the most technically demanding provision and has the later deadline: December 2, 2026. Providers of AI systems that generate synthetic audio, image, video, or text must mark outputs in a machine-readable format detectable as artificially generated.
Machine-readable marking is not a visible label. It requires embedding a provenance signal into the content itself — through technical approaches such as C2PA cryptographic provenance standards, frequency-domain watermarks, or provenance-aware model outputs. The December 2, 2026 deadline was confirmed in the May 7, 2026 Digital Omnibus provisional political agreement. Formal legal effect requires Official Journal publication, which is expected before the original August 2, 2026 deadline. Plan against December 2, 2026 — monitor the Official Journal for formal enactment.
Why the December deadline is closer than it looks
Enterprise product roadmaps typically plan 6–9 months ahead for technical compliance work. December 2, 2026 is under seven months away. Content generation pipelines that need to embed machine-readable watermarks require infrastructure decisions — which technical standard, which vendor, whether the GPAI provider handles embedding at the model layer or whether the product must handle it at the output layer. Those decisions cannot be deferred until October.
The Commission's draft implementation guidelines were in public consultation until June 3. Final guidelines are expected summer 2026. The guidelines specify which technical approaches satisfy the machine-readable marking requirement. Product teams should confirm their compliance path against the final guidelines when published.
The GPAI pass-through: your model provider is not your compliance layer
Product teams building on foundation model APIs (OpenAI, Anthropic, Google Gemini, and others) sometimes operate under the assumption that the model provider handles Article 50(2) watermarking requirements upstream.
This assumption needs to be confirmed, not assumed.
The GPAI Code of Practice — final version published August 1, 2025 — establishes obligations for GPAI providers who have signed the Code. Those obligations cover certain provenance commitments. But Article 50(2) creates independent obligations for product teams who deploy content generation features. Your product layer — how AI-generated content is assembled, presented, and delivered to users — carries its own compliance responsibility.
Confirm with your API provider what watermarking infrastructure they provide and when it will be available. Then map what your product layer needs to add on top.
Three steps for product and compliance teams
(Operational starting points — not a substitute for qualified legal review of your specific products and deployment context.)
Step 1: Map your Article 50 surface area. For every AI feature your product exposes to users, run three checks: Does it involve direct AI-to-human interaction? → Article 50(1). Does it generate synthetic audio, image, video, or text? → Article 50(2), and potentially (3). Does it produce text on matters of public interest? → Article 50(4). A 30-minute mapping session with product and legal using this checklist produces what most teams are missing — a complete Article 50 inventory.
Step 2: Set your August 2 target. Article 50(1), (3), and (4) apply from August 2, 2026. Disclosure changes to AI-interactive features, synthetic media tools, and AI-generated text in public interest contexts are needed before that date. Most of the required changes are not technically complex — they are product and UX changes. But they need legal sign-off, design review, and release planning. Begin that process now if you have not already.
Step 3: Confirm your December 2 watermarking path. Ask your GPAI API provider in writing: what is their Article 50(2) implementation plan? What watermarking infrastructure do they provide? By what date? Then assess what your product must add to satisfy the machine-readable marking requirement. This decision cannot wait until the final Commission guidelines are published — the lead time for implementation is real.
On the Commission guidelines
The Commission opened a public consultation on draft Article 50 implementation guidelines on May 8, 2026 — the first authoritative interpretive guidance on how Article 50 should be applied across all four provisions. The consultation closed June 3, 2026. Final guidelines are expected in the months following.
This issue summarises the Article 50 obligations as established in the EU AI Act text and confirmed in the May 7 Digital Omnibus provisional agreement. A future issue will incorporate the Commission's final interpretive positions once they are published.
Next issue: Conformity assessment — what the technical documentation process actually requires, and who on your team owns each component.