Article 50 Is Live: The EU AI Act Deadline Nobody Deferred
Article 50 of the EU AI Act (Regulation (EU) 2024/1689) became enforceable on August 2, 2026, imposing direct transparency duties on providers and deployers of chatbots, synthetic-media generators, emotion-recognition systems, and deepfake tools — regardless of whether the underlying system qualifies as "high-risk" under Annex III. The Cloud Security Alliance published a detailed research note on this on July 29, and the core message is one many compliance teams have gotten backwards: this obligation was never part of the delay.
Why this deadline survived when others didn't
The EU's Digital Omnibus package, approved by the Council on June 29, 2026, deferred the compliance deadline for standalone high-risk AI systems under Annex III from August 2026 to December 2027, and gave product-embedded high-risk systems under Annex I until August 2028. That deferral responded to concerns that harmonized CEN-CENELEC technical standards weren't ready. But Article 50 sits outside the risk-tiered structure entirely — it attaches to a functional category of AI system (conversational, generative, emotion-detecting, or deepfake-producing), not to a risk classification, and the Omnibus negotiations left that category untouched. The practical result, per CSA's analysis: many organizations spent the first half of 2026 absorbing the message that EU AI Act enforcement had been pushed out broadly, when one of the obligations most likely to touch everyday generative AI deployments — chatbots, image generators, AI writing assistants — arrived exactly on schedule.
The one piece of relief inside Article 50 itself is narrow: a four-month runway for the machine-readable watermarking sub-obligation specifically, running through December 2, 2026. The disclosure duties for interactive systems and the deployer-side deepfake disclosure requirement took effect immediately, with no such grace period.
Four duties, four different obligated parties
Article 50 creates four distinct requirements, each attaching to a different system type and party. Providers of systems that interact directly with people (chatbots, voice assistants) must ensure a reasonably informed user understands they're talking to an AI, unless that's already obvious from context. Providers of systems generating synthetic audio, image, video, or text must mark those outputs in a machine-readable, detectable format. Deployers of emotion-recognition or biometric-categorization systems must inform the people exposed to them. And deployers of deepfake content or AI-generated text on matters of public interest must disclose the artificial origin, unless the content is clearly artistic/satirical or has undergone genuine human editorial review with assigned editorial responsibility. Major generative AI providers — OpenAI's ChatGPT and Anthropic's Claude among them — fall squarely within scope, meaning this reaches well past niche deepfake use cases into mainstream enterprise AI deployments. Non-compliance carries fines up to €15 million or 3% of worldwide annual turnover, whichever is higher, with the burden of proving timely disclosure resting on the organization.
The obligation is built on a technology that keeps losing
The synthetic-content marking duty asks providers to make AI output "detectable" through machine-readable means, but watermarking research hasn't kept pace with that mandate. Academic work presented at ICML 2024 showed an attacker querying a watermarked model's public API could both strip and forge watermarks for under $50 per attack, against schemes the researchers called "previously considered safe." A 2025 follow-up demonstrated a different technique defeating watermarks that embed their signal in high-entropy tokens — the exact design choice made to preserve text quality. On images, a public tool now defeats Google's own SynthID detector on its current model outputs. CSA's read is pointed: the statute's own "as far as this is technically feasible" language implicitly concedes watermarking isn't a solved problem, which means compliance here can't be a one-time labeling project — it needs to be treated like any other security control with active, publicly documented bypass techniques: reassessed on a cadence, backed by a decision log tied to the state of the art at the time each choice was made.
What to do now
CSA's recommendation, and a reasonable one for any organization deploying generative AI features reaching EU users: inventory every customer- or employee-facing AI system against the four Article 50 categories specifically, since the obligation attaches by function and is easy to miss if compliance work has focused only on Annex III risk classification. Confirm disclosure happens no later than first user interaction, and treat watermarking/provenance technology choices as living, documented decisions rather than a shipped-and-forgotten checkbox.