標準與 AI 工程
EU AI Act Article 50 Takes Effect, Requiring Machine-Readable Marking of Generated Content
Starting August 2, the EU is enforcing transparency obligations covering chatbot notifications, machine-readable marking of generated content, and deepfake disclosures. Although rules for high-risk systems have been postponed, Article 50 has not been suspended in parallel, so product teams must still modify their output pipelines and frontend workflows.
Most provisions of the EU AI Act, including the transparency obligations under Article 50, become applicable on August 2, directly affecting products that provide chatbots, generative media, and deepfake tools in the EU. Providers must clearly inform users that they are interacting with AI. Systems that generate audio, images, video, or text must ensure their outputs carry machine-readable markings that can be detected automatically. Deployers that publicly disclose deepfake content, or publish text generated or manipulated for matters of public interest, are also subject to audience-facing disclosure requirements.
This brings compliance into the output pipeline rather than leaving it solely in the terms of service. Engineering teams must decide whether markings are added during model generation, media encoding, content delivery network processing, or download, and ensure they remain detectable after transcoding, cropping, and format conversion. Chat interfaces must also display the notification when the interaction begins instead of burying it in a privacy policy. For multi-agent systems, teams must additionally identify which component directly interacts with users and determine whether external models, image tools, or post-processors remove provenance information.
The EU’s voluntary Code of Practice recommends a layered approach combining signed metadata, invisible markings, and visible labels, but it does not establish a single mandatory technical standard. Following the legislative amendment, existing generative systems may receive a limited grace period for machine-readable marking until December 2026; chatbot notifications and deepfake disclosures have not received an equivalent blanket postponement. Another common source of confusion is that the principal obligations for high-risk AI have been postponed to December 2027 and August 2028, respectively. This does not mean that Article 50 has also been delayed. Developers should now focus on testing marking durability, defining responsibility boundaries for third-party APIs, and monitoring how national competent authorities assess whether markings are “detectable” and disclosures are “clear.”