A quiet legal shift on August 2 will force businesses across Europe to rethink how they present artificial intelligence to the public. The change arrives not with a sweeping regulatory launch, but through a set of transparency obligations that catch many organisations off guard while Brussels delays its broader high-risk AI framework.
These duties land at a strange moment. The European Union has pushed back the compliance deadline for high-risk systems to December 2027, and even further for AI embedded in regulated products. Yet the rulebook did not freeze. Instead, it fractured into multiple timelines.
Starting Sunday, any interactive AI system must clearly signal to users that a machine sits on the other side of the conversation. The only exception applies when the artificial nature of the interaction is plainly obvious from context. Systems that churn out synthetic text, audio, images or video must also carry machine-readable markers designed to survive downstream use.
The obligations split between builders and deployers. Deepfakes require prominent disclosure. AI-generated text published for public consumption needs a label unless human editors have reviewed and vetted it. Emotion-recognition and biometric-categorisation technologies trigger their own notice requirements when individuals encounter them.
These rules sweep in far more than tech firms. Publishers, ad agencies, political campaigns, retailers and government bodies all qualify as deployers when they push AI outputs into public view. The core legal test hinges not on who trained the model, but on how the resulting material reaches an audience.
The European Commission has released voluntary guidelines and a code of practice to standardise marking and labelling methods. Practical value remains uncertain. The real question is whether digital watermarks and metadata survive compression, editing and platform-to-platform sharing.
Compliance teams face a disjointed enforcement picture. National market-surveillance bodies hold key responsibilities, while the EU AI Office wields authority over general-purpose models. Cross-border cases will stress-test how quickly these bodies swap intelligence and settle jurisdictional turf battles.
For any organisation asking the immediate safety question, the calculus stays narrow. If an AI system speaks directly to people or generates content that could pass as human or genuine, the transparency provisions already bind. What happens with high-risk rules later does not change what the law demands now.
