New EU AI Act Transparency Requirements Take Effect on August 2
Key Highlights
The transparency provisions of the EU AI Act came into force on August 2. Two requirements sit at the core. Interactive AI systems such as chatbots must clearly tell users that they are talking to an AI. Deepfake content must carry both a label visible to the human eye and a machine-readable marker embedded in the file. On the same day the EU published the first list of organizations signing the Code of Practice on Transparency of AI-Generated Content, more than 180 in total, with Google, Microsoft and OpenAI among them, while Meta explicitly declined to join. The penalty ceiling for breaching the transparency obligations is 7.5 million euros or 1% of global annual turnover, whichever is higher.
How It Unfolded
The AI Act was designed to phase in over several stages. The prohibitions on unacceptable-risk systems landed first, obligations on general-purpose models followed, and the transparency duties fall in this round. What the August 2 date practically means is that every AI product serving users in the EU now has to make user-visible changes to the product itself, not merely refresh a compliance document buried somewhere on a corporate website.
The Code of Practice is the soft-law instrument that accompanies the statute. Signatories commit to implementing content marking according to a common specification, and in exchange they gain predictability in how enforcement will treat them. The list of more than 180 organizations spans the major cloud providers, the model companies, and a slice of content platforms. Meta declining to sign is the signal drawing the most attention. The company has previously argued in public that the Code demands more than the law itself requires and introduces additional uncertainty for model development in Europe. Refusing to sign is not in itself a violation, but it does mean Meta must demonstrate its own compliance path while forgoing the safe-harbor effect the Code confers.
Technical Details
Machine-readable marking is the genuinely hard engineering problem in this round. A visible corner badge is easy to add and just as easy to crop away, so the binding element is provenance information embedded inside the file. Two approaches dominate practice. The first is C2PA content credentials, which write the generation source, model information and edit history into file metadata and sign the whole package cryptographically. The second is invisible watermarking, which embeds a signal into the pixel distribution or audio waveform and aims to survive compression, cropping and transcoding. Each has weaknesses. Metadata is trivially lost through re-saving, screenshotting, or platform-side re-encoding, while invisible watermarks can degrade under heavy editing, so in practice the two are usually layered together.
The disclosure duty for interactive systems looks simple but its boundaries are fuzzy. Does an AI customer service agent satisfy the requirement by declaring itself once in the opening turn, or must every reply carry a reminder? How does a screenless voice assistant discharge the obligation? In a multi-agent workflow, which component bears responsibility for disclosure? These details will only be settled gradually through enforcement practice and subsequent guidance.
How It Compares
Measured against other jurisdictions, the EU has chosen horizontal legislation: a single statute covering every industry and use case, imposing tiered obligations according to risk level. The United States currently relies mainly on state-level legislation and industry self-regulation, with no unified federal mandate for content marking, producing what amounts to patchwork regulation. China has combined its Interim Measures for the Management of Generative AI Services with rules on labeling deep synthesis content, likewise requiring both explicit and implicit markers. That approach is conceptually close to the EU model but arrived earlier in practice, which means Chinese vendors already have a round of hands-on engineering experience with labeling pipelines.
Industry Impact and Applicable Scenarios
For AI products expanding overseas, the barrier to entry in the EU market has substantively risen as of August 2. Teams shipping AI companions, AI customer service, or AIGC content tools should immediately audit two things: whether any part of the product design could lead a user to believe they are interacting with a human, and whether generated images, video and audio carry compliant markers. The engineering effort involved is not large, but the cost of skipping it is calculated against global turnover.
For content platforms, responsibility is expanding from marking content they generate themselves to detecting whether upstream content already carries markers and preserving those markers through transcoding and distribution. That forces a reassessment of transcoding pipelines, thumbnail generation and compression policies, since a great many routine processing steps silently strip metadata. For the ecosystem as a whole, once a common marking specification actually works end to end, AI content detection shifts from probabilistic guesswork to verifiable fact-checking, which is a concrete win for journalism, the judiciary, insurance and any other field whose work depends on establishing whether a piece of media is authentic.