Two transparency architectures for AI content took effect on the same day
The European Union combines interaction notices, labels and machine-readable marks, while California requires large providers to offer detection, optional visible disclosure and a provider-specific hidden mark.
Artificial Intelligence··Midday
The European Union activated three layers of visibility together
The European Union's new AI Act transparency obligations began to apply on 2 August. The European Commission's announcement defines three layers, beginning with the interaction between a user and a system and extending to later detection of published content. Chatbots and other systems that interact directly with people must tell them that they are dealing with AI rather than a human. Deepfakes in the form of images, video or audio generated or altered with AI must be labelled. The same content must also carry machine-readable marks for automatic detection. Enforcement does not sit only with the AI Office inside the Commission; national authorities take part as well. The Commission says that more than 180 organisations have signed a voluntary code of practice on transparency for AI-generated content. Participation in that code is optional, while the duties in the law are mandatory. The announcement gives no penalty amount and sets out no transition period; it presents 2 August as the date on which the rules start to apply.[1]
California attaches the duty to large providers
California's law, operative on the same day, covers providers of publicly accessible generative AI systems in the state with more than one million monthly visitors or users. These providers must offer a free, public content-detection tool that does not collect personal data. Users must be given the option to add a visible AI disclosure to an image, video or audio file that the system generated. The content must also contain a hidden mark carrying the provider's name, the system version, the creation time and a unique identifier. Only the provider's own tool can detect that mark. A third party licensing the system must preserve the disclosure capability; its licence must be revoked within 96 hours if a modification disables the feature. The law provides for a penalty of $5,000 per violation per day. Separate duties begin on 1 January 2027 for large online platforms and hosting platforms that distribute model weights, and on 1 January 2028 for manufacturers of devices that capture images or audio.[2]
A shared aim with different technical links
Both regimes seek to make the origin of AI-generated or altered content discoverable later, but their technical and institutional links are not the same. The European Union framework combines identification by systems that interact directly with people, labelling of deepfakes, and machine-readable marks suitable for automatic detection. California assigns a large provider the task of offering its own free detection tool, makes visible disclosure an option for the user, and constructs a hidden mark readable only through that provider's tool. The European Commission's text separates a voluntary code joined by more than 180 organisations from mandatory legal duties. California's text expressly sets a daily $5,000 penalty and a 96-hour licence-revocation period when the disclosure feature has been disabled. Although the two approaches began on the same day, one establishes duties across interaction and content types; the other proceeds through provider scale, a provider-operated detector, and a later timetable for platforms and device manufacturers. The result is two distinct infrastructures for connecting generated content to information about its origin. That distinction also changes who performs detection and which technical tool performs it.[1], [2]
Related columns
For more information on this topic, you can read the related columns.
