The labelling duty lands on multimodal output first
The EU transparency rules now in force require generated or materially altered content to be labelled. Text is the easy case. The obligation bites hardest on image, audio and video, where the provenance chain is longest and the tooling is least mature.
The transparency obligations that became enforceable on 2 August require that content generated or materially altered by AI carries a clear label. Applied to text this is largely a UI question. Applied to image, audio and video it is a provenance engineering problem that most of the industry has not solved.
The difficulty is compositional. A single deliverable may involve a generated base image, a model-based upscale, a human edit, a generated audio bed and an automated caption. Which of those makes the artefact "materially altered"? The regulation does not enumerate, and the honest answer is that reasonable people will disagree until an enforcement decision settles it.
Watermarking and content credentials exist and are being adopted, and neither survives the pipeline reliably. Metadata is stripped by platforms on upload. Perceptual watermarks degrade through re-encoding, cropping and screen capture — which is to say through normal use.
The compliance posture that seems defensible is to label at the point of generation, preserve the credential where the pipeline allows, and document where it does not. That is a records exercise as much as a technical one. It is also the obligation that is genuinely live, unlike the high-risk regime being marketed against.
European Commission — Commission starts enforcing AI Act rules and new transparency requirements on 2 August → · Help Net Security — EU begins enforcing AI Act, putting AI models under the microscope →