On 28 April 2026 the Cyberspace Administration of China (CAC) disclosed that cyberspace authorities had found the CapCut and Maoxiang apps and the Jimeng AI website to have failed to effectively implement AI synthetic-content labelling requirements, in violation of the Cybersecurity Law, the Interim Measures for the Administration of Generative AI Services and the Measures for Labelling AI-Generated Synthetic Content. The CAC directed local cyberspace authorities to penalise these platforms in accordance with the law.
From “making the rules” to “checking implementation”
The synthetic-content labelling regime has been in place for some time. Companies have tended to treat it as a technical task to be “configured” — add a watermark, write some metadata. This enforcement shows the focus has shifted from whether a system exists to whether it is actually implemented.
The penalised products cover video editing, AI chat and image generation — three typical scenarios pointing at the most common problem: user- or model-generated content is displayed without being clearly labelled as AI-generated. This is usually not a technical impossibility but a product missing the label in export, sharing or re-editing.
Visible and hidden labels — both are required
The labelling duty has two layers. Visible labels are perceptible to users, typically rendered in text, sound or graphics in the content or interface; hidden labels are technical measures added to the file data of synthetic content, not readily perceptible.
A common mistake is doing only the visible label. But the hidden label supports content traceability and platform verification — once content is downloaded, forwarded or re-edited, the hidden label is the basis for judging whether it was AI-generated. Missing either layer may be found to be ineffective implementation. For the technical requirements, see our algorithm filing service.
Three easily missed points
First, export and sharing. The user sees the label in the app, but it is lost when exporting to a photo album or sharing to a third-party platform — a high-frequency problem.
Second, re-editing. Whether the label is retained when generated content is edited, spliced and subtitled before publication must be handled in the editing pipeline.
Third, API-call scenarios. Applications calling already-filed models via API must also label their final output, and cannot be exempted on the ground that “the underlying model is already labelled”.
Self-inspection checklist
First, map every generation and output path. List every content exit of the product and confirm the label lands on each.
Second, check how the visible label is presented. Whether the position, form and duration make it clearly perceptible.
Third, confirm the hidden label is written at generation, not annotated afterwards, so it survives editing.
Fourth, put labelling duties into cooperation terms. In agreements with distribution platforms, MCNs and channels, clarify who carries the labelling duty along the chain.
Fifth, keep remediation records. Where prompted or reported, the process and result of remediation must be verifiable.
Labelling is the most basic link in AI content governance, and the one most easily squeezed by product timelines. Regulation has moved from “notice to rectify” to “penalise in accordance with the law”. Doing labelling properly is the lowest-cost compliance action. For related duties and filing arrangements, see our algorithm filing service.
This article is based on public information concerning the Cyberspace Administration of China’s disclosure on 28 April 2026 that cyberspace authorities penalised CapCut and other platforms over synthetic-content labelling violations. It is for general reference only and does not constitute legal advice.