AI Content Labelling and Content Safety
Labelling is not about announcing that AI made this. It is about being able to trace the content after it has been forwarded, cropped and re-edited.
Labelling is tested after the content travels, not at publication
The Measures for the Labelling of AI-Generated Synthetic Content and the mandatory national standard GB 45438-2025 took effect on 1 September 2025. Most companies’ first instinct is to add “AI generated” in a corner. What regulators and platforms actually test is different: after the content has been downloaded, forwarded, re-edited and uploaded again, can you still show which service produced it and who generated it.
That requires two layers. Visible labels face the user, placed on the content or in the interface so they can be perceived. Implicit labels face traceability, written into file metadata or a watermark. With visible labels only, a screenshot loses provenance; with implicit labels only, users are never informed — neither satisfies the requirement alone.
Obligations are also distributed. Generation service providers, distribution platforms and app stores each carry obligations at different points, and AI companies often occupy the first two at once.
What we do
- Labelling scheme: placement, style and wording of visible labels; metadata fields and watermark strategy for implicit labels
- Technical implementation: where labels are written and preserved across generation, editing, export and upload
- Platform obligations: verification, notice and takedown workflows for unlabelled content
- Content safety rules: moderation criteria, handling of unlawful and harmful content, user reporting and review
- Minor protection: age assurance, minor mode, content grading and time limits
- Incident response: takedowns, batch retroactive review, regulator responses
Typical situations
Situation 1: AI short film distributed without its label
The producer adds an “AI generated” notice at the end card, but the distributor trims it for platform length. After launch the content is ordered to be corrected, and producer and distributor blame each other.
Our approach: place the visible label within the content itself or on a persistent frame rather than the end card; prohibit removal in distribution agreements; spot-check each distribution channel.
Situation 2: Metadata stripped by the platform
Images carry implicit labels, but metadata is lost after compression on a social platform. In a copyright dispute the company cannot prove the content came from its own model.
Our approach: use a robust watermark that does not depend on metadata, and retain generation logs (request ID, model version, timestamp) as an independent evidential chain.
Situation 3: AI companion product accused of targeting minors
The product has no age assurance. A minor uses it for long sessions and makes purchases. Following a parental complaint, the regulator asks what protections are in place.
Our approach: implement age assurance and minor mode, restrict purchases and session length, establish parental request and content grading procedures, and keep records of interventions.
Process and deliverables
| Stage | Work | Deliverables |
|---|---|---|
| 1 | Product flow review, obligation mapping | Obligation checklist + labelling scheme |
| 2 | Visible and implicit label implementation | Labelling specification + technical notes |
| 3 | Content safety and moderation rules | Moderation policy + handling workflow |
| 4 | Minor protection and response mechanism | Minor mode plan + incident playbook |
| Ongoing | Rule monitoring, spot checks, training | Spot-check reports + training records |
Frequently asked
Must all AI-generated content be labelled? Providers of generated content services must label their output. The precise scope of visible and implicit labels depends on content type and service form; not every scenario requires the same intensity.
Is a visible label enough? No. Visible labels inform users; implicit labels enable traceability. Enforcement and platform verification generally look at both.
Users generate content with our model — who is liable? The provider bears obligations for its service. Users who remove labels and redistribute separately bear their own liability. Terms should prohibit removal, and technical traceability should be retained.
Does labelling hurt the product experience? Somewhat. The goal is a balance between compliance and experience — placement, transparency and presentation can all be tuned, but the label cannot be omitted.
Legal basis
- Measures for the Labelling of AI-Generated Synthetic Content (effective 1 September 2025)
- GB 45438-2025, Cybersecurity Technology — Labelling Method for AI-Generated Synthetic Content
- Interim Measures for the Administration of Generative AI Services
- Provisions on the Administration of Deep Synthesis in Internet Information Services
- Law on the Protection of Minors and related online protection rules