AI-generated content labeling

Can AI Content Labels Be Removed? Is Helping Illegal?

What you may need to do

1. Do not strip label information by default in exporting, transcoding or compression steps
2. Avoid any one-click "clear AI label" action in editing features
3. State in your terms of service that you do not take on work to remove or forge AI labels
4. Keep evidence when your own labelled content is maliciously de-labelled, and use the complaint and reporting channels

Clause basis

One-line conclusion

No. No organisation or individual may maliciously delete, tamper with, forge or conceal the labels for generated or synthesized content provided for in these Measures, nor provide tools or services for others to carry out such malicious acts, nor use improper labeling means to harm the lawful rights and interests of others. Note that what the clause prohibits is not only stripping labels yourself; it also covers providing tools or services to others, so de-watermarking tools and label-removal services are themselves within the prohibition.

Basis

The second paragraph of Article 10: "No organisation or individual may maliciously delete, tamper with, forge or conceal the labels for generated or synthesized content provided for in these Measures, nor provide tools or services for others to carry out such malicious acts, nor harm the lawful rights and interests of others through improper labeling means."

Article 11: "Where a service provider conducts labeling activities, it shall also meet the requirements of relevant laws, administrative regulations, departmental rules and mandatory national standards." That is, the way labels are stored and handled must also stand up to comparison against other regulations and standards.

What you need to do

Product side: do not clear metadata or watermarks in export, transcoding, compression or splicing flows; when offering editing capability, do not build a one-click "clear the label" function.

Service side: make it clear in your terms of service and in requirements review that you do not take on work to remove, forge or conceal AI labels, and give sales and support staff clear boundaries on what they can say.

External cooperation: review the purpose of third-party API calls, so that your own tools are not used for bulk label stripping.

Enforcement side: when you find your own labelled content maliciously de-labelled and reposted, secure the evidence promptly and report it through the platform and regulator channels.

Internal side: include label integrity in version testing, so a label is not silently removed after an upgrade or refactor.

Common pitfalls

Thinking de-watermarking is merely a "technical process", without realising the clause expressly prohibits malicious deletion, tampering, forgery and concealment.

Tool-type products marketing "watermark removal" as a selling point, which lands you in providing tools or services for others to commit such malicious acts.

Cropping out the notice text or covering the label sticker during secondary creation and then publishing.

Using a false label to mislead others, for example putting an AI label on genuinely filmed content to dodge responsibility — that is an improper labeling means.

Labeling only in the production release while beta and canary versions are missed, so nothing can be traced once the content spreads publicly.

Common follow-up questions

Q: If format conversion causes metadata loss, will that be treated as tampering?

A: What the clause prohibits is "malicious" deletion, tampering, forgery and concealment; normal processing should still avoid losing the label as far as possible. Whether it is malicious depends on intent and the specific circumstances, and disputes go by the competent authority's determination.

Q: If I add another layer of labeling to content I generated myself, is that forgery?

A: Adding a label that meets the requirements of these Measures is itself performing an obligation; if the label added is untrue and misleads others, it may fall within harming the lawful rights and interests of others through improper labeling means.

Common follow-up questions

Is there risk in running a de-watermarking service for clients?

The clause expressly prohibits providing tools or services for others to maliciously delete, tamper with, forge or conceal labels.

If metadata disappears after transcoding, is that a violation?

Design to avoid losing the label; whether it amounts to malice depends on the circumstances and is decided by the authority empowered to make that call.

This entry belongs to the China AI Compliance & Filing cluster (42 entries in total): View all · Answers home