China's AI labelling rules meet the music apps, and NetEase says its labels are still a test

China has had one of the most detailed AI labelling rulebooks anywhere since 1 Sep 2025. Seven months later, one of its big music apps said that, for music at least, the labels were still a trial. NetEase Cloud Music told Zi Niu News (Yangtse Evening Post) that "AI labelling is currently still in a small-scale testing phase and relies more on creators proactively declaring it" (our translation).

That one sentence explains a lot about how AI music gets labelled in practice, in China and elsewhere.

What do China's rules require for AI music?

The Measures for Labelling AI-Generated Synthetic Content were issued by the Cyberspace Administration of China with the industry, public security and broadcasting regulators, and took effect on 1 Sep 2025. They cover text, images, audio, video and virtual scenes. Going by the China Law Translate English version, they set up two kinds of label:

• Explicit labels, which people can perceive, required where content could confuse or mislead the public. For audio, the rules ask for a spoken notice or an "audio rhythm" cue at the start, end or middle of the track, or a conspicuous notice in the app interface.
• Implicit labels, written into the file's metadata: what the content is, the name or code of the service that generated it, and a content reference number. Digital watermarks are encouraged, not required.

Platforms that distribute content (music apps included, on our reading) have three duties. They must check uploads for that metadata and label confirmed AI content. Where there is no metadata but the uploader says the track is AI, they must label it as possibly AI. And where neither applies but the platform detects traces of generation, they must label it as suspected AI. Platforms must also give uploaders a way to declare AI content and remind them to use it. Uploaders, for their part, must declare AI content, and nobody may maliciously strip or fake a label.

One loophole matters for producers. A user can ask a generator for output without the explicit label, as long as the service sets out the user's labelling duties in its terms and keeps logs for at least six months. The Oxford China Policy Lab points out that the metadata label still has to stay in the file in that case.

What is NetEase Cloud Music actually doing?

The Zi Niu News report, published on 17 Apr 2026, started from a listener who found out that an R&B singer she loved might be AI-generated. She learned this from comments on the singer's social media, not from the app, and took three of the songs out of her playlist.

The reporter then asked NetEase. Its customer service said the platform counts works synthesised, generated or co-created with AI as AI music, and reviews content with both automated systems and people. Content that is detected, or that users declare, gets "an explicit label in the interface". Asked why some songs carried no label, the service said the labels were in a small-scale test.

The report's most telling detail comes from Yuan Ye, a composer with a doctorate in AI from the Central Conservatory of Music. He said he filled in the AI information NetEase asks for when he uploaded his own AI music. The reporter checked his NetEase page and found that the listening pages still showed no clear AI label. So even creators who declare properly may not see a label reach listeners.

Yuan Ye also told the paper that tools such as Suno add a hidden signature that listeners can't hear but computers can detect. Even so, he said, platforms find it hard to label everything. A lawyer quoted in the piece, Cui Wu, argued that the measures are administrative rules rather than mandatory ones. In his view, enforcing them track by track would cost too much, so in practice they depend on platforms and creators policing themselves. That is one lawyer's reading, not a ruling.

How does Tencent Music compare?

Tencent Music Entertainment (TME), which runs QQ Music and Kugou, offers AI creation tools as well as distributing music. On its Q4 2025 earnings call on 17 Mar 2026, management said more than 10 million users and over 150,000 professional creators use its one-stop AI music production platform. Chief executive Ross Liang said AI songs had appeared on its charts in the previous three months and called the growth explosive, adding that most of them were covers.

Chief financial officer Shirley Hu then drew a line on money. Going by the call transcript published by MarketBeat, songs made with AI tools but sung by a human voice share revenue the same way as traditional songs. Songs with a fully AI virtual voice are not paid from the human-voice royalty pool and are handled another way. TME didn't say how it decides which side of that line a track falls on.

Where do the rules break down?

The Oxford China Policy Lab's review published on 26 Jun 2026, by research associate Zilan Qian, found that a regime that goes further than the EU or US on paper has clear gaps in practice. Its main findings:

• Metadata is fragile. Format conversion, re-exporting through editing software or re-uploading to another platform can strip it entirely.
• Detection fails both ways. Unlabelled AI videos still circulate on major platforms, while human creators get wrongly flagged as AI.
• There is no accuracy target. The rules and the national standard, GB 45438-2025, set out how to label but not how accurately platforms must detect AI content already in circulation.

Most of the review's evidence is about images, video and text rather than music. But a bounced stem or a re-encoded master is exactly the kind of export that loses metadata.

What we think

The non-obvious lesson here is that a law can put the AI label on the file and still not decide what listeners see. Between the generator's metadata and the play screen sits a platform interface, and NetEase's own answer shows that this last step is where labelling stalls. When the system "relies more on" creators declaring AI, the label ends up marking the people honest enough to declare it. Anyone who strips the metadata and stays quiet uploads as if they were human.

That gap starts to matter for money once platforms split payouts. If TME really pays fully AI vocals outside the human-voice pool, then whether a track counts as human or AI is a royalty decision, not just a transparency one. We infer the following, because neither company has described its process. Honest declarers risk less traffic and a different payout. Undeclared AI uploads compete with human artists on equal terms. Human artists wrongly flagged by a detector could, in a worst case, be moved into the wrong pool with no clear way to appeal. China's rules require labels, but say nothing about appeals.

For listeners and human artists, China's rules point in the right direction: tell people what they're hearing, and build provenance into the file. The weak points are enforcement on the play screen and the lack of any appeal route. Western services have their own versions of the same problem.

What artists can do now:

• Declare honestly where a form asks you to. It's the one signal you control, and in China it is a legal duty.
• Keep your own provenance record, because platform metadata won't survive every export. Keep project files, stems and dated session history. Whether you produce in a desktop DAW or in a browser-based AI-assisted tool like Veena, your saved sessions are the best evidence of which parts were yours.
• Ask your distributor and the platform how a wrong AI flag is reviewed, and whether a label changes how a track is recommended or paid. Neither NetEase nor TME has said so publicly.