A Second Wave of Enforcement
On September 2, 2026, the Cyberspace Administration of China (CAC) announced the results of the second phase of its "Qinglang: Rectifying AI Application Chaos" campaign: more than 5.61 million pieces of "unlawful and non-compliant" AI-generated content removed, over 49,000 accounts penalized, and action taken against more than 2,400 websites and apps. The campaign, launched April 30, 2026, targets what regulators call "AI slop" — fabricated news, deepfake impersonation of public figures, AI-remixed classic literature turned into clickbait, and coordinated inauthentic "water army" bot activity — across Douyin, Kuaishou, RedNote, and WeChat, per the South China Morning Post.
The legal basis is not new. China's 2023 Interim Measures for the Management of Generative AI Services — the world's first dedicated generative-AI statute, jointly issued by seven agencies including CAC, MIIT, and the Ministry of Public Security — already required providers to label synthetic content and screen outputs against a list of prohibited categories. A follow-on rule, the Measures for Labeling AI-Generated and Synthetic Content, took effect September 1, 2025, mandating both visible labels (text, watermarks, audio cues) and invisible metadata tags on AI outputs, with platforms required to build detection systems for unlabeled synthetic media. This campaign is CAC enforcing rules that have been on the books for one to three years — not writing new law by press release, which is itself worth noting given how much tech policy globally happens through soft guidance rather than statute.
The Case for the Rules Themselves
It is worth stating plainly what the underlying regulation gets right, because dismissing it as pure censorship theater would be dishonest. Labeling requirements for synthetic media are not a uniquely authoritarian idea — the EU's AI Act (Article 50) imposes comparable disclosure duties on deepfakes and AI-generated text, and the C2PA industry coalition (Adobe, Microsoft, OpenAI, Google) has spent years building the same explicit/implicit provenance-marking architecture voluntarily. A user who cannot tell whether the "news clip" in their feed is real footage or a fabricated deepfake of a politician is a genuine harm, not a hypothetical one — China's own regulators cite AI-fabricated disaster hoaxes and impersonation scams as the proximate trigger. Platforms hosting 2,400+ apps and sites that skipped labeling entirely, per CAC, were not meeting even the minimal 2023 obligations. On child-safety and fraud-impersonation grounds specifically, enforcement here tracks harms that any jurisdiction would legitimately act on.
Where the Model Breaks Down
The problem is not the labeling requirement — it's the enforcement architecture wrapped around it. CAC's campaign-style numbers (5.61 million pieces, 49,000 accounts) are reported in aggregate with no published case list, no named appeals process, and no judicial review comparable to a court order or an independent regulator's adjudication. Compare this to the EU's DSA enforcement against platforms, which — whatever its own flaws — proceeds through formal investigations with published decisions and a right of judicial appeal to the European Court of Justice. A Chinese account operator whose content is removed under the "AI application chaos" campaign has no equivalent path to contest the classification, and the campaign's own target list bundles genuinely dangerous content (child exploitation, fraud) with categories that are far more subjective — "vulgar" remixes of classic literature, or coordinated posting patterns that could just as easily describe an organic grassroots campaign as a bot network.
That ambiguity is the actual risk, not a rhetorical one. Vague standards enforced through opaque, quota-driven sweeps are a poor substitute for narrowly drawn rules enforced case-by-case with reasons published. The 2023 Interim Measures already require generative AI outputs to reflect "socialist core values" and avoid content that "subverts state power" — categories broad enough that a platform facing a 49,000-account purge has every incentive to over-remove political and satirical speech alongside actual fraud and deepfakes, since the cost of under-compliance (being one of the 2,400 flagged apps) is asymmetric and immediate, while the cost of over-removal falls on users with no recourse.
The Innovation Cost
There's a competitive dimension too. Chinese AI labs — Doubao, Qianwen, Ernie Bot — are already tightening output review in response to this campaign, per reporting on the crackdown's platform-side effects. Compliance overhead scales with ambiguity: a bright-line rule ("label synthetic video with visible watermark X") is cheap to implement; a standard that also requires guessing whether content might later be judged part of a "chaos" sweep is not. Beijing wants both an AI industry that out-competes the U.S. on model capability and tight information control — and the more its enforcement leans on discretionary campaign sweeps rather than predictable rules, the more it taxes exactly the frontier-model developers it is also trying to promote industrially.
The fix, for regulators anywhere adopting similar labeling regimes — including the EU as it operationalizes AI Act Article 50 — is procedural, not substantive: keep the transparency mandate, publish the enforcement standard with specificity, and give account holders a real appeal before the takedown, not after.
The technology-neutral case for AI content labeling is sound. What China's campaign demonstrates is that the same authority can enforce a defensible rule through a process with none of the safeguards that make enforcement legitimate elsewhere.