China's Cyberspace Administration (CAC) announced on September 2, 2026 that the second phase of its "Qinglang: Rectifying AI Application Chaos" campaign has removed 5.61 million pieces of illegal or non-compliant AI-generated content, penalized more than 49,000 accounts, and sanctioned over 2,400 websites and apps, according to Xinhua. The targets read like a genuine list of harms: AI face- and voice-swapping used to impersonate public figures, fabricated disaster and rescue imagery designed to trigger panic, AI-generated child sexual abuse material, and networks of AI-managed accounts simulating authentic engagement ("water army" astroturfing).
A Two-Phase Campaign, Escalating in Scope
CAC deployed Qinglang's AI-chaos campaign on April 30, 2026, structured across two phases over roughly four months, per the original CAC deployment notice. Phase one, whose results Chinese state media reported in early July, targeted upstream technical compliance: unregistered large language models, inadequate safety review, training-data poisoning, and poor AI-content labeling. That phase alone removed over 14,000 AI products — websites, apps, and agents — scrubbed more than 6 million pieces of content, suspended 26,000+ accounts, and pulled 1,300+ noncompliant AI merchandise listings, per Global Times. Phase two pivoted downstream, to information content itself: deepfakes, disinformation, and platform-hosted abuse. Combined, the two phases account for roughly 11.6 million content removals and 75,000 penalized accounts in under five months — a scale no Western regulator has approached for AI-specific enforcement.
The Case For It
The strongest argument for this campaign is not hypothetical. AI-generated disaster and rescue footage circulating during real emergencies is a documented public-safety hazard — it diverts attention and resources during the exact windows when accurate information matters most. Non-consensual face- and voice-swapping used to impersonate real people, and AI-generated CSAM, are harms essentially every jurisdiction criminalizes; the debate is over enforcement mechanism, not whether the underlying conduct should be stopped. China's approach also isn't improvised: it sits on top of the Provisions on the Administration of Deep Synthesis of Internet Information Services (effective January 2023) and the Measures for Labeling of AI-Generated Synthetic Content, which took effect September 1, 2025 and require explicit and metadata-embedded labels on AI content, per Inside Privacy's summary of the CAC labeling rules. A regulator moving against synthetic-media fraud and child exploitation ahead of a rule taking full effect is, on its face, closer to enforcement than novelty.
Where the Model Breaks Down
The problem is not the target list — it's the mechanism. Qinglang campaigns operate by administrative directive, not judicial process: CAC sets quotas and categories, platforms self-audit against them, and takedowns and account suspensions happen without published case-level review or an appeals path visible to outside observers. Categories like "AI-modified classics" producing "digital garbage" (低质庸俗内容) sit in the same enforcement bucket as CSAM and disaster disinformation — content that is merely tacky gets swept by the same administrative machinery as content that is genuinely dangerous, with no proportionality signal in the aggregate numbers CAC publishes. Scholarship on China's platform-moderation architecture notes that platforms facing this kind of top-down pressure tend to over-comply — enforcing beyond what the rule text requires to avoid regulatory risk — which as one analysis in the German Law Journal observes, tends to restrict more legitimate speech than the underlying rule intends. A platform told to hit a takedown number under a four-month deadline has every incentive to over-remove and no incentive to contest a wrongful strike.
What Better Governance Would Look Like
Proportionate AI-content regulation does not require choosing between "nothing" and "5.61 million removals with no visible appeals mechanism." The elements worth importing are narrow and procedural: transparency reports that break down removals by category and legal basis rather than a single aggregate figure; a disclosed appeals channel for account penalties, distinct from platform-internal moderation queues; and content-neutral technical standards — like mandatory AI labeling, which is genuinely useful for provenance regardless of political system — kept separate from campaigns that also sweep in vaguely defined "digital garbage." The EU's approach under the AI Act and DSA, whatever its own compliance-burden problems, at least publishes rule text and case rationale that outside parties can contest. Synthetic-media fraud is a real and growing problem worth regulating aggressively. But a takedown regime that reports its own success purely in removal counts, with no independent measure of wrongful takedowns, is optimizing for throughput — not for getting the line between harmful and merely disfavored speech right.