China Made 'Digital Slop' an Official Enforcement Category and Deleted 5.61 Million Items: The Take Inside China

On September 2, 2026, the Cyberspace Administration of China posted the phase-two report of its “Qinglang — Fixing AI Application Chaos” special action, and in doing so promoted a piece of internet slang into an official enforcement category: 数字泔水, “digital slop.” 5.61 million items, gone. Our WeChat column’s take ran under the headline “5.61 million ‘digital slop’ posts cleared — something you scrolled past last night was probably in the pile” (2026-09-03). This entry translates the take with the official record pinned underneath.

The numbers

• Notice dated 2026-09-02, phase two of the special action first deployed 2026-04-30 (Xinhua)
• 5,610,000+ items cleaned; 49,000+ accounts punished; 2,400+ websites and apps disposed of (CAC)
• 7 problem categories; #1 is AI “remixing” of classics (Three Kingdoms, Journey to the West) — the notice’s own use of “digital slop”
• Platforms: 46 cumulative governance bulletins; multimodal recognition; dynamically expanded faceprint and voiceprint libraries (Douyin, Kuaishou, Weibo, Tencent, Baidu, Bilibili, Xiaohongshu, Zhihu, Douban, Taobao); model layer: corpus review and output limits at Doubao, Yuanbao, Qwen, Ernie Bot; app stores: Huawei, Xiaomi, OPPO, vivo screening incoming apps
• Local layer: Beijing upgraded review; Shanghai self-inspection plus AI “knowledge kits”; Zhejiang one-firm-one-policy compliance clinics; Guangdong point-to-point rectification meetings
• English-language record: SCIO’s own English press-room post and SCMP covered the sweep; English coverage of phase one (April 2026) reported ~14,000 non-compliant AI products and ~6 million posts removed [as relayed]
• Lexical backdrop: Merriam-Webster’s 2025 Word of the Year was “slop” — “digital content of low quality that is produced usually in quantity by means of artificial intelligence”

The take inside China

What slop is. Three identifying marks, per the take: it is mass-produced (one template, a swapped face and title, hundreds of copies — an account posting dozens a day is not a person); it is high-stimulation, zero-nutrition (anything goes except information); and a layer of it is outright toxic (fake disaster footage, fake celebrities, “poison cartoons” targeting children). Its distinction from ordinary bad video is industrial: slop never intended to be any good — it only has to be swallowed.

The slop factory, five links. Money is the engine (views are income); AI zeroes the marginal cost (one prompt replaces script, shoot, edit); the 49,000 punished accounts are mostly organized matrices, not 49,000 individual slip-ups; the recommender only measures dwell time, and slop is engineered to stop the thumb; and — the dirtiest link — scrapers harvest the slop back into the training data of the next generation of models. “AI raised on swill will taste of swill.” Cut the capacity end all you like; until the money flow changes, someone refills the bucket.

Three filters for readers. Check the label (AI-content marks have been mandatory since September 2025 — treat unlabeled uncanny content as guilty until proven otherwise, and report stripped labels). Check the artifacts (fingers, teeth, lip-sync, shadows, audio too clean to be real). Check the account (machine-rate posting, homogeneous output, dead comments). And the closing rule: pause three seconds before forwarding — hot-topic fakes pay the best, so hot topics deserve the extra ten seconds of cross-checking.

The campaign targets slop, not AI. Read the named categories and the boundary is clear: unbounded remixing, fabrication, impersonation, harm to minors — all “slop,” not “using AI.” The take’s own trade-off note: the blurry edge is platform over-compliance, and the notice’s simultaneous pledge to “foster a favorable environment for AI innovation and development” is the counterweight. The blade should be fast — and land on the slop.

What the Chinese take left out

The fifth link — slop recycled into training data — is the columnist’s inference; the notice says nothing about training corpora. “Digital slop” is not a CAC coinage from nothing: it is the settled Chinese rendering of the English “slop” that Merriam-Webster had already crowned, and the real signal is a regulator adopting foreign slang as a category name with published body counts. The take’s phase framing (“phase one governed sources — model registration, security review, data poisoning”) is its gloss of the April deployment notice, not the September text. And enforcement economics cut both ways, as the take itself concedes: 5.61 million is the caught portion; while payouts track dwell time, the supply re-forms. One more omission: the official notice redacts the offending account names, so the vivid examples circulating in secondary coverage are media embroideries, not the notice’s text.

Why it matters outside

“Slop” is already the English-speaking world’s word for the same phenomenon — the dictionary made it official a year before the regulator did. What China adds is operationalization: slop as an enforcement category with counts, named platform obligations (labeling, faceprint libraries, app-store screening) and local-government scorecards. That platform playbook is a preview of defenses every feed-based platform will eventually need, and the recycling loop the columnist infers — slop in, slop out — is a training-data problem for every lab that crawls the open web, not just Chinese ones.

Sources

• CAC: phase-two report, Qinglang special action on AI misuse (2026-09-02)
• SCIO press room (English): CAC cleanup of AI-generated content (2026-09-03)
• SCMP: China cracks down on AI deepfakes and clickbait on WeChat, RedNote, Douyin
• Xinhua: Qinglang AI-misuse special action deployed (2026-04-30)
• Merriam-Webster: 2025 Word of the Year — slop

Provenance & disclosure. Originally published in Chinese on our WeChat channel on 2026-09-03 (“561万条”数字泔水”被清走，你昨晚刷到的可能就在里面”); drafted with AI assistance under human editorial direction. Translated to English on 2026-09-04 (AI-assisted, human-reviewed). This entry goes beyond translation: the item/account/site counts, the seven categories, the platform and local actions and the 46-bulletin figure were verified against CAC’s official notice; the Merriam-Webster word-of-the-year claim was verified against the dictionary’s own page; the training-data-recycling link is labeled as the columnist’s inference; phase-one removal figures are marked [as relayed] via English coverage. This is translated commentary — not a SigPulse measurement. Our first-party measurements live in the dispatches and the /data/ ledger.