China’s AIGC labeling system has now been in place for a full year, and regulators have moved from rulemaking into enforcement. On Sept. 1, the Beijing Cyberspace Administration said 68 key companies had completed mutual recognition of visible labels and invisible machine-readable watermark metadata. It also said 597 billion pieces of information had been properly labeled on the generation side, while more than 800 million AI items on the distribution side were carrying labels and codes.
At nearly the same moment, the technology itself took another step forward. According to Sept. 2 reports from STAR Market Daily and Tencent News, Fei-Fei Li’s World Labs released Atlas in San Francisco. The product can generate a one-minute 1440p video from six photos. In the company’s official blind test, Atlas posted a 94% win rate against Seedance 2.5 in a specified camera-motion comparison, with Atlas reading camera coordinates directly while the competing method relied on text prompts for camera movement.
The article’s central argument is that the problem is not that governance has chosen the wrong direction. The problem is that the rules now being enforced were designed around an earlier generation of AIGC tools. Over the past year, the framework has largely targeted text-to-text and text-to-image systems, where visible tags and C2PA metadata at least leave some traceable markers. Video produced by world models such as Atlas may be different. After secondary editing and transcoding, visible watermarks can be diluted by camera motion, and invisible watermarks can be damaged by re-encoding, creating a situation where both human review and automated detection may fail.
Enforcement has become real over the past year
The piece lists several cases from the last year to show that labeling rules are no longer just guidance on paper.
- In April 2026, Jianying, Maoxiang App and Jimeng AI were publicly summoned by the Cyberspace Administration of China for failing to effectively implement labeling rules.
- In May 2026, the Liangjiang New Area Market Supervision Administration in Chongqing issued the city’s first penalty tied to a fully AI-fabricated commercial scene. The ad involved no real actors, used AI-generated crowding and purchase scenes, and lacked prominent labeling.
- On Sept. 2, 2026, the Cyberspace Administration’s second-phase “Qinglang” campaign update said more than 5.61 million illegal or non-compliant pieces of information had been removed and more than 49,000 accounts had been dealt with.
The article treats this period as the stage in which the institutional base for AIGC traceability was built. Yet it also says the current enforcement model has limits. Penalties after the fact can restrain compliant companies that leave records, but they do little against hidden generation, stripped labels and cross-platform circulation. In that sense, punishment-led governance is reaching its ceiling.
Why punishment alone is not enough
To make that point, the article turns to older regulatory histories. It cites the UK Motor Vehicle Act of 1865, which required a person carrying a red flag to walk ahead of a car and limited speed to 2 miles per hour. In the author’s reading, that was a carriage-era rule imposed on mechanical transport. Regulation only advanced when license plates, driver licensing and insurance systems took shape, making identity traceable and responsibility divisible.
Chinese e-commerce followed a similar path, the article says. Store closures and fines did not stop fraud when bad actors could simply reopen under another name. Real-name verification and credit scoring proved more effective in separating trusted sellers from the rest.
Germany is used as another comparison. Amendments to its Road Traffic Act in 2017 allowed L3 systems, and its Autonomous Driving Act in 2021 permitted regular L4 operations. The article sums that up as a three-step approach: sandbox first, standards next, credit after that.
From that perspective, AIGC labeling at the one-year mark is near the end of the second phase. The penalty framework has been established, but penalties alone cannot carry the system much further.
The article’s core view: make fraud uneconomic
The piece argues that under information asymmetry, punishment identifies people who are not afraid of penalties, but it does not identify people who are trustworthy. The “AI-generated” mark, in this framework, is a digital-era signal. A creator who labels content voluntarily is offering a traceable form of credit backing for that work.
The article draws on Erving Goffman’s ideas on sign vehicles and impression management, and on Michael Spence’s theory of separating equilibrium. The stated goal of institutional design is not to eliminate fakes entirely. It is to make fakery uneconomic. Technology will keep moving, and so will methods of arbitrage. What governance should anchor is the distribution of incentives.
Under that logic, honest creators absorb short-term traffic friction in exchange for a longer-term traceable record. Content factories face a different trade-off. If they label output, they expose the source. If they avoid labeling, they fall into an unsourced suspicious pool. Labeling, the article says, is not a moral appeal. It is a rational separation mechanism.
Platform review has distorted that mechanism
The article then argues that the mechanism has been bent at the platform screening layer. Regulators handed the first filter to platforms, but platforms often apply it like a broad security checkpoint. Everyone gets screened, even though the target is a far smaller set of high-risk actors.
One practical problem is the boundary between AI polishing and mixed human-machine creation. A human writer may use AI to refine a short product description and still be flagged as suspected AI-generated content, leading to reduced distribution. Appeals are difficult because platforms often decline to explain algorithmic logic, citing trade secrets.
The piece cites a typical Beijing Internet Court case, docket (2023) Jing 0491 Min Chu No. 16846. In that dispute, user Tang had posted a little more than 200 characters of purely human, instant writing, but the platform misclassified it as AI-generated and imposed traffic restrictions and muting. The court ultimately found the platform in breach.
That, in the article’s view, leaves rule-following users trapped while serial content factories can return under fresh identities.
When the label itself becomes a low-quality signal
The article says another complication is emerging inside content platforms: the words “AI-generated” are drifting from a compliance marker toward a signal of lower quality.
It uses Xiaohongshu as an example. The platform handled 56,500 AI-generated notes impersonating celebrities and 1,372 related accounts during the year. The article argues that malicious operators never intended to follow the labeling rules in the first place. What they care about is whether the traffic-arbitrage model still works. In contrast, if platform algorithms still treat a label as a negative signal, honest creators may be punished precisely because they disclose.
In that case, labeling no longer narrows the suspicious set. It screens out the most compliant participants instead.
Platform policy is beginning to shift
The article points to some changes in platform behavior. In an August 2026 notice, Xiaohongshu said that voluntarily labeling AI-generated content or AI virtual identities would not affect traffic, and that high-quality creators who disclose early could gain priority access to new AI tools and features.
Douyin has also moved its AI self-disclosure entry to the first-level submission page. The article reads this as a shift from treating labeling as a deduction item to treating it as a liability shield, or even as a ticket to early access. Once honest disclosure starts to accumulate credit points, creators have a reason to think in longer time frames.
That is why the article says the issue is no longer only about governance. It is becoming an industry story as well.
Four possible business layers around AIGC labeling
1. Provenance infrastructure
The first layer is provenance infrastructure, including watermark SDKs and issuance gateways tied to C2PA and China’s GB 45438 standard. The article says these tools may move from optional features to compliance necessities.
Its estimate for mid-sized AI content producers such as MCNs, corporate marketing teams and digital publishers is an annual technical upgrade cost of roughly RMB 500,000 to RMB 1.5 million to connect to a labeling pipeline, with operations and verification costs on top.
It also cites The Business Research Company, which estimated in its 2026 report that the global market for C2PA content provenance solutions rose from $1.63 billion in 2025 to $2.06 billion in 2026 and could reach $5.12 billion by 2030. The article says domestic demand in China is currently more concentrated in implicit labeling and metadata gateways under GB 45438-2025, while C2PA issuance is more relevant for export content and multinational brand workflows.
2. Compliance audits
The second layer is compliance auditing. As of the end of June 2026, the Cyberspace Administration of China had recorded 988 filed generative AI services and 598 registered services, according to the article. It argues these services must show that their labeling systems comply with GB 45438-2025.
The article lists four core lines of review: visible label placement and font size, invisible metadata fields, tamper-resistant signatures and interfaces for regulatory verification. Missing any one of them could expose a provider to traffic restrictions or a formal summons in a later inspection round.
It says Alibaba Cloud, Venustech and NetEase Yidun currently bundle audit functions together with watermarking and cloud security products, while the standards institute platform handles regulator-side self-check functions. The comparison here is with the early stage of vehicle inspections, when the public security system ran its own testing stations. The market worked, but neutrality remained open to question.
The article notes that vehicle inspection only became a major private-sector market after China’s Road Traffic Safety Law in 2004 established socialized annual testing and a 2014 policy required government exit from direct inspection operations. It argues that AIGC audit services could follow a similar path if regulators begin accepting compliance reports from independent third parties. It cites publicly reported labeling reports issued by Shanghai Motor Vehicle Inspection Center for in-vehicle large models used by Li Auto and Mercedes-Benz as an example of the kind of role an independent verifier could play.
3. Credit layering
The third layer is credit layering. The article compares a label to a content insurance policy, with credit scoring acting like variable car-insurance premiums. Creators with a long record of compliant labeling could receive higher traffic weighting and greater brand preference, while accounts that repeatedly fail to label could face lower exposure.
It sees early signs in Xiaohongshu’s August statement that active labeling would not hurt traffic and that quality creators would get priority access to new functions, as well as in Douyin’s move to surface AI disclosure earlier in the posting flow.
Three business paths are outlined here. The first is evidentiary services against false positives. After the Beijing Internet Court case, human creators misjudged by algorithms need “proof of human authorship” for appeals. The article says United Trusted Timestamp has launched an AIGC PAS platform with a base price starting at RMB 10 per file and cites adoption in more than 120,000 judicial documents. The second is software for internal platform credit stratification. The third is a cross-platform credit passport. The article says brands may eventually require creators with credit scores above 90, even though data walls between platforms are unlikely to disappear soon. A shared standard is not required for internal scoring systems to start working.
4. Verification and inspection
The fourth layer is verification. “Labeling is writing; verification is reading,” the article says. Once a mark is attached, platforms, regulators, brands and ordinary users still need tools to inspect it.
The capabilities named include C2PA manifest parsing, extraction of SynthID or Alibaba hidden watermarks, and GCmark verification for national-standard numbering segments. The article says verification and inspection applications are the fastest-growing subsegment in the global C2PA provenance market.
It also says China’s broader digital watermark market is worth about RMB 5.28 billion, with verification work now mainly handled by the standards institute’s GCmark, Alibaba Cloud AI tools, Venustech MACCW and NetEase Yidun. Independent third-party verification SaaS remains limited.
The comparison here is to anti-counterfeit systems in consumer goods. The code itself costs very little. The continuing fees sit in the verification infrastructure and the brand-side risk-control backend. Unlike the first layer, which sells tools up front, verification can produce recurring revenue through per-call billing and annual subscriptions. The article adds that once metadata is stripped during reposting across platforms, robust watermark extraction may become the last meaningful gate.
Its conclusion: smarter credit matters more than harsher punishment
The article closes with an image from the early automotive era. A decade from now, it says, today’s “AI-generated” labels may look like the first license plates on early cars: awkward, conspicuous and open to scrutiny from onlookers.
But the point of that first plate was not elegance. It was traceable identity for rule-followers and less room to hide for those who were not following the rules. In the article’s framing, the state builds the road and the labeling rules set the direction. The real question for founders and investors is which stretch of road can collect tolls.
Its final judgment is clear. The first anniversary of AIGC labeling is not an endpoint. It is the opening ticket to a credit-based passage system. As Fei-Fei Li and others push generative technology toward world models, the thing most likely to keep up is not heavier punishment, but a smarter credit system.