NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

China's amended cyber law has an AI clause. It is addressed to the state, not to you

The cyber law amendment took effect on 1 January 2026 with a new AI article. It binds the state, not companies, but the new penalty ladder reaches RMB 10 million.

2026-11-16 · 1119 words · NeuroAI
China's amended cyber law has an AI clause. It is addressed to the state, not to you

Consider a compliance lead at a Chinese model company who opens the revised Cybersecurity Law (网络安全法) in January, looking for the AI section the board asked about. She finds it early in the amendment text: Article 20, four lines, and every verb belongs to the state. The law supports, advances, improves, strengthens and promotes. Nowhere does it say her company must file anything, test anything, or register anything.

That gap between expectation and text is the most commonly misread part of China's first amendment to its foundational cyber law since it took effect on 1 June 2017.

What was actually passed

On 28 October 2025 the Standing Committee of the 14th National People's Congress adopted the Decision on Amending the Cybersecurity Law of the People's Republic of China (《关于修改〈中华人民共和国网络安全法〉的决定》) at its 18th session. The decision contains 14 numbered amendments. It took effect on 1 January 2026, and the full law was re-promulgated in amended form.

Stage matters here, because a lot of English coverage has blurred it. This is not a draft, not a consultation paper, and not a five-year plan. It is an adopted amendment to a statute, in force for more than ten months as of this writing. What it is not is a new AI regulation. The operative AI rules remain the Interim Measures for Generative AI Services (生成式人工智能服务管理暂行办法), the labelling rules for AI-produced content, and the Network Data Security Regulations (网络数据安全管理条例). The amendment sits above them and changes their legal backdrop rather than their text.

The AI article, read closely

Article 20 has two paragraphs, and both are instructions to government rather than to firms:

  • The state supports basic theoretical research and R&D on algorithms and other key technologies for AI.
  • The state advances infrastructure for training data resources and computing power (算力).
  • The state improves ethical norms for AI, and strengthens risk monitoring, assessment and safety supervision.
  • The state promotes the application and healthy development of AI.
  • In a second paragraph, the state supports new approaches to cyber security management, including using AI to raise protection levels.

Read as a compliance obligation, this is empty. Read as a budget signal, it is not. Every clause names something a ministry or province can now fund without arguing about legal basis: compute, datasets, evaluation tooling. If you are raising money for an evaluation lab or a data pipeline in China, this sentence is the citation you were missing.

The other side of the amendment is less quoted and more consequential. A new paragraph in Article 42 states that when a network operator processes personal information, it must comply with the Civil Code and the Personal Information Protection Law (个人信息保护法). That is not new substance — it is a restatement — but it places AI training data squarely inside the cyber law's own enforcement machinery, which is where the fines live.

Where the money actually is

The penalty ladder was rebuilt, and it is the part that will be quoted in boardrooms.

SituationPenalty from 1 January 2026
Ordinary operator fails security duties, first findingwarning, plus a discretionary RMB 10,000–50,000
Not corrected, or harm resultsRMB 50,000–500,000, and RMB 10,000–100,000 for the responsible person
Critical information infrastructure (关键信息基础设施) operator fails duties, first findingwarning, plus a discretionary RMB 50,000–100,000
Not corrected, or harm resultsRMB 100,000–1,000,000, and RMB 10,000–100,000 for the responsible person
Mass data leak, or loss of partial CII functionRMB 500,000–2,000,000 (≈US$69,000–278,000 / HK$543,000–2.17 million), and RMB 50,000–200,000 for the responsible person
Loss of main CII functionRMB 2,000,000–10,000,000 (≈US$278,000–1.39 million / HK$2.17–10.87 million), and RMB 200,000–1,000,000 (≈US$27,800–139,000 / HK$217,000–1.09 million) personally

Two structural shifts sit inside that table. The warning stage is no longer free — a first finding can now attract an immediate fine where the old text stopped at a warning. And the ordinary-operator ceiling for uncorrected breaches moved from RMB 100,000 to RMB 500,000, a fivefold increase that lands on companies that never classified themselves as critical infrastructure at all.

For an AI company, the exposure is not a hypothetical "AI fine". It is the ordinary breach that becomes expensive because the data is a training corpus. A leaked fine-tuning dataset is a data leak; if it is large enough to count as mass data leakage (大量数据泄露), the case can jump three rows in the table.

Apps, not just websites

One small edit travels further than it looks. Where the law previously let regulators order a site to be closed, it now says "close the website or application" (关闭网站或者应用程序). Six added characters give an app-store removal a statutory footing it previously borrowed from other rules.

The same amendment also rewrote the penalties for selling network equipment that has not passed security certification — with a turnover-based multiplier of one to five times illegal gains, and licence revocation for serious cases. That matters to hardware startups shipping AI accelerators or edge boxes into regulated buyers.

One mitigation clause is worth noting because it is new: Article 73 says that where a violation falls under the Administrative Penalty Law's grounds for lighter, reduced or waived penalties, those grounds apply. Building an evidence trail before an incident is how that clause becomes usable.

The amendment's real message is quieter than its headline. China did not write a new AI law in October 2025. It put AI into the body of its oldest cyber statute, funded the plumbing, and then raised the price of getting the plumbing wrong.

Honest limitations

I have read the adopted decision text and the official explanatory articles published by the National People's Congress and the Cyberspace Administration of China, not any internal enforcement guidance. I have no data on how many cases have been brought under the new ladder since January 2026, and I would distrust any figure that circulates without an official release. The tier boundaries — what counts as "mass", what counts as "loss of partial function" — are not defined in the amendment, and I cannot tell you how a provincial regulator will draw them. My reading that Article 20 creates no corporate duty is a reading of the text, not a court holding.

What readers can do now

  • Do not build an "Article 20 compliance" workstream. There is nothing to comply with. If a vendor is selling you one, ask which clause creates the obligation.
  • Re-price your data inventory. The question is no longer whether a leak violates PIPL, but whether the corpus is large enough to reach the RMB 500,000 tier.
  • Check the app removal exposure. If your product ships as an app, the remedies list now names it directly. Review takedown contingencies accordingly.
  • Keep the mitigation file. Records of assessments, training and remediation are what Article 73 operates on.

Related coverage

Cite this article

NeuroAI (脑机网), "China's amended cyber law has an AI clause. It is addressed to the state, not to you," 16 November 2026, https://neuroai.site/a/na-policy-cybersecurity-law-ai-clause

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Policy & Governance” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories