China AI Law Explained: What the Full-Lifecycle Rules Mean for DeepSeek, Moonshot and Alibaba

China AI Law Explained: What the Full-Lifecycle Rules Mean for DeepSeek, Moonshot and Alibaba

China is about to get a single, comprehensive China AI law — and it will govern artificial intelligence from the data used to train a model all the way to the way that model is deployed and used. On 3 November 2026, China’s top economic planner, the National Development and Reform Commission (NDRC), said it will fast-track legislation covering the entire lifecycle of AI, with safety standards and ethical guidelines spanning every stage.

The announcement matters far beyond Beijing. The rules are expected to reach DeepSeek, Moonshot AI and Alibaba — the Chinese labs whose open-weight models power tools used by developers and businesses around the world, including in the United States. If you use Kimi, Qwen or DeepSeek models, or you build products on top of them, this law will shape what those models can do, how fast they ship, and what their makers must disclose.

This guide explains the China AI law in plain language: what was actually announced, what “full lifecycle” means in practice, who is in scope, what developers will have to do, how the new law compares with the EU AI Act and the American approach, and — just as importantly — what is still unknown. No draft text has been published and no enactment date has been set, so this is a story about direction, not a finished rulebook.

China AI law — the Great Hall of the People in Beijing, where China's AI legislation is being shaped

Table of Contents

What Happened on 3 November 2026

On 3 November 2026, Wang Ruomeng, a deputy director-general of the NDRC’s Department of Innovation and High-Tech Development, spoke at a press briefing in Beijing. The briefing was the first in a series on the goals of China’s 15th Five-Year Plan, which runs from 2026 to 2030. His message on AI was short and direct: China will accelerate work on a basic, comprehensive China AI law.

Bloomberg reported the remarks first, and the NDRC’s own transcript of the briefing confirmed them. That confirmation matters, because it turns a press report into an official statement of intent from the country’s main economic planning agency. The NDRC is not a niche technology regulator. When it signals a legislative direction, the rest of China’s bureaucracy takes notice.

Wang set out four areas of work. The first, and the one that made headlines, is the governance rule system itself: speed up foundational, comprehensive AI legislation; strengthen rules on the common inputs of AI, including data, computing power and algorithms; and establish safety standards and ethical guidelines covering the entire lifecycle of AI. Implementation will be assessed periodically, and the provisions will be revised as the technology moves.

The second area is the obligations placed on companies. China will push AI developers and operators to strengthen risk management, disclose safety risks promptly, and proactively report safety incidents and vulnerability information. The third is support: reasonable transition periods, compliance guidance by sector, and evaluation, consulting and certification services designed to cut firms’ compliance costs. The fourth is the sandbox idea — sector-specific regulatory sandboxes where new AI applications can be tested before they are deployed broadly.

None of this is a law in force yet. No draft text was released at the briefing, no timetable for enactment was given, and no penalties were described. What 3 November settled is the direction and the architecture of the China AI law: one comprehensive statute, covering the full lifecycle, built on top of the rules China already has.

What “Full Lifecycle” Actually Covers

The phrase that defines the China AI law is “full lifecycle”. It sounds like bureaucratic language, but it describes a genuinely broad scope. Rather than regulating only algorithms, or only AI services after they reach users, Beijing wants rules that follow an AI system from its raw inputs through development, deployment and use.

Start with the inputs. Wang named three explicitly: data, computing power and algorithms. Data means the training data that models learn from — where it comes from, whether it was lawfully obtained, and how personal information inside it is handled. Computing power means the chips and data centres that train and run models — the hardware layer that has become a strategic resource in its own right. Algorithms mean the model architectures and the recommendation and decision-making logic built on top of them.

China AI law — computing power and data centres are among the inputs the new law will govern

Then comes development. This is the stage where a lab trains a model, tests it, and decides whether it is safe to release. Under the planned China AI law, this stage would be covered by safety standards and ethical guidelines — the kind of rules that decide what testing a frontier model must pass before it ships.

Deployment is the next stage, and it is where most of China’s existing rules already bite. A model that reaches the public in mainland China must already pass through filing and, in many cases, a security assessment before launch. The new law would bring that existing machinery under one statute rather than leaving it spread across separate measures.

Finally, there is use — what happens after a model is live. That includes how AI-generated content is labelled, how incidents are reported, and how operators manage the risks of a system that keeps learning and changing after release. The lifecycle framing matters because AI systems do not stay still. A model deployed in January may behave differently by June, and a law that only checks it once, at launch, would miss most of its life.

For readers used to the European approach, the contrast with the China AI law is instructive. The EU AI Act sorts systems by the risk of their use case — a hiring tool is high risk, a spam filter is not. The China AI law sorts by the stage of the lifecycle — data, compute, algorithms, development, deployment, use — and attaches duties to each stage. The two approaches can overlap in practice, but they start from different questions, and that difference will shape how companies comply.

Who Is in Scope: DeepSeek, Moonshot and Alibaba

The three names attached to the China AI law announcement are the three Chinese AI developers that matter most to the global market. Bloomberg reported that the rules would reach DeepSeek, Moonshot AI and Alibaba Group. It is important to be precise about attribution here: Wang Ruomeng did not name any company in the briefing itself. The names come from Bloomberg’s reporting of the announcement, and they describe the expected scope — not a final list.

That said, the three names are no surprise. Each of them sits at the centre of a different part of China’s AI ecosystem, and each of them is already covered by the site’s own reporting.

DeepSeek: the efficiency pioneer

DeepSeek became globally prominent by building frontier-class models at a fraction of the usual cost. Its open-weight models and its famously cheap API pricing forced the whole industry to rethink what AI training should cost. We covered its latest efficiency release in our guide to DeepSeek V4.1 Flash, and the open-source inference engine behind those prices in our explainer on DeepSeek DeepInfer.

For DeepSeek, a China AI law built around the full lifecycle would touch the company at every stage. Its training data practices, its computing infrastructure, and its incident-reporting duties would all fall inside the new statute. The company’s open-weight releases are also a test case for a question the law will have to answer: when a model’s weights are published openly, who carries the compliance duty — the lab that trained it, or the thousands of developers who download and deploy it?

Moonshot AI: the long-context specialist

Moonshot AI is best known for its Kimi models and its long-context work. Kimi K3.1 Pro, its latest flagship, is built around very large context windows that let it work through long documents and long-running tasks in a single session. We covered it in detail in our Kimi K3.1 Pro explainer.

Moonshot’s angle under the China AI law is the deployment stage. Kimi is a consumer-facing product with millions of users, which means it already sits squarely inside China’s existing generative-AI filing and labelling rules. A comprehensive China AI law would consolidate those duties and add the new lifecycle obligations — incident reporting, risk disclosure, and the safety standards that apply to long-context agents that act on a user’s behalf.

Alibaba: the open-weight giant

Alibaba Group is the largest of the three by revenue, and its Qwen family of models is among the most widely used open-weight systems in the world. Qwen models power enterprise tools, developer products and consumer apps across dozens of countries. We covered the image-generation side of the family in our guide to Qwen-Image-2.1-Turbo, and the way Qwen is reaching Apple users in China in our article on Apple Intelligence in China.

Alibaba is also the company most exposed to the China AI law across the board. It trains models, runs one of China’s largest cloud businesses, and distributes AI through consumer and enterprise products. Data, computing power, algorithms, development, deployment and use — Alibaba touches every stage the law names, in several different businesses at once. For a company of that size, the transition periods and sector guidance Wang promised will matter as much as the China AI law itself.

Beyond the three named companies, the exact scope of application and the specific obligations for each company will be set by future decrees and detailed criteria. Other Chinese labs — Baidu, ByteDance, Tencent, Z.ai and the newer frontier labs — will almost certainly fall inside the regime as well. The November announcement should be read as the frame, not the finished picture.

What Developers Will Have to Do

The most concrete part of the 3 November announcement is also the most demanding for companies: the disclosure and reporting duties. According to the NDRC transcript, China will push AI developers and operators to strengthen risk management, disclose safety risks promptly, and proactively report safety incidents and vulnerability information.

Each of those three phrases carries real weight. Strengthening risk management means building the internal processes to find problems before they reach users — testing, red-teaming, and documented safety reviews. Disclosing safety risks promptly means telling regulators, and in some cases users, about known risks rather than waiting for them to surface in the press. Proactively reporting incidents and vulnerabilities means that when something goes wrong — a model leaks personal data, a vulnerability is found in a deployed system, an agent takes an unauthorised action — the company reports it itself, quickly, instead of waiting to be caught.

China AI law — developers will have to disclose safety risks and report AI incidents under the new rules

This is a significant shift in emphasis for the China AI law. China’s existing rules are mostly gatekeeping rules: file your algorithm, pass a security assessment, label your generated content, and then you may launch. The new duties are ongoing duties. They apply for as long as the model is in use, not just at the moment of launch, and they assume that problems will keep appearing after deployment.

The model for this kind of duty already exists in other fields. Cybersecurity laws in many countries require companies to report data breaches within a set window. Financial regulators require banks to report operational incidents. What the China AI law would do is bring that same logic to AI — treating a serious model failure as an event that must be reported, not a private embarrassment to be managed quietly.

There are open questions here, and they are large ones. The transcript does not say whether disclosure and incident reporting will be mandatory or voluntary, what counts as a reportable incident, how quickly a report must be filed, or what happens to a company that fails to report. Those details will decide whether the duties are a genuine accountability mechanism or a paper exercise. They will be set in the decrees and detailed criteria that follow the law itself.

One more point deserves emphasis. The duties apply to both developers and operators — the labs that build models and the companies that run them as services. That distinction matters for open-weight models. If a Chinese lab publishes weights openly and a company in another country deploys them, the operator duties may fall on the deployer, while the developer duties stay with the lab. The China AI law will need to draw that line carefully, and how it draws it will shape the open-weight ecosystem far beyond China.

Transition Periods, Sector Guidance and Sandboxes

The NDRC’s announcement was not all stick. Wang also described a set of measures designed to make compliance cheaper and more predictable — and for smaller companies, these may matter more than the headline duties.

First, transition periods. The transcript says regulators will set reasonable transition periods, though no duration was given. A transition period is the gap between a rule being published and the date it becomes enforceable. For a comprehensive law, that gap is where companies rebuild their processes, retrain their teams, and adjust their products. The length of the gap — months or years — will determine how disruptive the law is in practice.

Second, sector-specific guidance. AI is not one industry. A model used in a hospital faces different risks from a model used to write marketing copy, and a model that trades stocks faces different risks from one that tutors schoolchildren. Sector guidance means the general law will be translated into practical checklists for each industry, so a healthcare company and a fintech company are not left guessing what “safety standards” mean for them.

Third, evaluation, consulting and certification services. Wang said regulators will build these services to cut firms’ compliance costs. In practice, that means third-party bodies that can test a model, certify that it meets a standard, and advise a company on what it still needs to fix. For a small startup, a certification service can be the difference between compliance being affordable and compliance being a reason to stop building.

Fourth, regulatory sandboxes. A sandbox is a controlled environment where a new AI application can be tested with real users under regulator supervision, before it is deployed broadly. The idea is to let innovation happen without letting an untested system loose on the public. Sandboxes are already used in finance in many countries, and China has used them for fintech. Applying the same idea to AI — especially to agents that take actions on a user’s behalf — is a logical next step.

Taken together, these four measures describe a regulator that wants the China AI law to be a framework companies can actually live inside, not just a list of prohibitions. Whether that ambition survives contact with the drafting process is the question to watch as the China AI law takes shape. Transition periods can be short, sector guidance can be vague, and sandboxes can be hard to enter. The announcement promises the architecture; the details will decide the experience.

How China Got Here: The Rulebook So Far

The China AI law is not starting from zero. China has been building AI rules piece by piece since 2022, and it did so faster than almost any other country. The new comprehensive China AI law is best understood as the roof on a building whose floors are already in place. Here is how those floors were laid.

March 2022: recommendation algorithms

The first major instrument was the Provisions on the Administration of Algorithm Recommendation in Internet Information Services, effective 1 March 2022. It created the algorithm filing registry — the system under which recommendation algorithms with public-opinion properties or social-mobilisation capacity must be filed with regulators. It also gave users opt-outs and transparency rights over the recommendations they receive. This was the first time a major country had put a filing system around the algorithms that decide what people see.

January 2023: deep synthesis

The Provisions on the Administration of Deep Synthesis of Internet Information Services took effect on 10 January 2023. Deep synthesis covers technologies that generate or edit text, voice, images, video, digital people and virtual scenes — in other words, deepfakes and synthetic media. The rules required labelling and consent before recreating a real person’s face or voice, and extended filing duties to deep-synthesis services. China was regulating synthetic media while most other countries were still debating it.

August 2023: generative AI services

The Interim Measures for the Administration of Generative Artificial Intelligence Services, effective 15 August 2023, are the core rule for the chatbot era. They cover generative AI services provided to the public in mainland China, and they require lawful training data, content moderation with provider responsibility, protection of personal information, and adherence to content rules. Crucially, a service with public-opinion attributes must pass a security assessment and complete algorithm filing before it can launch. Filing is a real gate: by 31 August 2026, 1,112 generative AI services had completed CAC filing, and 731 apps or features built on them had registered. No filing, no public launch.

September 2025: content labelling

The Measures for Labelling AI-Generated Synthetic Content took effect on 1 September 2025, paired with the mandatory national standard GB 45438-2025. Every piece of AI-generated content must carry both a visible label and hidden metadata marking it as synthetic. Distribution platforms and app stores share the duty. China moved first on mandatory labelling — more than a year before the EU’s equivalent transparency duties took full effect.

July 2026: human-like AI companions

The Interim Measures for the Administration of Anthropomorphic AI Interaction Services took effect on 15 July 2026. They cover continuous emotional-interaction services that simulate human traits and communication — the AI companion apps that have become a large consumer category. This is the newest floor of the building, and it shows the pattern: each time a new class of AI application reaches the public, a dedicated instrument follows.

Alongside the binding rules sits the advisory layer. The national cybersecurity standards committee, TC260, released version 3.0 of its AI Safety Governance Framework on 14 September 2026, under the guidance of the Cyberspace Administration of China. The framework is advisory, not binding, but it signals where the binding rules are heading. Version 1.0 appeared in September 2024, so the framework has been revised twice in two years — a pace that matches the speed of the technology itself.

The pattern is clear. China regulates vertically: each new application class gets its own instrument, enforced by the Cyberspace Administration of China (CAC) together with sectoral ministries. The China AI law would change the shape of that system without replacing it. Instead of a new instrument for each new application, there would be one comprehensive statute underneath them all — setting the lifecycle duties, the safety standards and the ethical guidelines that every instrument must respect.

The 15th Five-Year Plan Connection

The announcement did not come out of nowhere. It came at the first NDRC briefing in a series on the goals of China’s 15th Five-Year Plan, which runs from 2026 to 2030. Five-year plans are the documents that set China’s economic and technological direction, and AI sits near the centre of this one.

China AI law — the announcement came at the first NDRC briefing on China's 15th Five-Year Plan

The legislative groundwork was laid months earlier. On 11 May 2026, the State Council published its 2026 legislative work plan, which for the first time included language about a comprehensive AI law. And on 23 October 2026, the national informatisation plan for 2026 to 2030 called for advancing foundational legislation on AI. The 3 November briefing turned those plan documents into a public commitment with a named owner — the NDRC — and a named direction: speed it up.

The five-year-plan context also explains the timing. The plan period begins in 2026, and the NDRC briefing series is the moment when each part of the plan is explained to the public. Announcing the China AI law at the first briefing in the series signals that AI governance is a first-order priority for the plan, not an afterthought to be handled later.

There is a second timing factor. The global AI race is now a race between regulatory systems as much as between models. The EU AI Act became fully applicable on 2 August 2026. California signed its frontier-model transparency law, SB 53, in 2025, and followed it in September 2026 with laws creating independent AI auditors. South Korea and Vietnam now have AI laws in force. China, which moved first on the individual rules, does not want to be the country without a comprehensive statute while everyone else has one. The NDRC announcement is, in part, a catch-up move.

For the 15th Five-Year Plan, the China AI law also fits a broader economic story. China’s plan emphasises technological self-reliance — domestic chips, domestic models, domestic cloud. A comprehensive AI law is the governance counterpart to that industrial policy: rules that make the domestic AI industry legible, standardised and, in the government’s view, safe enough to scale. The law is not separate from the industrial strategy. It is part of it.

China AI Law vs the EU AI Act vs the US Approach

Three of the world’s largest AI markets now regulate AI in three fundamentally different ways. Understanding the China AI law means understanding how it differs from the other two — because any company that sells AI globally will have to live inside all three at once.

Dimension China European Union United States
Core instrument Comprehensive AI law in preparation; layered binding rules already in force (2022–2026) AI Act, Regulation (EU) 2024/1689, fully applicable since 2 August 2026 No federal AI statute; state laws plus agency rules
Architecture Vertical and lifecycle-based: duties attach to each stage, from data to deployment Horizontal and risk-based: duties attach to the risk tier of the use case Patchwork: sector rules, state laws and voluntary frameworks
Philosophy State control over information and security Rights-based and precautionary Market-led and pro-innovation at federal level
Model-level duties Security assessments and filings before public launch; incident reporting planned Transparency and documentation for general-purpose models; systemic-risk tier Frontier-model disclosure under California SB 53; no federal equivalent
Content transparency Mandatory visible labels and hidden metadata since 1 September 2025 Transparency duties for synthetic content and chat interfaces since August 2026 California AI Transparency Act operative; no federal labelling law
Enforcement Cyberspace Administration of China with sectoral ministries Member-state authorities plus the EU AI Office State attorneys general, FTC, EEOC and sector agencies
Maximum penalties Service rectification, fines, app-store removal; no schedule published for the new law Up to 7% of global turnover or €35 million Varies by state; no federal cap
Comprehensive law Drafted, not enacted; on the State Council legislative plan Enacted; implementation under way None at federal level

The most important difference between the China AI law and the EU AI Act is the axis of regulation. The EU asks: how risky is this use case? A system that decides who gets hired or who gets a loan is high risk, and it faces a conformity assessment before deployment. A system that filters spam is minimal risk, and it faces almost nothing. The China AI law asks a different question: where in the lifecycle is this system, and what duties attach to that stage? Data, compute, algorithms, development, deployment, use — each stage carries its own obligations, regardless of the use case.

The American approach is different again. There is no federal AI statute, and the federal posture in 2025–2026 has been explicitly deregulatory, framing rules as a competitiveness risk and even threatening to pre-empt the states. The real action is at state level: California’s SB 53 requires frontier developers to disclose safety frameworks and report critical incidents, and its September 2026 laws created independent verification organisations and an auditor registry. Colorado repealed and re-enacted its AI Act, with obligations from 1 January 2027. The result is a patchwork that a company must navigate state by state.

For a global AI company, the practical consequence of the China AI law is that the same model faces three different compliance logics. In Europe, it is graded by the risk of its use. In China, it is tracked through its lifecycle. In the United States, it is governed by whichever state’s law applies to its customers. The China AI law adds a fourth consideration: a company that trains in China, deploys in China, or serves Chinese users will need lifecycle documentation that neither the EU nor the US currently demands in the same form.

One area where the regimes converge is content labelling. China has required labels since September 2025, the EU’s Article 50 transparency duties apply from August 2026, and California’s transparency law is now operative. Whatever else divides them, the world’s three largest markets now agree on one thing: AI-generated content must be marked as AI-generated. That convergence is worth noting, because it suggests where the next convergence may come — incident reporting, which California already requires and China is now planning.

What It Means for the Open-Weight Models You Use

The China AI law matters to readers outside China for one simple reason: the models it will govern are the models many of them already use. Kimi, Qwen and DeepSeek are not niche products. They are among the most downloaded open-weight models in the world, and they power tools, apps and workflows far beyond China’s borders. A law that changes how those models are built will change what they can do.

Release speed may slow — or become more predictable

The most likely near-term effect is on release cadence. Today, a Chinese lab can train a model, publish its weights, and ship an API in a matter of weeks. Under a lifecycle regime with safety standards, incident reporting and pre-deployment assessments, each release would carry more process. That could mean slower releases — or it could mean more predictable ones, with documented testing that gives enterprise users more confidence to deploy. Both outcomes are possible, and the details of the law will decide which one dominates.

Open weights face the hardest question

The hardest question the China AI law must answer is who carries the compliance duty when weights are published openly. A lab that releases a model under an open licence cannot control what thousands of downstream developers do with it. If the law places the duty on the lab, open releases become riskier and rarer. If it places the duty on the operator — the company that deploys the model as a service — then open releases can continue, but every deployer inherits a compliance burden. The NDRC’s language covers both developers and operators, which suggests the duty will be shared. How it is shared is the detail that will shape the open-weight ecosystem for years.

Documentation may improve for everyone

There is a genuine upside for users. Lifecycle regulation forces labs to document what their models were trained on, how they were tested, and what risks were found. That documentation — model cards, safety evaluations, incident reports — is exactly what enterprise buyers and independent researchers have been asking for. If the China AI law makes that documentation standard, users everywhere benefit, including users who never touch a Chinese product directly, because the norms spread across the industry.

Compute and data rules could reshape training

The law’s coverage of data and computing power is the sleeper issue. Rules on training data — where it comes from, whether it was lawfully obtained, how personal information is handled — can change what a model learns. Rules on computing power can change where and how models are trained. For labs that train on large clusters, even small changes in the rules around data and compute can have large effects on cost and capability. This is the part of the China AI law that technical readers should watch most closely, because it operates upstream of everything else.

What It Means for US Readers and Businesses

If you are reading this in the United States, the China AI law may feel distant. It is a Chinese law, announced by a Chinese agency, aimed at Chinese companies. But there are four concrete ways it reaches American readers and businesses.

First, the tools you use. If your workflow includes Kimi, Qwen, DeepSeek or any product built on them, the law will shape those tools’ capabilities, release speed and documentation. A model that is slower to release but better documented is a different product from one that ships fast with no paperwork. Your vendor choices may change as the law takes effect.

Second, the compliance map. If your company sells AI products globally, or serves customers in China, you now face a third regulatory logic alongside the EU’s risk tiers and the American state patchwork. Lifecycle documentation — what data you used, what compute you ran, what testing you did, what incidents you reported — is a capability your compliance team will need to build. Companies that build it once, for China, will find it reusable in California and Europe, where incident reporting and documentation duties are converging in the same direction.

Third, the competitive picture. The law’s cost falls hardest on smaller labs, which have the least compliance capacity. If transition periods are short and certification is expensive, the Chinese AI market could consolidate around the largest players — Alibaba, ByteDance, Tencent — while smaller labs slow down or merge. That consolidation would change who your suppliers and competitors are, and what models are available to build on.

Fourth, the precedent. China’s 2023 generative-AI rules were followed, in spirit, by labelling rules elsewhere. Its 2025 labelling measures preceded the EU’s transparency duties. When China adopts a regulatory idea, that idea tends to travel. Lifecycle regulation, incident reporting and sector sandboxes are ideas worth watching precisely because they may appear in other jurisdictions next — including American states, where the legislative energy on AI is currently highest.

None of this is a reason to panic, and none of it is a reason to ignore the story. The practical move for a US business is simple: map which of your vendors and models touch China, ask those vendors how they are preparing, and start keeping the lifecycle records — data sources, testing results, incident logs — that every major regime is now converging on. That preparation is useful regardless of what the final China AI law says.

What Is Still Unknown — and What to Watch

Honesty requires a long list of unknowns. The China AI law is an announcement, not a statute, and the gap between the two is where most of the real story will be decided. Here is what to watch, in order of importance.

China AI law — many details, including penalties and the enactment date, are still unknown

No enactment date. No timetable for the law has been published. Comprehensive legislation in China can take years from announcement to enactment, and the NDRC has not said whether this law is months away or years away. The May 2026 State Council legislative plan included the law for the first time, which suggests it is a genuine priority — but a priority is not a deadline.

No draft text. Without a draft, nobody outside the drafting process knows which duties will be mandatory, which will be voluntary, and which will be left to later decrees. The China AI law announcement’s language — “push developers to disclose”, “set reasonable transition periods” — is the language of direction, not the language of obligation. The draft will be the moment the story becomes concrete.

No penalty structure. The existing Chinese rules carry real penalties: service rectification, fines and app-store removal. The new China AI law’s penalties have not been described. Whether it will carry fines comparable to the EU’s 7% of global turnover, or something narrower, is unknown — and it matters enormously to the companies that must comply.

No lead drafting body named. The NDRC made the announcement, but the NDRC is an economic planner, not a technology regulator. The Cyberspace Administration of China enforces the existing AI rules. The China AI law could be drafted by the CAC, by the NDRC, by the Ministry of Industry and Information Technology, or by a combination of agencies. Who drafts it will shape what it emphasises.

The open-weight question is unanswered. As discussed above, the division of duties between a lab that publishes weights and a company that deploys them is the single most consequential detail for the global open-weight ecosystem. Nothing in the announcement resolves it.

The named-companies list is not final. Bloomberg named DeepSeek, Moonshot and Alibaba. Wang did not name anyone. The final scope will be set by future decrees and detailed criteria, and it may be broader — or narrower — than the three names in the press coverage.

Enforcement capacity is untested. Lifecycle regulation requires regulators who can evaluate models, audit training data, review incident reports and run sandboxes. That is a large institutional build, and it is not clear how quickly China’s regulators can staff it. A law that outruns its enforcers is a law that is enforced unevenly.

The milestones to watch are specific: the publication of a draft law for comment, the naming of a lead drafting body, the release of the first sector guidance, and the opening of the first regulatory sandboxes. Each of those will be a news event in its own right, and this article will be updated as they happen.

China AI Law Timeline

  • 1 March 2022 — Algorithm Recommendation Provisions take effect; the algorithm filing registry is created.
  • 10 January 2023 — Deep Synthesis Provisions take effect; synthetic media rules and labelling begin.
  • 15 August 2023 — Interim Measures for Generative AI Services take effect; security assessments and filing become the gate for public launch.
  • 9 September 2024 — TC260 releases version 1.0 of the AI Safety Governance Framework.
  • 1 September 2025 — AI content-labelling measures and the mandatory standard GB 45438-2025 take effect.
  • 11 May 2026 — The State Council’s 2026 legislative work plan includes a comprehensive AI law for the first time.
  • 15 July 2026 — Anthropomorphic AI Interaction Measures take effect, covering AI companions.
  • 14 September 2026 — TC260 releases version 3.0 of the AI Safety Governance Framework.
  • 23 October 2026 — The national informatisation plan for 2026–2030 calls for advancing foundational AI legislation.
  • 3 November 2026 — The NDRC announces it will fast-track a basic, comprehensive China AI law covering the full lifecycle, at the first briefing on the 15th Five-Year Plan.
  • Next — Draft law publication, lead drafting body, sector guidance and sandboxes: all still to come.

Frequently Asked Questions

Is the China AI law already in force?

No. The China AI law is in the legislative planning phase. On 3 November 2026 the NDRC said it will accelerate work on a basic, comprehensive law, but no draft text has been published and no enactment date has been set. China’s existing AI rules — the 2022 algorithm provisions, the 2023 generative AI measures, the 2025 labelling measures and the 2026 companion-AI measures — remain the binding law today.

Which companies will the China AI law cover?

Bloomberg reported that the rules would reach DeepSeek, Moonshot AI and Alibaba Group. Those names come from Bloomberg’s reporting, not from the NDRC briefing itself, and the exact scope will be set by future decrees and detailed criteria. Other major Chinese labs, including Baidu, ByteDance and Tencent, are likely to fall inside the regime as well.

What does “full lifecycle” mean?

It means the law will cover every stage of an AI system: the data, computing power and algorithms that go into it; the development and testing stage; deployment to users; and ongoing use, including incident reporting and risk disclosure. That is broader than regulating only the finished product or only the moment of launch.

How is the China AI law different from the EU AI Act?

The EU AI Act sorts systems by the risk of their use case and attaches duties to each risk tier. The China AI law sorts by lifecycle stage and attaches duties to each stage. The EU law is enacted and fully applicable since 2 August 2026, with fines up to 7% of global turnover. The Chinese law is still being drafted, and its penalties have not been published.

Will the China AI law affect me if I live in the US?

Yes, indirectly. If you use Kimi, Qwen, DeepSeek or products built on them, the law will shape those models’ release speed, documentation and capabilities. If your company sells AI globally, the law adds a third regulatory logic to the EU and US regimes you already navigate. And because Chinese regulatory ideas have travelled before — labelling rules being the clearest example — lifecycle duties and incident reporting may appear in other jurisdictions next.

What should companies do now?

Start keeping the lifecycle records the China AI law will demand: where your training data came from, what testing you ran, what risks you found, and what incidents occurred. Map which of your vendors and models touch China. Ask those vendors how they are preparing for the China AI law. And watch for the draft law, the sector guidance and the first sandboxes — those are the milestones that will turn this announcement into obligations you can plan around.

The Bottom Line

The China AI law is the most consequential AI regulation announcement of November 2026 — not because a law has passed, but because of what it signals. China, which already regulates AI more actively than any other country through its layered rules, is now building the roof: a single comprehensive statute covering the full lifecycle, from data and computing power to deployment and use.

The announcement names a direction, not a destination. No draft exists, no date is set, no penalties are published, and the hardest questions — above all, who carries the duty when model weights are published openly — remain unanswered. But the direction is clear, the named companies are the ones that matter most to the global market, and the supporting measures — transition periods, sector guidance, sandboxes — show a regulator thinking about how companies will actually live inside the law.

For users of Kimi, Qwen and DeepSeek, the practical takeaway is to watch the draft. For businesses selling AI globally, it is to start building lifecycle documentation now. And for everyone following AI regulation, it is to recognise that the China AI law means the world’s three largest AI markets now regulate in three different ways — by risk, by lifecycle, and by patchwork — and that the companies which learn to live inside all three will be the ones that keep shipping.

Sources: the NDRC briefing of 3 November 2026 as reported by ForkLog and FourWeekMBA; the EU–China comparison draws on Caybles Research. Company names in scope are attributed to Bloomberg’s reporting, as the NDRC transcript itself names no company.

Leave a Comment

Your email address will not be published. Required fields are marked *