China's Ten Departments Jointly Implement AI "Review-Before-Creation": What the World's First Mandatory Pre-Research Ethics Review Mechanism Means

On April 2, 2026, the Ministry of Industry and Information Technology, together with ten departments including the National Development and Reform Commission, the Ministry of Education, the Ministry of Science and Technology, the National Health Commission, the Cyberspace Administration of China, and the Chinese Academy of Sciences, officially issued the AI Science and Technology Ethics Review and Service Measures (Trial) (MIIT Joint Science [2026] No. 75), effective the same day. On August 26, Vice Minister of Industry and Information Technology Xin Guobin disclosed at a State Council Information Office press conference that China has successfully developed nearly 200 key AI standards and is organizing multiple cities to carry out AI ethics review and service practices. The convergence of these two milestones marks China's AI governance transitioning from "policy release" to "compliance enforcement."

Why has the review checkpoint been moved to "before R&D initiation"

According to Article 12 of the Measures, for any AI science and technology activity that may pose ethical risks, the person in charge must submit an application to their institution's ethics committee or an accredited service center before commencement, rather than waiting until after the product is completed to go through market approval. Application materials must include a complete description of algorithmic mechanisms, data sources, testing and evaluation methods, and intended application domains, along with a science and technology ethics risk assessment report and prevention and control plans.

The Measures require higher education institutions, research institutes, medical and health institutions, and enterprises engaged in AI activities to establish AI science and technology ethics committees, which must have multidisciplinary backgrounds covering AI technology, applications, ethics, and law, with no fewer than five committee members attending meetings. Institutions unable to establish their own committees independently may entrust professional "AI Science and Technology Ethics Review and Service Centers" accredited by competent authorities to conduct reviews, but such centers may not provide both review and re-review services for the same project.

For high-risk scenarios, the Measures specifically list three categories that must undergo expert re-review by local or competent authorities: human-machine integration technologies, public opinion guidance algorithms, and highly autonomous decision-making systems in high-risk scenarios. What these three categories share is that once ethical control fails, the impact is broad and difficult to reverse.

Compared with the EU AI Act: Both talk about risk, but diverge on timing

The core logic of the EU AI Act is pre-market compliance — high-risk AI systems must complete conformity assessment before entering the market or being put into use, pass third-party review, and remain subject to ongoing market surveillance. The maximum fine is 7% of global turnover or €35 million, whichever is higher.

Both China and the EU adopt risk-based tiered review logic, but the timing of review constitutes a fundamental divergence: the EU regulates "before market entry," while China regulates "before research." The former allows iterative exploration during the development process and unified verification after the product takes shape; the latter requires a full cycle of ethical assessment before the design plan is finalized and the algorithmic framework is built.

This timing difference is particularly critical for large model development. Pre-training itself is a highly uncertain exploratory process, making it difficult to accurately predict the final ethical impact of training data at the design stage. The "pre-review system" requires developers to structure uncertainty into reviewable documentation before exploration even begins.

The gray areas in implementation details

The Measures are classified as an administrative normative document rather than regulations or departmental rules with stronger binding force. In its analysis, law firm MMLC Group pointed out that according to the document's wording, each institution should "independently determine" based on its own business type and risk characteristics whether it needs to establish or improve an ethics review mechanism.

As of August 2026, the Measures have not yet clarified which risk levels would lead to project rejection, nor have they specified concrete rectification measures or pass/fail determination criteria. International legal compliance platform Reg Intel noted in a comparative study that the practical challenge facing multinational enterprises is that China's pre-research review requirements have no direct mapping to existing compliance frameworks in the EU or the US, and companies must build a separate compliance architecture for the Chinese market.

What it means for multinational AI companies

For multinational AI companies operating in China or planning to enter the Chinese market, the compliance complexity brought by this mechanism is multi-layered.

First is the issue of fragmented compliance interfaces. Unlike the EU's unified authorization through a single law, China's AI regulation adopts a vertical, scenario-based stacked structure: algorithmic recommendation, deep synthesis, generative AI, content labeling, and ethics review each have their own independent rules under different competent authorities. A single large model product may simultaneously trigger three sets of regulatory requirements: algorithmic recommendation, generative AI, and ethics review.

Second is the restructuring of cost structures. Pre-research review means compliance personnel must be involved from the project initiation stage, rather than dealing with requirements centrally before product launch. The Measures also allow small and micro enterprises to outsource review to accredited institutions, which has objectively given rise to a market for AI ethics review outsourcing services.

Furthermore, China's competitors in the AI field — leading companies such as Alibaba and Baidu — had already established internal science and technology ethics review committees under regulatory pressure as early as 2022, and according to the South China Morning Post, this system has become the new compliance norm for leading domestic companies. Foreign enterprises face the pressure of catching up with existing compliance practices.

Deeper signal: A watershed in governance philosophy

If the EU AI Act represents "rule-driven compliance civilization," China's "pre-review system" reflects a different governance logic: embedding regulatory power into the R&D process itself, rather than standing at the end point of the product waiting for verification.

This logic has its inherent rationality. The ethical risks of large models are often "encoded in" during the training phase — biases in training data and the design of objective functions are difficult to correct after product release. In this sense, the "pre-review system" attempts to intervene at a point that truly reaches the root of risk earlier than "pre-market review" does.

But the cost is equally significant: in the early stages of R&D when algorithmic mechanisms have not yet been validated and data configurations are still being adjusted, it mandates the submission of an "ethics impact assessment for intended application domains," requiring the production of certainty documents amid profound uncertainty.

The implementation of nearly 200 AI national standards provides partial technical anchors for this review mechanism; but many standards, few enforcement details, and ambiguous pass/fail determinations remain the largest sources of uncertainty in the current system. Enforcement practice over the next 12 to 18 months — especially the first cases of rejection or rectification — will tell us more about the actual boundaries of this mechanism than the document itself.