--- title: "China's ethics-review rule: build the committee before you build the model" date: 2026-10-22 category: Policy & Governance site: NeuroAI canonical: https://neuroai.site/a/na-policy-tech-ethics-review-measures language: en --- # China's ethics-review rule: build the committee before you build the model > A 2023 trial measure makes Chinese universities, hospitals, and AI firms form an ethics committee and send the riskiest projects to national review. A researcher trains a model that can shape what large audiences read. Another builds a system that decides, without a human in the loop, whether a machine applies the brakes. Both are now asked to pause and explain themselves first. China's answer is not a ban and not a model-approval gate. It is an internal review board — the kind of committee long required for medical trials, now extended to certain kinds of artificial intelligence. The rule is quiet, procedural, and easy to miss. Its reach, once you read the fine print, is not small. ## What the measure does The *Trial Measures for Ethical Review of Science and Technology* (科技伦理审查办法(试行)) took effect on December 1, 2023. It tells universities, research institutes, medical institutions, and companies that they are the ones responsible for checking the ethics of their own work — and for building the committee that does the checking. The logic is preventive. Before a sensitive project scales, it should pass an internal review that weighs benefits against risks to people, society, and the environment. The state supplies the framework; the institution supplies the committee. ## Who must comply The measure names four kinds of responsibility subjects: higher-education institutions, research institutes, medical and health institutions, and enterprises. Crucially, any unit working in **life sciences, medicine, or artificial intelligence** on ethically sensitive topics is expected to establish a science-and-technology ethics (review) committee. Other organizations with a review need may set one up voluntarily. The committees must be staffed, funded, and — the document insists — allowed to operate independently of the teams whose work they judge. ## The review machinery Review is not a single step. The measure lays out several procedures: - A standard **application and acceptance** step. - A **general procedure** for ordinary cases. - A **simplified procedure** for work whose risk is no higher than minimal, or for minor modifications to an already-approved plan. - An **expert review procedure** for items on a special high-risk list. - An **emergency procedure** for urgent cases. Approved work is not forgotten. Committees are told to conduct follow-up review, generally at intervals no longer than twelve months, and they can pause or terminate a project if things go wrong. ## The red-list that triggers national review The measure attaches a list of activities that must be sent up for expert review rather than handled locally. Several items land squarely on AI: - Research and development of **human-machine fusion systems** that strongly affect human behavior, emotion, or life and health. - Research and development of **algorithmic models, applications, and systems with the capacity to mobilize public opinion or steer social consciousness** — a category that comfortably covers large models (大模型) built for persuasion, recommendation, or mass communication. - Research and development of **highly autonomous automated decision systems** used in scenarios carrying safety or personal-health risk. In other words, the state keeps a short list of the projects it most wants to see itself, while letting institutions handle the long tail. ## Why AI is in the frame The inclusion of AI is consistent with a broader shift: algorithms that influence people at scale are no longer treated as pure engineering. A model that can shape attention, or a system that makes consequential decisions without a human, is now seen as carrying societal risk that warrants a paper trail. For foreign researchers collaborating with Chinese institutions, the practical effect is that joint AI projects touching these categories will pass through the same committee gate — even if the model is trained elsewhere. ## How it differs from an AI law This is a governance layer, not a licensing regime. It does not pre-approve models for public release, and it is not the long-discussed comprehensive AI statute. It is a standing obligation attached to institutions, focused on the research and development phase rather than the shipped product. That distinction matters. A company can still release a model; what changes is the internal obligation to have asked, and documented, the hard questions first. ## Honest limitations This article explains a procedural measure, not an audit of how strictly it is enforced. The list of review-triggering activities is framed in broad language — "steer social consciousness," "highly autonomous" — whose exact boundaries are not fully defined in the text and may be clarified by later guidance. We have not claimed that every AI project in China now requires review; only those meeting the sensitivity and category tests do. Compliance depth likely varies by institution, and no public consolidated enforcement report was verified for this writing. ## What readers can do now - **For research institutions:** confirm whether your AI work falls under the life-science, medicine, or artificial-intelligence sensitive categories, and stand up or join a qualifying ethics committee before starting such projects. - **For international collaborators:** treat the committee review as a real gate in joint AI proposals, and budget time for it in project plans. - **For compliance teams:** track the dynamic red-list and any sector-specific rules, since the measure explicitly says the list can be adjusted. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-policy-tech-ethics-review-measures Free to quote with attribution and a link to the original.