Zhejiang University Proposes AI Law Built Around Matching Systems to Their Tasks

A 99-Article Proposal for Deciding What AI Should Be Allowed to Do

Zhejiang University researchers have released an expert proposal for an artificial intelligence law containing eight chapters and 99 articles, with a central idea: the authority given to an AI system should match its reliability and the consequences of the task it performs.

The Zhejiang University Version 1.0 proposal was presented at a September 24 launch event. The university's official event notice scheduled the gathering for September 24, 2026, from 2 p.m. to 5:30 p.m. at its Zijingang campus. Its agenda placed industry development alongside rights protection and accountability, reflecting the two demands the proposal seeks to address.

For people affected by AI, the practical concerns are immediate. Can a model operator use an illustrator's work for training without permission? Who carries responsibility when a language model gives a harmful, incorrect answer? How can someone preserve evidence and seek a remedy after a video using their face without permission spreads online?

Businesses face another set of concerns: overlapping compliance requirements, burdensome procedures before products can be used, and costs that smaller companies may struggle to meet. The researchers say their approach aims to protect rights while responding to those pressures.

The document is an academic proposal for discussion and further revision. Its most developed idea is a method for judging AI in a particular setting, connecting what a system can do, how reliably it can operate, and what is at stake for the people affected.

How the Draft Took Shape

The drafting team spent nearly six months developing the proposal and completed five rounds of concentrated revision. An initial draft of 174 articles became 99, a reduction of 75 articles, or about 43%. That shows how much the text was condensed during development, although the article count alone says little about the demands individual provisions would place on users or companies.

Researchers visited more than 20 AI businesses in Hangzhou as well as leading internet companies. Their work also drew on discussions at the Zhijiang legislative forum, exchanges with relevant ministries and commissions, and several rounds of expert review. Research proceeded alongside work on two local draft regulations concerning Hangzhou's AI industry and its development as a city of innovation.

The organizers were Zhejiang University Guanghua Law School, the university's Digital Law and Governance Research Institute, the Zhejiang Legislative Research Institute and the Zhejiang University Legislative Research Institute. The official agenda assigned the proposal's release and introduction to Li Wenlong, a researcher at Guanghua Law School.

The development and launch sequence was:

• Over nearly six months, the team conducted business visits, consultations and five rounds of revision.
• On September 24, the proposal was presented at Zhejiang University, with the official notice dating the event to 2026.
• After the meeting, the team planned to incorporate feedback and develop further versions.

Three Questions Behind the Proposed Rules

Sun Xiaoxia, director of Zhejiang University's Digital Law and Governance Research Institute, describes the framework through three linked elements: execution capability, cognitive maturity and the nature and impact of a task. In simpler terms, the questions are what an AI system can do, how dependable it is, and how much its decisions matter.

Sun explained his view of the source of AI risk in Chinese:

We believe that the risks of artificial intelligence do not come from how powerful machines are, but from a mismatch among capability, cognition and tasks.

A system might be able to perform a long series of operations without a person intervening, yet struggle to recognize an unusual situation. Assigning that system a task involving personal safety would create a dangerous gap between its freedom to act and its ability to judge when something is wrong.

The proposal therefore looks beyond a single label attached to a technology. A regulator would examine the particular version of a system, its intended use and the environment in which it operates. The same underlying technology could warrant different controls when used for different tasks.

Capability and Reliability Would Be Assessed Separately

The proposed capability scale has four levels: assistance, limited automation, high automation and autonomous execution. Assessment would consider how independently a system acts, how widely its decisions reach, which tools it can access, how long it can operate continuously, and whether it can change a digital or physical environment.

Tool access matters because an AI system that merely suggests an action has a different reach from one authorized to carry it out. Permission to operate other software or equipment can turn a mistaken answer into a sequence of real actions.

Cognitive maturity would also have four levels, described as basic, general, higher and highly controlled. The assessment would examine accuracy, consistency, the ability to recognize unusual inputs and correct errors, cybersecurity, traceability, and the conditions under which a person can take control.

Another factor is confidence calibration. In plain language, that means checking whether a system's expressed confidence matches how often it is right. A system that sounds certain when it is frequently wrong can be especially difficult for users to supervise.

The framework would also examine whether important outputs can be explained and whether a system can remain within defined boundaries. Separating capability from reliability allows a regulator to identify a system that is powerful enough to act independently but insufficiently dependable for a sensitive assignment.

Which Tasks Would Receive Greater Scrutiny?

The proposal divides tasks into three impact levels: low, general and high. The assessment would consider the importance of affected rights, the likelihood and severity of harm, the number of people affected, how long the effects last, whether a result can be reversed, and whether people can obtain a remedy.

Vulnerability would also matter. A decision affecting someone with limited ability to challenge it can require closer attention than a similar decision affecting someone with ready access to help.

Tasks involving life and health, personal liberty, human dignity, equal opportunity, major property interests, education, employment, social security, justice, administrative enforcement or public safety would be presumed to have high impact. Tasks capable of substantially affecting large numbers of people would receive the same presumption.

The proposal also states that prohibited AI applications must not be developed, deployed or used, regardless of their task impact classification. The descriptions released with the proposal do not provide the detailed list of prohibited applications or the precise legal answers to each copyright, incorrect output and face substitution dispute raised at the outset.

A Medical Example Explains the Matching Principle

Sun used medical imaging and pathology analysis to illustrate why a system's assignment matters as much as its technical strength. AI can process large volumes of images and medical data quickly, while its reasoning may remain limited when faced with difficult or unusual conditions.

A suitable assignment would take both features into account. Giving the system complete responsibility for a difficult diagnosis, while placing excessive trust in its judgment, would create the mismatch the framework seeks to address.

Sun described the aim in Chinese:

Requiring a match is different from simply imposing restrictions; it seeks the best match among capability, cognition and tasks.

A system that performs poorly on a demanding task could, in his view, be reassigned to a less complex one instead of being discarded. The practical choice is therefore wider than allowing unrestricted use or preventing use altogether. Authorities and operators could adjust the task, the system's permissions or the degree of human involvement.

For a hospital, that approach would focus attention on the exact work assigned to AI and the safeguards around it. For a patient, the central concern would be whether the system's role is appropriate to the seriousness of the decision.

Approval Would Need to Change as AI Changes

The proposed governance process covers six stages: research and development, testing, deployment, operation, iteration and withdrawal. Classification would continue throughout those stages because a system's capabilities and circumstances can change after it enters use.

A model update might increase the range of actions a system can perform. Expansion into a new setting might expose more people or more sensitive interests to its decisions. New data can also change how a model behaves, including reducing reliability in ways that were not apparent during earlier testing.

The proposal calls for reassessment when the system or task changes. A judgment would be tied to a concrete version, purpose and operating environment, giving operators a reason to revisit their obligations when any of those elements changes.

Sun identifies eight areas where the matching principle could guide legal rules: classification, allocation of duties, assessments, identification of high impact uses, regulatory intensity, risk controls, legal responsibility and adjustment over time. Connecting those decisions through one framework is the proposal's main attempt to make AI governance more coherent.

Rights Protection and Business Duties Meet in the Details

The questions about illustrators, incorrect answers and fabricated videos show why accountability matters beyond technical testing. A person seeking compensation or removal of harmful content needs a workable route from the harm they experienced to the parties responsible for it.

The proposal places legal responsibility within its matching framework. That would make the suitability of a system for its assigned task relevant to the discussion of duties and accountability. The relationship between assessment results and existing rules on civil liability, however, is among the issues Sun says require further detailed work.

Researchers also need to specify how cognitive maturity should be tested and how large a gap between capability, reliability and task demands should count as an unsuitable match. Those choices will determine how consistently companies can apply the framework and how effectively regulators can check their judgments.

The launch agenda brought those competing practical needs into the same programme. Its industry session listed legal specialists from Meituan and Douyin alongside academic and technical participants. Its rights and accountability session listed an Alibaba compliance director, lawyers and university scholars. Their inclusion shows the range of perspectives the organizers sought for the discussion; the agenda gives no basis for assigning those participants a shared position on the draft.

For smaller businesses, the concrete stakes are the number of assessments required, the evidence they must produce and how obligations change as a product develops. For people affected by AI, the stakes are whether those obligations help prevent harm and support a usable remedy when something goes wrong.

The research team plans to revise the proposal using feedback from the meeting. The central work now is to turn the matching principle into tests, thresholds and responsibility rules that people can apply in actual disputes and product decisions.

Key Points

• Zhejiang University's expert AI law proposal contains eight chapters and 99 articles, condensed from an initial 174 articles.
• The framework links a system's execution capability, cognitive reliability and assigned task.
• Health, liberty, employment, education and other sensitive tasks would be presumed to have high impact.
• Assessments would cover the full development and use cycle and change when a system or its task changes.
• The academic proposal will undergo further revision, with testing methods, thresholds and liability rules requiring more detailed work.