- Read
~/.claude/plugins/config/claude-for-legal-zh/ai-governance-legal/CLAUDE.md→ contract review position, acceptable risk thresholds, red-line clauses. - Run the following workflow.
- Check AI-specific risks item by item—training data → liability → model changes → compliance propagation.
- Output: risk summary + red-line flags + negotiating position (approved/conditional/blocked).
/ai-governance-legal:vendor-ai-review
[paste the vendor AI terms]
AI Supplier Contract Review
Matter Context
Matter context. Check the practice-level CLAUDE.md for ## 事务工作区. If 已启用 is ✗, skip the rest of this section. If enabled and there is no active matter, ask which matter it belongs to. Load the active matter's matter.md. Unless 跨事务上下文 is 开, never read files from other matters.
Purpose
AI supplier contracts introduce risk dimensions that traditional technology contracts do not have—whether the supplier uses your data to train models, whether you will be notified when the model changes, who bears the risk if the AI generates infringing content, and whether the supplier has completed the statutory algorithm filing and security assessment (Article 17 of the Administrative Measures for Generative AI Services [法条原文]). This skill systematically reviews these risks.
Load Current State
Read ~/.claude/plugins/config/claude-for-legal-zh/ai-governance-legal/CLAUDE.md：
## 合同审查配置— company position, risk appetite, red lines## 监管注册表— applicable regulatory framework## 已批准的供应商— existing relationships and previously reviewed clauses
Review Framework
Step 1: Service Characterization
First clarify what the supplier is providing:
| Model delivery method | Description | Key risks |
|---|---|---|
| API interface | Call the model via cloud API | Data transmission security, whether data is recorded for training |
| On-premises deployment | Model deployed on own servers | High security and controllability, but updates and upgrades depend on the supplier |
| SaaS product | Use the supplier's AI-enabled product | Terms of use may be unclear, and data use clauses require particular attention |
| Model licensing/customization | License a foundation model for fine-tuning | IP ownership, compatibility of model updates |
Step 2: Training Data Check
This is the most important part of AI contract review.
Core question chain
- Does the supplier use customer data to train models?
- If so, is the customer informed and consenting?
- Does the training data include personal information or sensitive personal information?
- Can customer data be retrieved after it flows out (the practical feasibility of the "right to be forgotten")?
Review Checklist
| Check item | Ideal state | Risk flag |
|---|---|---|
| Training data clause | Clearly agree not to use customer data for model training, or only with the customer's express written consent | 🔴 The contract is silent, or the clause vaguely claims that the supplier may "use data for service improvement." |
| Personal information training | Do not use data containing personal information for training, or separate consent has been obtained from the individual (Article 23 of the Personal Information Protection Law [法条原文]） |
🔴 Data types are not distinguished, and authorization is one-size-fits-all |
| Guarantee of the legality of training data | The supplier guarantees that its training data comes from lawful sources and does not infringe third-party intellectual property rights (Article 7 of the Measures for the Management of Generative Artificial Intelligence Services [法条原文]） |
🟠 The supplier provides only a "best efforts" guarantee or provides no guarantee |
| Data deletion | After contract termination, the supplier deletes customer data and destroys model copies containing customer data | 🟠 Only commits to "cease use" but not to deletion |
| Ownership of intellectual property | Clearly stipulate the ownership of the fine-tuned model (customer-owned/supplier-owned/shared) | 🟠 The contract is silent |
Red lines for training data clauses
- The supplier unilaterally reserves the right to use all customer data "for the purpose of improving services," and this is non-negotiable
- The supplier refuses to provide any guarantee regarding the lawful source of training data
- Personal information is involved and the supplier refuses to sign a data processing agreement (refer to Article 21 of the Personal Information Protection Law
[法条原文]）
Step 3: Allocation of liability
The particular nature of AI output may make traditional liability clauses directly inapplicable. Special attention should be paid to:
| Liability scenario | Supplier's ideal position | Risk |
|---|---|---|
| AI output infringement(intellectual property) | The supplier bears infringement liability caused by its training data or the model itself | 🔴 The supplier shifts all infringement risk to the customer |
| AI output illegal/harmful content | The supplier bears content safety responsibility based on the Measures for the Management of Generative Artificial Intelligence Services | 🔴 The supplier claims to be only a "technologically neutral tool" |
| AI output errors cause commercial losses | Liability is allocated reasonably, and special or indirect losses are reasonably excluded | 🟠 The supplier is completely exempt from liability and the customer bears all losses |
| Model downtime/service interruption | The SLA is clear, with availability commitments and service credits/compensation mechanisms | 🟡 The SLA is vague or missing |
| Model performance degradation | The supplier guarantees that the quality of model output will not materially decline | 🟠 The supplier reserves the right to unilaterally modify the model and has no notification obligation |
Special considerations under Chinese law: Article 9 of the Measures for the Management of Generative Artificial Intelligence Services requires providers to bear the producer's responsibility for generated content [法条原文]—If a supplier claims that it only provides a "technical tool" and is not responsible for AI output, that position has weak regulatory support. In practice, however, suppliers may transfer part of the risk to users through contract clauses, and this needs to be analyzed case by case.
Step 4: Model change notification
- Can the supplier unilaterally modify the model? (Usually yes—the key lies in notification and impact assessment)
- How far in advance must model changes be notified? (Industry practice: 30-90 days)
- If a change materially reduces performance or compliance, does the customer have the right to terminate?
- If a supplier stops supporting a certain model version, is there a reasonable exit mechanism?
| Check item | Minimum acceptable standard |
|---|---|
| Notice of material changes | At least 30 days' advance written notice |
| Remedy for performance degradation | If a change causes performance degradation >X%, the supplier must remedy it within a reasonable period |
| Right to terminate | If the supplier cannot remedy the issue or the change affects compliance status, the customer has the right to terminate |
| Compliance impact assessment | The supplier should provide a compliance impact summary before the change (at least a brief description) |
Step 5: Transmission of compliance obligations
As an AI service user, you need to ensure the supplier is able to support your compliance obligations:
| Regulatory compliance requirements | Obligations the supplier should fulfill | Contract review points |
|---|---|---|
Algorithm filing(Article 24 of the Provisions on the Administration of Algorithm Recommendations for Internet Information Services [法条原文]） |
The supplier has completed filing and provided the filing number | Does the contract clearly state the supplier's algorithm filing status? Is there a continuous compliance assurance? |
Security assessment(Article 17 of the Measures for the Administration of Generative Artificial Intelligence Services [法条原文]） |
The supplier has conducted a security assessment | Can a summary of the assessment conclusion be obtained (a complete report is not required, but confirmation that it has been completed is needed)? |
Science and technology ethics review(Measures for Science and Technology Ethics Review (Trial)[法条原文]） |
The supplier has completed an ethics review (if applicable) | Does it involve scenarios requiring ethics review? |
Personal information protection(Personal Information Protection Law[法条原文]） |
If data processing is involved, a data processing agreement should be signed | Adequacy of the data processing agreement |
Security measures(Article 51 of the Personal Information Protection Law [法条原文]） |
The supplier undertakes to take necessary security measures | SOC2/MLPS report, security incident notification time limit, data breach notification obligation |
| Audit rights | The supplier should accept audits or provide third-party audit reports | Scope and frequency of audit rights |
| Content security | The supplier should have mechanisms for identifying and handling illegal and harmful information | Description or certification of information security management capabilities |
Step 6: Assessment and output
Risk summary table
| Category | Legal risk | Commercial friction | Key risk points |
|---|---|---|---|
| Training data | 🔴🟠🟡⚪ | 🔴🟠🟡⚪ | [Key points] |
| Allocation of liability | 🔴🟠🟡⚪ | 🔴🟠🟡⚪ | [Key points] |
| Model changes | 🔴🟠🟡⚪ | 🔴🟠🟡⚪ | [Key points] |
| Compliance transmission | 🔴🟠🟡⚪ | 🔴🟠🟡⚪ | [Key points] |
Output format
[工作成果头 — 按照插件配置 ## 输出]
# AI供应商合同审查：[供应商名称] — [服务类型]
**日期：** [日期]
**供应商：** [名称]
**服务：** [API / SaaS / 本地部署 / 模型授权]
**总体结论：** [经核准 / 附条件核准 / 不适合当前条款]
---
## 一、服务概况
[一段话]
## 二、训练数据风险
### 发现
[具体条款语言及风险分析]
### 建议
[谈判立场 + 备选条款语言]
## 三、责任分配风险
### 发现
[参照第3步清单]
### 建议
[谈判立场 + 备选条款语言]
## 四、模型变更风险
### 发现
[参照第4步清单]
### 建议
[谈判立场 + 备选条款语言]
## 五、合规传导风险
### 发现
供应商是否具备支持客户合规义务的能力：[描述]
### 建议
[需要供应商补充的材料、条款修订建议]
## 六、谈判立场
| 条款 | 当前 | 我方立场 | 最低接受标准 | 谈判优先级 |
|------|------|----------|-------------|-----------|
| [条款] | [现状] | [理想] | [底线] | 致命/高/中/低 |
## 七、红线标记
[列出触发红线的条款——需升级法律顾问]
Closing
Conclude with the next-step decision tree specified in CLAUDE.md ## 输出 . Customization options: negotiate with the supplier based on the review comments, escalate red-line clauses to legal counsel for decision-making, accept the current clauses (if there are no red lines), obtain more supplier information.
What this skill does not do
- Does not cover general contract review elements (governing law, dispute resolution, confidentiality clauses, etc.)—focuses only on AI-specific risks
- Does not assess the model's technical performance—this is a legal and compliance review, not technical due diligence
- Does not replace algorithm filing verification—this skill checks the supplier's filing commitments, but does not verify the authenticity of the filing information (this should be verified independently through the Cyberspace Administration's public channels or the supplier's filing number)