第五章——通用人工智能

欧盟人工智能法案中的通用人工智能（GPAI）义务：第五章指南

欧盟人工智能法案的第五章规范了通用人工智能（GPAI），即大多数人所说的基础模型。它分为两个层级：第五十三条规定的标准GPAI义务，以及第五十五条规定的对具有系统性风险的模型提出的额外义务。

上次更新: 2026年7月4日

什么算作GPAI？

欧盟人工智能法案对通用人工智能模型进行了广泛的定义：

一种显示出显著通用性，并且能够胜任各种不同任务的人工智能模型，无论其如何投放市场。大多数大型语言模型、图像生成模型和多模态基础模型都符合这一条件。

第五十三条：标准GPAI提供者义务

所有GPAI提供者必须满足以下基本要求：

维护技术文档，包括培训和测试过程
向集成该模型的下游提供者提供信息
遵守欧盟版权法，包括制定尊重文本和数据挖掘退出政策
发布足够详细的培训数据摘要
如果公司不在欧盟设立，则指定一名欧盟代表

第五十五条：系统性风险GPAI义务

符合系统性风险阈值的模型（目前为10²⁵ FLOPs的培训计算量，仅供参考）将面临额外的义务：

模型评估，包括对抗性测试
系统性风险评估和缓解
网络安全保护
向人工智能办公室报告重大事件
能源消耗披露

行为准则

委员会制定了一项GPAI提供者自愿行为准则：

采用该准则具有自愿性，但可以对其中涵盖的义务提供符合性推定。

开源豁免

开源GPAI模型在法规下享有优惠待遇：

根据自由和开源许可协议发布的GPAI模型，基本上可以免除第五十三条的规定，但版权和培训数据摘要义务除外——而且，该豁免条款完全不适用于系统性风险层级的模型。

GPAI Compliance Timeline

The Chapter V obligations phase in on their own schedule, distinct from the high-risk deadlines:

2 August 2025 — the GPAI provider obligations under Articles 53 and 55 begin to apply to models placed on the market from this date
2 August 2026 — the AI Office's supervisory and enforcement powers over GPAI providers become applicable
2 August 2027 — GPAI models already on the market before 2 August 2025 must be brought into compliance by this date
Signing the Code of Practice — providers who adhere to the GPAI Code of Practice can rely on it to demonstrate conformity while harmonised standards are still being developed

Dates reflect Regulation (EU) 2024/1689 as it stands; confirm against the latest Official Journal text and AI Office guidance, which may evolve.

How AIAgentree helps

AIAgentree governs how your agents use GPAI models in production, giving you the downstream documentation and incident evidence Chapter V expects.

Tamper-evident decision records log how a GPAI-backed agent reasoned toward each decision, giving you durable evidence for downstream-provider documentation and serious-incident reporting
Human-oversight and approval workflows plus outcome tracking help you evidence the risk-mitigation and monitoring practices Article 55 asks of systemic-risk deployments
Python and TypeScript SDKs with REST, MCP, A2A, and OpenTelemetry integrate model-usage evidence into your compliance stack, with EU data residency in Germany and audit-fit retention of at least six months

开始免费试用

Frequently Asked Questions

What counts as a general-purpose AI model under the EU AI Act?

A GPAI model is one that displays significant generality and can competently perform a wide range of distinct tasks, regardless of how it is released. Most large language models, image-generation models, and multimodal foundation models fall within this definition.

What is the systemic-risk threshold for GPAI?

A GPAI model is presumed to carry systemic risk when the cumulative compute used for its training exceeds 10^25 floating-point operations, an indicative threshold the Commission can adjust. Such models face extra duties under Article 55, including model evaluations, adversarial testing, risk mitigation, cybersecurity, and serious-incident reporting.

Is the GPAI Code of Practice mandatory?

No. The Code of Practice is voluntary. Providers who adhere to it can use it to demonstrate conformity with their Chapter V obligations while harmonised standards are still being finalised; providers who do not sign it must demonstrate compliance by other adequate means.

Are open-source models exempt from the GPAI rules?

Only partly. GPAI models released under a free and open-source license are largely exempt from the Article 53 obligations, but they must still meet the copyright-policy and training-data-summary duties. The exemption does not apply at all to models that meet the systemic-risk threshold.