在本页3 分钟阅读
概述
More than 10^25 training FLOP creates a presumption of high-impact capability, not an unchallengeable classification; systemic-risk providers have extra evaluation, risk-management, incident-reporting, and cybersecurity duties.
深入探讨
The AI Act’s systemic-risk category is a subset of general-purpose AI models. Article 51 provides two routes: a model has high-impact capabilities assessed with appropriate tools, indicators, and benchmarks; or the Commission designates it because capabilities or impact are equivalent, using Annex XIII criteria. Training compute above 10^25 floating-point operations creates a presumption of high-impact capability. The number is a threshold in the law and is subject to delegated adjustment; it is not the only way to qualify. Article 52 requires a provider whose model meets the compute condition to notify the Commission without delay, and no later than two weeks after the condition is met or it becomes known that it will be met. The provider may submit substantiated arguments that the model does not, due to its specific characteristics, present systemic risks. The Commission can reject that case or designate a model on its own initiative or following a qualified scientific-panel alert. Commission guidelines describe its interpretation and enforcement approach but are not binding law. Article 55 adds four duties beyond the general GPAI requirements: standardized model evaluation, including documented adversarial testing; assessment and mitigation of possible systemic risks at Union level; documentation and reporting of serious incidents and corrective measures; and adequate cybersecurity for both model and physical infrastructure. These duties are not waived just because a model is distributed under a free and open-source license. Providers can rely on an approved code or harmonised standards while available, or demonstrate alternative adequate means for Commission assessment. GPAI obligations began applying on August 2, 2025. The Commission’s guidance states that its enforcement powers apply from August 2, 2026, and qualifying models already on the market before August 2, 2025 have a compliance transition until August 2, 2027. Providers should track training-compute evidence, expected threshold timing, notification, and any designation decision. This is an overview, not a legal opinion on an individual model.
战略影响
成本与预算
多年来，架构决策决定着性能和运营成本。
更清晰的判决
技术教育帮助团队选择正确的堆栈，而不仅仅是最新的堆栈。
质量控制
更好的工程选择可以减少生产中的可靠性事故。
The Future of GPAI Models with Systemic Risk Under the EU AI Act
The Act empowers the Commission to revise compute thresholds and benchmarks as capabilities and hardware efficiency evolve. A provider should not assume today’s FLOP threshold will remain fixed or that falling below it resolves designation risk. Monitor official Commission guidelines and any delegated acts, and review status when a model’s training run, capabilities, distribution, or deployment context changes. Track delegated updates to Annex XIII and any official changes to the compute presumption. Document threshold calculations consistently across training runs. Review changes before each release.
现实世界的实施
A provider forecasts training compute above 10^25 FLOP and prepares the Article 52 notification before the threshold is reached.
A provider crossing the threshold submits evidence with its notification explaining why the model’s specific capabilities do not create systemic risk.
A safety team documents adversarial testing, EU-level risk assessments, incident response, and infrastructure security for a designated model.
A company with an open-source systemic-risk model checks Article 55 duties rather than assuming Article 53’s limited exception covers it.
风险与防护栏
优化一项基准测试可以隐藏更广泛的系统弱点。
基础设施和维护成本常常被低估。
随着系统变得更加复杂，安全性和可观察性差距可能会扩大。
实施路线图
在实施之前定义延迟、质量和成本目标。
在实际负载和数据条件下进行基准测试。
仪器监控错误、漂移和用户影响。
在扩展之前准备回滚和事件响应路径。
不断探索
Free newsletter
Get the daily AI briefing
Three verified AI stories every weekday morning, written in plain English. Free forever, no ads.
One email each weekday. Unsubscribe in one click. We never sell or share your address.
Test yourself
Take the GPAI Models with Systemic Risk Under the EU AI Act quiz
Instant feedback on every answer, and a shareable certificate with a verifiable ID once you pass a course.
Support free AI education. AI Understanding is a 501(c)(3) nonprofit — no ads, no paywall, ever. Make a donation
常见问题
What is GPAI Models with Systemic Risk Under the EU AI Act?
The EU AI Act treats a general-purpose AI model as systemic-risk when it has high-impact capabilities or the Commission designates it for equivalent capabilities or impact. More than 10^25 training FLOP creates a presumption of high-impact capability, not an unchallengeable classification; systemic-risk providers have extra evaluation, risk-management, incident-reporting, and cybersecurity duties.
What does training compute above 10^25 FLOP do under Article 51?
Article 51(2) presumes high-impact capabilities above the compute threshold, while Article 51 and 52 allow other evidence and procedures.
Which second route can lead to systemic-risk designation besides the compute presumption?
Article 51(1)(b) and Article 52(4) allow Commission designation based on equivalent capabilities or impact using Annex XIII criteria.
When must a provider notify the Commission after a model meets or is expected to meet the compute condition?
Article 52(1) requires notice without delay and no later than two weeks after the condition is met or known to be expected.
Which item is an Article 55 duty for a systemic-risk GPAI provider?
Article 55(1)(a) requires model evaluation under state-of-the-art protocols, including documented adversarial testing.
What does Article 55 require about serious incidents?
Article 55(1)(c) requires tracking, documenting, and reporting serious-incident information and possible corrective measures.
继续学习
相关指南
为此主题精选的更多指南