Summary:
Un nouveau document politique de l’ACM Europe traite de la transparence climatique liée à l’IA. Il propose des recommandations pour améliorer la responsabilité environnementale du règlement sur l’IA de l’UE, en soulignant l’importance de la divulgation de l’énergie d’inférence, le reporting des émissions indirectes et la consommation d’eau spécifique à l’IA. Cinq recommandations principales et plusieurs secondaires visent à améliorer la transparence et l’impact environnemental de l’IA, tout en appelant à des améliorations au-delà de la simple transparence.
Original Link:
Original Article:
Our New ACM Europe Policy Brief on AI & Climate Transparency
Jointly with Nicolas Alder, Kai Ebert, and Ralf Herbrich, I have written a new policy brief for the ACM Europe Technology Policy Committee on how the EU AI Act can foster true environmental accountability. This comes at a critical moment, particularly concerning the AI Act CoP, and the EU Water Resilience Strategy.
Thanks also to Bran Knowles, Lynn Kaack, Alejandro Saucedo for comments and substantive contributions!
We argue that the Act, while well-intentioned, fails to capture the full climate impact of AI. Based on a joint technical and legal assessment, we propose five primary and four secondary recommendations to remedy this.
Primary Recommendations:
· Inference Energy Disclosure
The Act currently focuses on energy for training. This misses the far larger emissions during inference, i.e., the actual use of AI. We propose extending Annexes IV (high-risk), XI, and XII (GPAI) to include standardized benchmarks and inference-phase disclosures.
· Indirect Emissions Reporting
AI used in carbon-intensive sectors (e.g., fossil fuel exploration) could also trigger disclosure obligations. The Act should cover indirect emissions by tying reporting to downstream use cases.
· AI-Specific Water Consumption
Cooling AI workloads consumes vast water quantities. The Energy Efficiency Directive does not fully cover this. Annexes IV and XI should explicitly require estimates of water use tied to AI models and data centers.
· No Blanket Exemption for Open-Source Models
Current law exempts open-source GPAI models unless they pose a “systemic risk.” This risks excluding powerful, high-consumption systems. We propose a threshold-based approach: once a model exceeds certain FLOP or user counts, climate transparency rules should apply.
· Standardized Energy Measurement
Without a clear methodology, transparency is useless. We propose measurement at the cumulative server level using Power Distribution Units. Standard benchmarks should also be introduced for task-specific inference energy.
Secondary Recommendations:
· Sustainability Risk Assessment
Expand Articles 9 and 55 to explicitly include environmental impact assessments.
· Time-of-Use Constraints
Encourage training/inference during off-peak hours to optimize grid use.
· Mandatory Renewable Buildout (Additionality)
Require providers to invest in new renewable capacity proportional to their energy consumption.
· AI Energy Credit Market
Explore a cap-and-trade system for AI energy use, modeled on the ETS.
Ultimately, we will have to go beyond transparency. But it is a necessary first step.
Download the full brief:
🔗 https://lnkd.in/eQnBYR8Q
#EUAIAct #ClimateTransparency #GreenAI #ACM #DigitalPolicy #AIRegulation #SustainableTech #EnergyDisclosure #EnvironmentalAccountability