AI, Climate, and Transparency: Operationalizing and Improving the AI Act

Nicolas Alder,Kai Ebert,Ralf Herbrich,Philipp Hacker

2024-08-28

Abstract:This paper critically examines the AI Act's provisions on climate-related transparency, highlighting significant gaps and challenges in its implementation. We identify key shortcomings, including the exclusion of energy consumption during AI inference, the lack of coverage for indirect greenhouse gas emissions from AI applications, and the lack of standard reporting methodology. The paper proposes a novel interpretation to bring inference-related energy use back within the Act's scope and advocates for public access to climate-related disclosures to foster market accountability and public scrutiny. Cumulative server level energy reporting is recommended as the most suitable method. We also suggests broader policy changes, including sustainability risk assessments and renewable energy targets, to better address AI's environmental impact.

Computers and Society,Artificial Intelligence

What problem does this paper attempt to address?

The paper aims to address significant shortcomings and challenges in the Artificial Intelligence Act (AI Act) regarding climate-related transparency requirements. Specifically, the paper identifies the following key issues:

1. **Energy consumption during the inference phase not included in the reporting scope**: The current Act only requires reporting on the computational resource usage during the development, training, testing, and validation phases of AI models, but does not explicitly require disclosure of energy consumption during the inference phase, leading to an incomplete estimation of the environmental impact of high-risk systems.

2. **Indirect greenhouse gas emissions not covered**: The Act fails to cover indirect greenhouse gas emissions generated by AI applications (such as oil and gas exploration), which may have a significant impact on climate change but whose environmental impact is not reported.

3. **Lack of standardized reporting methods**: There is currently no unified reporting method to measure and report the energy consumption of AI systems, making it difficult to compare different systems.

4. **Exemption for open-source AI models**: Open-source general AI models are often exempt from transparency obligations unless they pose systemic risks. However, the energy consumption information of these models is equally important but not required to be disclosed.

5. **Uncertainty in fine-tuning operations**: The Act's provisions are unclear on whether entities performing fine-tuning should be considered as providers of new models, which may result in small entities bearing excessive reporting obligations.

6. **Public inaccessibility to climate-related disclosures**: Currently, climate-related data is only accessible to regulatory agencies, while downstream providers and the public cannot access this information, limiting market effects, independent research, and public oversight.

To address these issues, the paper proposes the following policy recommendations:

1. **Explicitly include energy consumption during the inference phase in the reporting scope**: Clearly require the reporting of energy consumption during the inference phase in Annex XI and XII.

2. **Expand reporting obligations**: Include indirect greenhouse gas emissions and water consumption of AI applications.

3. **Clarify reporting obligations for fine-tuning operations**: Link reporting obligations to computational costs and training mechanisms, and set a minimum computational cost threshold.

4. **Remove the exemption for open-source models**: Ensure that all AI models are subject to climate-related reporting.

5. **Adopt cumulative server-level energy consumption measurement methods**: Reflect the total computation-related electricity usage and separately report the PUE factor of each data center.

6. **Make all climate-related disclosures public**: Make all climate-related data publicly accessible to promote market pressure, reputational effects, and public oversight.

Through these measures, the paper hopes to improve the transparency requirements of the AI Act and better address the environmental impact of AI technology.