How CO2STLY Is CHI? The Carbon Footprint of Generative AI in HCI Research and What We Should Do About It

Research Background and Issues

Problems and Challenges

The authors identify that the widespread use of Generative AI (GenAI) in the stages of method development and research has led to significant energy consumption, which in turn raises notable carbon emission concerns. In the field of Human-Computer Interaction (HCI), although GenAI is often employed for model evaluation and user studies, there is a significant lack of transparency in reporting the carbon footprint resulting from its use in related research literature.

Importance of the Problem

As global attention focuses on climate change, the environmental impact of the technology industry—particularly the development, training, and application of algorithms and models—has become an issue that cannot be ignored. The popularity of GenAI technology has further amplified the demand for computational resources, leading to a significant increase in electricity consumption. For instance, the electricity usage of data centers has grown by 20%-40% in recent years. This conflicts with the current climate crisis and may hinder energy policies at national or regional levels.

Research Motivation and Related Work

The motivation of this study is to raise awareness of the carbon footprint caused by the use of GenAI in HCI research and to promote more transparent and responsible actions regarding energy consumption by estimating carbon emissions. Previous related studies in the field have explored how to assess the environmental impact of technology, but this paper seeks to specifically address the following key research questions:

• How is GenAI used in HCI research?
• What is the scale of the carbon footprint caused by GenAI?
• What specific recommendations can be proposed to address this issue?

Solution

Methodology

The authors conducted a systematic review of both accepted and rejected papers from the 2024 CHI conference, analyzing a total of 282 papers that utilized GenAI technology, including estimates for rejected submissions. The research process involved the following specific steps:

1. Keyword Search: Papers containing keywords related to GenAI technology were filtered from the ACM database.
2. Preliminary Screening: Articles that only discussed social impacts or did not directly use GenAI technology were excluded, retaining only those reporting specific application scenarios.
3. Classification and Annotation: By reviewing the methodology sections of the papers, the applications of GenAI were categorized into different research stages, and specific technical parameters used were recorded.
4. Carbon Emission Estimation Experiments: Open-source models were run on local hardware to estimate the energy consumption of model usage.

Innovations

• Carbon Footprint Transparency: Conducted a detailed data audit of the carbon footprint resulting from GenAI usage in HCI research and proposed recommendations for standardized transparent reporting.
• Comprehensive Tool Application: Used multiple open-source tools to measure energy consumption during the inference phase, obtaining precise data for result analysis.

Implementation Steps and Techniques

• Model Evaluation Experiments: Measured energy consumption using tools like Carbontracker for tasks such as text-to-image (Stable-Diffusion-XL), text-to-text (Llama-3.1-Instruct), and speech-to-text (Whisper).
• Classification Framework Development: Systematically categorized the research stages and functions of GenAI usage in each paper, proposing a seven-stage model for HCI research.
• Experimental Data Synthesis: Combined the measured carbon footprint per functional use case with reported usage frequencies in the literature to estimate the overall carbon footprint.

Research Findings

Specific Results

1. Usage Patterns and High-Consumption Stages: GenAI is most commonly used in the prototyping and user research stages of HCI research, but the data collection and model training stages consume the most energy.
2. Estimated Carbon Emissions: Papers submitted to the 2024 CHI conference resulted in approximately 10,769.63 to 10,925.12 kilograms of CO2e emissions, equivalent to the carbon footprint of driving approximately 100,000 kilometers.

Advantages Over Existing Solutions

• Fine-Grained Data Analysis: The authors provided average carbon footprint estimates across different model types and usage stages, addressing the lack of specific data in existing research.
• Comprehensive Action Recommendations: Proposed multi-level action recommendations encompassing individual researchers, the academic community, and policy-making institutions.

Experimental or Evaluation Results

Through classification analysis, the authors found that prototyping and user research are the primary stages where GenAI is used in HCI research, while data generation and model training are the phases with the highest carbon footprint. Further experiments demonstrated that the energy consumption of the inference phase increases linearly with the input task volume.

Limitations and Future Directions

• Lack of Data Transparency: Most research papers do not report key data such as model usage frequency or hardware environment, limiting the accuracy of carbon emission estimates.
• Conservative Assumptions in Estimation: The estimation of the number and distribution of rejected papers involves some uncertainty.
• Future Directions: Encourage researchers to proactively report environmental impacts and develop tools to support simple and efficient carbon footprint estimation processes.

Conclusion

This paper systematically audits the use of GenAI in HCI research and its environmental impact, revealing its significant carbon footprint. It provides actionable recommendations for the HCI research community to transparently report carbon footprints and calls for researchers and institutions to take collective measures to mitigate the negative environmental impacts of technology while continuing responsible research on generative AI in the field of human-computer interaction.