In order to limit global warming to a safe level of 1.5℃, individual action is required. The project Data-driven understanding of low-carbon lifestyles (LOW-AI) analyzes the extent and drivers of low-carbon consumption using social media data. LOW-AI deploys social media data (Reddit) to monitor attitudes towards climate solutions and especially individual action, developing tools that can be implemented with a higher geographical reach and are less costly than traditional approaches.
Demand reduction through behavioral changes is considered an important mitigation option, and the crucial role of social systems in rapid decarbonization is increasingly acknowledged. Complementing available empirical sources with global social media data to explicitly capture bottom-up behavioral and opinion change will enable assessing the feasibility of widespread behavioral changes and social tipping mechanisms.
LOW-AI’s research is based on the use of social media data, which helps to gauge the interest of the public in topics related to climate change. The project will use this data to infer information about behavior change, and public attitudes towards it. That will complement already existing survey-based data both in terms of temporal and geographical coverage.
LOW-AI will first validate the use of social media data for the analysis of climate change opinion, attitudes and low-carbon consumption choices against traditional data sources. In a second stage, the validated methods will be used to extend the available behavior information to countries, population groups and timeframes for which traditional data is missing. Finally, the results will be displayed in an interactive dashboard openly accessible to anyone.
For the analysis of social media data, LOW-AI develops AI methods that make large-scale text analysis possible with minimal human input. A key bottleneck in supervised text classification is creating a high-quality training dataset, which is usually done by human annotators. In LOW-AI, we replace this manual labeling step with large language models (LLMs) that can reliably label climate-relevant text for specific categories (e.g., topic frames or stance). However, rather than deploying LLMs directly at inference time, we use them to generate high-quality, task-specific training data, which is then used to fine-tune compact transformer models (e.g., DistilBERT, RoBERTa) to identify topics, emotions, and sentiment expressed in the social media content.
The topic classifiers we developed can be found on Hugging Face:
Publications
Gaupp, F. & Eker, S. (2024). Climate Activism, Social Media and Behavioural Change: A Literature Review. IIASA Working Paper. Laxenburg, Austria: WP-24-007
Eker, S. , Garcia, D., Valin, H. , & van Ruijven, B. (2021). Using social media audience data to analyse the drivers of low-carbon diets. Environmental Research Letters 16 (7) 074001. 10.1088/1748-9326/abf770.
News
11 December 2024
Advancing climate insights with social media data
IIASA recently hosted an interdisciplinary research workshop under the LowAI project to explore how social media and digital platforms can advance climate and sustainability research. The participants tackled the challenges of leveraging digital data to shape effective climate actions and foster social change. IIASA researcher Sandeep Chowdhary shares his insights and experiences from the event.
Annual Report 2023
Harnessing the benefits of digitalization for reducing energy demand
Various IIASA studies have highlighted the need for societal and technological transformations that could form the basis for new lifestyles to create a more just, net-zero world through innovation.