Facilitating system-level behavioural climate action using computational social science

R ecently, a Comment in Nature Human Behaviour recommended the collection of large-scale behavioural datasets through public data observatories to enable system-level climate action. Computational social science (CSS) approaches offer important tools to use these large-scale datasets to facilitate climate action. The time for substantial climate action is now, and public opinion matters in people-centric system-level change. The urgency of this message means that sweeping social, political and economic changes are necessary. Improving the public’s understanding of climate change mitigation and adaptation strategies is crucial. Jenny and Betsch proposed these large-scale behavioural datasets can enable system-level climate action through public data observatories. Although important, their proposal raises considerable challenges in ensuring the political neutrality of the data, protecting privacy in open datasets, adhering to internationally agreed-upon scientific practices, making the data relevant to policy, and sustaining the funding required to collect international and representative cross-sectional and longitudinal surveys. The more important problem is that simply providing large-scale behavioural data is insufficient. Although the attempt to scale individual-level behavioural insights to system-level policy-design applications through this proposed public observatory is timely and valuable, we find that Jenny and Betsch limit their scope on how these observatory datasets will be used to produce actual climate action that can drive consensus for mitigation and the adaptation of policy actions. This gap in assessment presents a welcome opportunity to bring together data scientists and social scientists to build actionable research that can mitigate the effects of climate change. CSS can act as a critical interdisciplinary bridge that advances theories of human behaviour using large and sometimes disparate datasets, generating key insights using observational, experimental and machine-learning methods. For instance, recent studies have used CSS approaches to measure the reactiveness of climate policies in hard-to-decarbonize sectors, identifying growing polarization in public opinions regarding climate change and classifying contrarian claims in communication networks. CSS provides appropriate tools to extract behavioural insights at scale from the data observatories (Fig. 1), enabling systems design that counters misinformation and polarization and supports consensus building for climate mitigation and adaptation policies. We propose an operationalizing framework for aiding system designs that can reduce climate misinformation, remove scepticism and recover trust in science (Fig. 1) through data-driven processes, including clustering, categorization and forecasting. For example, reducing misinformation using a CSS approach can help to establish bidirectional links between information ‘producers’ and ‘consumers’ at a system level. Existing CSS applications show that misinformation and polarization can be detected using deep learning models. Newer applications such as OpenAI’s ChatGPT, which aims to mimic human conversation built using a human-in-the-loop reinforcement learning algorithm, provide promising advances for researchers. However, these methods are dependent on the use of training data from internet sources such as Wikipedia and BookCorpus databases, which are controversial owing to embedded bias and data justice issues. Another challenge in countering misinformation and polarization is the lack of standardized measurement instruments applied to online conversations. CSS-driven large language models can help to develop epidemiological models for identifying, treating and preventing misinformation, which require specialist interdisciplinary behavioural and data science skills. Many important problems, such as climate scepticism, require the use of population-level data. The dissemination of climate misinformation and conspiracy theories about climate change is difficult to detect in small-n survey datasets, requiring instead natural language processing and network modelling. To effectively manage climate scepticism, population-level data and analysis are required for the identification and prediction of contrarianism. CSS provides the necessary tools. For instance, a recent natural-language processing study developed a fact-checking taxonomy to remove contrarian claims based on a large-scale, multistakeholder Check for updates

Facilitating system-level behavioural climate action using computational social science R ecently, a Comment in Nature Human Behaviour recommended the collection of large-scale behavioural datasets through public data observatories to enable system-level climate action 1 . Computational social science (CSS) approaches offer important tools to use these large-scale datasets to facilitate climate action.
The time for substantial climate action is now, and public opinion matters in people-centric system-level change. The urgency of this message means that sweeping social, political and economic changes are necessary. Improving the public's understanding of climate change mitigation and adaptation strategies is crucial. Jenny and Betsch proposed these large-scale behavioural datasets can enable system-level climate action through public data observatories 1 . Although important, their proposal raises considerable challenges in ensuring the political neutrality of the data, protecting privacy in open datasets, adhering to internationally agreed-upon scientific practices, making the data relevant to policy, and sustaining the funding required to collect international and representative cross-sectional and longitudinal surveys 1 .
The more important problem is that simply providing large-scale behavioural data is insufficient. Although the attempt to scale individual-level behavioural insights to system-level policy-design applications through this proposed public observatory is timely and valuable, we find that Jenny and Betsch 1 limit their scope on how these observatory datasets will be used to produce actual climate action that can drive consensus for mitigation and the adaptation of policy actions. This gap in assessment presents a welcome opportunity to bring together data scientists and social scientists to build actionable research that can mitigate the effects of climate change.
CSS can act as a critical interdisciplinary bridge that advances theories of human behaviour using large and sometimes disparate datasets, generating key insights using observational, experimental and machine-learning methods 2 . For instance, recent studies have used CSS approaches to measure the reactiveness of climate policies in hard-to-decarbonize sectors 3 , identifying growing polarization in public opinions regarding climate change 4 and classifying contrarian claims in communication networks 5 .
CSS provides appropriate tools to extract behavioural insights at scale from the data observatories ( Fig. 1), enabling systems design that counters misinformation and polarization and supports consensus building for climate mitigation and adaptation policies. We propose an operationalizing framework for aiding system designs that can reduce climate misinformation, remove scepticism and recover trust in science ( Fig. 1) through data-driven processes, including clustering, categorization and forecasting.
For example, reducing misinformation using a CSS approach can help to establish bidirectional links between information 'producers' and 'consumers' at a system level. Existing CSS applications show that misinformation and polarization can be detected using deep learning models 6 . Newer applications such as OpenAI's ChatGPT, which aims to mimic human conversation built using a human-in-the-loop reinforcement learning algorithm 7 , provide promising advances for researchers. However, these methods are dependent on the use of training data from internet sources such as Wikipedia and Book-Corpus databases, which are controversial owing to embedded bias and data justice issues 7,8 . Another challenge in countering misinformation and polarization is the lack of standardized measurement instruments applied to online conversations 9 . CSS-driven large language models can help to develop epidemiological models for identifying, treating and preventing misinformation 9 , which require specialist interdisciplinary behavioural and data science skills.
Many important problems, such as climate scepticism, require the use of population-level data. The dissemination of climate misinformation and conspiracy theories about climate change is difficult to detect in small-n survey datasets, requiring instead natural language processing and network modelling 10 . To effectively manage climate scepticism, population-level data and analysis are required for the identification and prediction of contrarianism 5   Correspondence climate-communication dataset 5 . This application shows the potential of behavioural data repositories.
Restoring public trust in climate research using behavioural science can be facilitated using large population-scale modelling that produces insights that can drive individual-level attitudinal and behavioural change 1,10 . CSS provides the tools to analyse population-scale data to reinforce and coach public and policy actors to distribute trustworthy information 4 . For example, approaches using human-in-the-loop deep-learning systems, known as reinforcement learning with human feedback 7 , can take advantage of artificial intelligence models and human evaluators to analyse large datasets and generate more-trustworthy behavioural insights.
We welcome the growing focus on systemic policy-level changes within the behavioural science community through large-scale public datasets -although CSS research challenges remain, for example, in developing approaches to deal with the injustices and representation biases 8 that we discussed earlier. However, in doing so, we need to combine CSS and behavioural science to make climate action possible -in particular, behavioural and computational social scientists need to collaborate.
Accordingly, there is an urgent need to train future leaders and researchers to be better equipped to tackle interdisciplinary problems through CSS. Importantly, researchers must come together to pursue mission-driven projects that reduce misinformation and scepticism and restore public trust. These efforts must extend beyond laboratories and university campuses to shape the science-policy interface. Only then, with an adequately funded, multidisciplinary and concerted research programme, will we be able to leverage the promise of big data and behavioural science to encourage appropriate climate action.