Past
Behind the scenes of digital climate disinformation: methods for detecting communities and tracing information flows
Date & Time
04/06/2025 1:15 pm – 2:45 pm
Simon Lindgren
Umeå University
I am Professor of Sociology and director of DIGSUM, an interdisciplinary academic research centre for the study of social dimensions of digital technology.
My research is about politics, power, and resistance at the intersection of society and digital technologies. I use critical discourse approaches, computational text analysis, and social network analysis to study issues relating to movements, mobilization, opinions, and identities.
About the Event
In this session, Simon shared a detailed account of how he and his co-author, Felicia Lundstedt, examined the spread of climate disinformation across more than 12 million tweets collected during COP26 and COP27. The project combined machine learning, network analysis, and close reading to study how problematic content circulates within and between interaction networks.
The seminar focused not only on methods, but on decisions: how categories are defined, how influence is modelled, and how ambiguity is handled throughout the research process. It offered a clear example of how computational approaches and interpretive judgement can work together in the study of large-scale, politically sensitive data.
For researchers working with classification, social media platforms, or the politics of information, the session provided valuable insight into the practical and conceptual demands of doing this kind of work with care.
5 Key highlights
The study did not rely on a pre-existing or externally validated training set. Instead, the researchers combined two small public datasets and used few-shot learning to construct a custom classifier. Throughout, classification remained an interpretive task, requiring decisions about label structure, balance, and thresholding.
Rather than analyzing who interacts with whom, the researchers reversed the direction of Twitter interactions. Retweets, replies, and quotes were treated as signals of attention toward the original account. This small methodological shift reoriented the analysis away from activity and toward influence, helping surface patterns of reception and agenda-setting rather than simple interaction volume.
Disinformation is not a clearly bounded category. By working with probability scores rather than fixed types, the researchers retained ambiguity in the dataset, enabling post-classification review and thresholding based on context. Manual inspection and sampling were used throughout to validate model performance, reinforcing that computational output alone cannot define complex concepts
Community detection algorithms identified over 280 clusters within the network, but algorithmic output was only a starting point. Classifying communities as disinformation-heavy, neutral, or mixed involved a combination of threshold-setting and close reading of representative tweets. The interpretive work came after the computational step, not instead of it.
Some of the most structurally central accounts in the network were not actively promoting disinformation. In certain cases, a single viral post, taken up by others for their own purposes, placed an account at the center of disinformation flow. This finding highlighted the difference between communicative intention and network position, and underscored the risks of interpreting structural centrality as alignment or endorsement.
Sign up to our seminars calendar
When you sign up, we’ll email you a link to the Data Methods Initiative Events Calendar Feed, where you can access Zoom links and stay updated on all future seminars.
You can also subscribe to our newsletter to receive detailed information, event reminders, and the latest news about our initiative directly in your inbox.
Your data is safe with us—we’ll never share it with anyone else. You can unsubscribe from our emails anytime by using the link in our emails.
For more details, check out our Privacy Notice.