Improving Machine Learning to Detect and Understand Online Conspiracy Theories
Conspiracy theories circulated online via social media contribute to a shift in public discourse away from facts and analysis and can contribute to direct public harm. Social media platforms face a difficult technical and policy challenge in trying to mitigate harm from online conspiracy theory language. As part of Google’s Jigsaw unit’s effort to confront emerging threats and incubate new technology to help create a safer world, RAND researchers conducted a modeling effort to improve machine-learning (ML) technology for detecting conspiracy theory language. They developed a hybrid model using linguistic and rhetorical theory to boost performance. They also aimed to synthesize existing research on conspiracy theories using new insight from this improved modeling effort. This report describes the results of that effort and offers recommendations to counter the effects of conspiracy theories that are spread online.
- The hybrid ML model improved conspiracy topic detection.
- The hybrid ML model dramatically improved on either single model’s ability to detect conspiratorial language.
- Hybrid models likely have broad application to detecting any kind of harmful speech, not just that related to conspiracy theories.
- Some conspiracy theories, though harmful, rhetorically invoke legitimate social goods, such as health and safety.
- Some conspiracy theories rhetorically function by creating hate-based “us versus them” social oppositions.
- Direct contradiction or mockery is unlikely to change conspiracy theory adherence.
- Engage transparently and empathetically with conspiracists.
- Correct conspiracy-related false news.
- Engage with moderate members of conspiracy groups.
- Address fears and existential threats.
Table of Contents
- Chapter OneIntroduction: Detecting and Understanding Online Conspiracy Language
- Chapter TwoMaking Sense of Conspiracy Theories
- Chapter ThreeModeling Conspiracy Theories: A Hybrid Approach
- Chapter FourConclusion and Recommendations
- Appendix AData and Methodology
- Appendix BStance: Text Analysis and Machine Learning