Dissertation Proposal Defense - Prerna Juneja
Towards Understanding, Measuring and Defending Against Algorithmically Curated Misinformation: A Socio-technical Approach
Abstract: Search engines and online social media platforms have become important sources of information for users worldwide. Despite their popularity and ubiquitousness, online platforms are not always trustworthy sources of information. The platforms are driven by black box algorithms that optimize for engagement over the credibility of information. There are increasing concerns that online platforms amplify inaccurate information, making it easily accessible via search results and recommendations. If the search interfaces incentivize the public towards conspiratorial content or fail to flag or moderate misinformative content from their platforms, the consequences for democracies around the world are enormous. My dissertation research takes a step toward countering algorithmically curated misinformation by studying multiple dimensions of the problem. I consider misinformation as an algorithm problem, a content moderation problem, a fact-checking problem, and a design problem.
The first thread of my research addresses the algorithm problem by investigating and auditing the online platforms to understand the role of algorithms driving these platforms in surfacing misinformative content to users. I have conducted controlled experiments to audit online platforms such as YouTube and Amazon. Through my audits, I identified the conditions under which algorithms present misinformative content to users as well as vulnerable user populations who could be targets for certain misinformative topics on online platforms. I found that performing certain real-world actions on misinformative content on platforms (e.g. watching a conspiratorial video on YouTube, or adding a misinformative book to the cart on Amazon) could drive users into problematic echo chambers of health misinformation.
My second research thread focuses on the content moderation problem by studying how effectively platforms enact their content moderation policies against online misinformation. I have conducted a large-scale crowdsourced investigation of the YouTube platform to determine how effectively YouTube implemented its policies about election misinformation. The results of the study suggest that while YouTube largely seems successful in regulating election misinformation in the search results, YouTube still offers some pathways to misinformative videos in the up-next video recommendation trails. In the third research thread, I focus on the fact-checking problem by identifying ways to support fact-checking of online misinformation. I interviewed 26 participants belonging to 16 fact-checking organizations across 4 continents to understand how fact-checking is practiced in the real world and identify the barriers to fact-checking online misinformation. Through the study, I establish that improving the quality of fact-checking requires systematic changes in the civic, informational, and technological contexts.
For the final thread of my dissertation research, I use the findings of the previous research threads to design and build a fact-checking system that would allow fact-checkers to monitor online platforms for algorithmically curated misinformation. More specifically, I propose to build an online tool, YouCred that would assist fact-checkers with misinformation discovery and credibility assessments on the YouTube platform. Overall, this thesis adopts a socio-technical approach to understand and defend against algorithmically curated online misinformation. It also opens avenues for future research in investigating the role of algorithms in surfacing other problematic information, such as extremism, and designing interventions aimed at redressing algorithmic harm.
Supervisory Committee:
Chair: Tanushree Mitra, Assistant Professor, Information School
GSR: Sean Munson, Associate Professor, HCDE
Member: Chirag Shah, Professor, Information School
Member: Bill Howe, Associate Professor, Information School