Stanford Seminar - Interest-based & Ideological Biases in Online Content Consumption

Key Concepts

• Partisan news exposure
• Misinformation and radicalization
• Social media platform algorithms
• Ideological bias in algorithms
• Interest bias in recommender systems
• Quality news consumption
• User nudges
• Algorithmic nudges
• Pro-social ranking

Main Topics and Key Points

Introduction

• The speaker, Magdalena Wojcieszak, discusses the fear that the current media environment facilitates exposure to likeminded information, hyper-partisan news, and misinformation, leading to polarization and extremism.
• Scholars have studied partisan news, misinformation, and radicalization, with recent focus shifting to social media platforms and their algorithms.

The Role of Algorithms

• Platforms create unique environments with curated information flows filtered through social networks and reinforced by algorithms.
• Algorithms are designed to optimize user engagement by recommending content aligning with inferred interests based on past behavior.
• The dominant fear is ideological bias, where algorithms recommend politically congenial, potentially extreme, or misinformative content.
• YouTube, with 70% of its content recommended by algorithms, faces scrutiny as a potential radicalizing instrument.

Study on YouTube Recommendations

• A study using 100,000 trained sock puppets (automated accounts) across the political spectrum (very left to very right) analyzed YouTube's algorithm.
• The study examined recommendations on the homepage and "up next" sidebar.
• A pipeline was developed to estimate video ideology by analyzing audiences on Twitter (data access since restricted).
• Findings:
  • Videos on the homepage are largely congenial with the sock puppet's ideology (e.g., 77% of recommendations for very left sock puppets were left or very left).
  • Congeniality increases deeper in the watch trail, especially for very right users.
  • Recommendations to radical, conspiratorial, or problematic channels increase deeper in the trail, with over 36% of sock puppets encountering such recommendations.

Critique of the Dominant Focus

• The speaker argues that the focus on radicalization, filter bubbles, and echo chambers is too narrow and overshadows other dangers.
• Research shows that exposure to problematic content is limited among actual users (0.1% to 5% of the American population).

Interest Bias in Recommender Systems

• The speaker introduces the concept of "interest bias," where content selection is primarily driven by user interests (e.g., sports, cooking) rather than political ideology.
• Many people are politically disengaged, finding news too complex, boring, or negative, leading them to avoid it.
• News comprises a small percentage of total online consumption (e.g., 4.2% on mobile and desktop).
• Data from a large grant involving 7,000 participants in the US, Netherlands, and Poland, who shared their browsing history for nine months, revealed that news exposure averaged only 3.4% of their online activity.
• Political content on news websites accounted for less than 1% of browsing activity.
• Platform reports show even lower news consumption rates on platforms like TikTok (1%) and Facebook (1.4%).

Consequences of Low News Exposure

• Low news exposure can lead to low levels of knowledge, misaligned voting, susceptibility to populism, and misinformation.
• News exposure is linked to better-informed citizens, stable attitudes, the ability to distinguish between fake and credible information, and greater support for democratic norms.
• Interest bias in recommender systems further redirects users away from news and public affairs.
• An audit of YouTube found that even politically interested users are redirected from news toward entertainment by the algorithm.

Interventions to Promote News Exposure

• The speaker presents two interventions aimed at incentivizing greater exposure to verified news and public affairs.

Field Experiment on Twitter

• A field experiment on Twitter aimed to engage users with news through their non-political interests (sports, entertainment, lifestyle).
• Bots were used to engage with users who frequently tweeted about these topics, directing them to topical sections of verified news organizations.
• The study found small but statistically significant effects: treated users followed more news accounts and liked more news content, especially in the female bot condition.
• These effects were primarily confined to users already interested in politics.

Study on YouTube

• A study on YouTube aimed to disentangle user-driven and algorithm-driven factors in low news consumption.
• A browser extension was developed with two interventions:
  • User Nudge: Banners reminding users of the benefits of news consumption.
  • Algorithmic Nudge: The extension automatically watched videos from high-quality news organizations in the background, injecting news into the user's watch history.
• The algorithmic nudge significantly increased recommendations to news and subsequent news consumption, while the user nudge had no effect.
• The algorithmic intervention also diversified the ideological slant of news recommendations and watches, especially for conservative users.
• The study found that recommendations drive exposure, which in turn drives recommendations.

Ongoing and Future Research

• Incentivizing WhatsApp and Instagram users to follow news accounts.
• Pro-social ranking challenge to uprank personally relevant news on X, Reddit, and Facebook.
• Building cross-platform extensions to promote verified, factual, and diverse public affairs information.
• Using AI and LLMs to identify and downrank harmful content on platforms.

Conclusion

• While echo chambers and radicalization are important issues, the more pressing problem is that most people consume no political content.
• Platforms should minimize interest-based selectivity to ensure users encounter more verified and factual political content.
• This can make them more resilient to misinformation, populism, and other democratic threats.
• The speaker poses questions about defining "good" public affairs information, determining the threshold for informed citizens, and balancing user autonomy with the need for news exposure.
• Structural solutions, such as education and addressing inequality, are also crucial.

Important Examples, Case Studies, or Real-World Applications Discussed

• YouTube Radicalization: The concern that YouTube's algorithm directs users to conspiratorial or extreme content.
• Twitter Field Experiment: Using bots to engage users with news through their non-political interests.
• YouTube Browser Extension Study: Using user and algorithmic nudges to promote news consumption.
• WhatsApp and Instagram Study: Incentivizing users to follow news accounts.
• Pro-social Ranking Challenge: App ranking personally relevant news on social media platforms.

Step-by-Step Processes, Methodologies, or Frameworks Explained

• Sock Puppet Training: Training automated accounts to mimic user behavior across the political spectrum.
• Ideology Estimation Pipeline: Analyzing audiences on Twitter to determine the ideological slant of videos.
• Twitter Bot Intervention: Using GPT-2 to generate contextual responses to users tweeting about non-political topics, directing them to news sources.
• YouTube Browser Extension Design: Developing a browser extension with user and algorithmic nudges to promote news consumption.

Key Arguments or Perspectives Presented, with Their Supporting Evidence

• Argument: The focus on radicalization and echo chambers is too narrow.
  • Evidence: Research shows that exposure to problematic content is limited among actual users (0.1% to 5% of the American population).
• Argument: Interest bias in recommender systems is a significant problem.
  • Evidence: News comprises a small percentage of total online consumption (e.g., 4.2% on mobile and desktop).
  • Data from a large grant revealed that news exposure averaged only 3.4% of online activity.
• Argument: Algorithmic nudges can increase news consumption and diversify exposure.
  • Evidence: The YouTube browser extension study showed that the algorithmic nudge significantly increased recommendations to news and subsequent news consumption, while the user nudge had no effect.

Notable Quotes or Significant Statements with Proper Attribution

• "YouTube has been termed the most radicalizing instrument of the past century." (Attribution: General claim, not attributed to a specific person)
• "The important problem is not that people consume bad political content online. This more important problem is that most people consume no political content." (Attribution: Magdalena Wojcieszak)
• "This democratic broccoli which is verified democratically good broccoli which is the verified factual news or public affairs information are a tiny fraction of people's online diets." (Attribution: Magdalena Wojcieszak)

Technical Terms, Concepts, or Specialized Vocabulary with Brief Explanations

• Sock Puppets: Automated accounts used to mimic user behavior in online experiments.
• Word Embeddings: A technique in natural language processing to represent words as vectors in a high-dimensional space, capturing semantic relationships between words.
• GPT-2: A large language model used to generate contextual responses in the Twitter bot intervention.
• Web Historian: An open-source tool used to access and visualize users' browsing history.
• LLMs (Large Language Models): AI models trained on vast amounts of text data, capable of generating human-like text.
• Multimodal LLMs: LLMs that can process and generate content from multiple modalities, such as text, images, and audio.

Logical Connections Between Different Sections and Ideas

• The speaker begins by outlining the concerns about online polarization and the role of algorithms.
• She then presents a study on YouTube recommendations, highlighting the presence of ideological bias.
• She critiques the dominant focus on echo chambers and introduces the concept of interest bias.
• She presents data showing low levels of news consumption and discusses the consequences.
• She then presents two interventions aimed at promoting news exposure, one on Twitter and one on YouTube.
• Finally, she discusses ongoing and future research and poses questions about the challenges and ethical considerations of this line of work.

Data, Research Findings, or Statistics Mentioned

• 70% of YouTube content is recommended by the algorithm.
• 0. 1% to 5% of the American population are exposed to problematic content online.
• News comprises 4.2% of total online consumption on mobile and desktop.
• Political content on news websites accounts for less than 1% of browsing activity.
• News comprises 1% of content on TikTok and 1.4% on Facebook.
• A study of 1.5 million Twitter users found that the majority do not follow any politicians, news media organizations, or pundits.

Synthesis/Conclusion of the Main Takeaways

The presentation argues that while concerns about echo chambers and radicalization are valid, the more pressing issue is the widespread lack of news consumption due to interest bias in recommender systems. The speaker presents research showing that news exposure is limited and that algorithmic nudges can be effective in increasing news consumption and diversifying exposure. She concludes by posing questions about the ethical considerations and challenges of promoting news exposure and calls for a broader focus on addressing the fundamental challenges in our information environment.