Making Sense of Algorithms in News Feeds: Varied Perspectives

# Making Sense of Algorithms in News Feeds: Varied Perspectives **Video Category:** Human-Computer Interaction / Technology Research ## 📋 0. Video Metadata **Video Title:** Making Sense of Algorithms in News Feeds: varied perspectives **YouTube Channel:** Stanford Center for Professional Development **Publication Date:** January 20, 2017 **Video Duration:** ~1 hour 11 minutes ## 📝 1. Core Summary (TL;DR) This presentation explores how users perceive, misunderstand, and interact with the invisible algorithms that curate social media news feeds. It reveals that the vast majority of users are entirely unaware that their feeds are algorithmically filtered, leading them to form incorrect mental models and blame themselves for missed social connections. By introducing "seamful" design interventions that explicitly reveal algorithmic filtering, developers can dramatically increase user awareness, trust, and long-term satisfaction while empowering external researchers to audit these opaque systems for critical biases and discrimination. ## 2. Core Concepts & Frameworks * **Concept:** Algorithm Awareness -> **Meaning:** The degree to which a user understands that a computer system is actively filtering, prioritizing, and curating the content they are exposed to, rather than presenting a raw chronological list. -> **Application:** Designing interfaces that explicitly state when content is sorted by "Top Stories" versus "Most Recent," allowing users to understand why they are seeing specific posts. * **Concept:** Seamful Design (Seamed Interaction) -> **Meaning:** An interface design philosophy that intentionally reveals the underlying mechanics, limitations, or data sources of a system (the "seams"), rather than hiding them behind a perfectly smooth, opaque user experience. -> **Application:** Displaying a side-by-side comparison of "All Content" versus "Algorithmically Filtered Content" to make the system's curation process visible to the user. * **Concept:** Folk Theories -> **Meaning:** The informal, non-expert mental models and hypotheses that users invent to explain how a complex, opaque system works when they lack access to the actual rules or code. -> **Application:** Users developing a "Personal Engagement Theory," assuming that Facebook only shows them posts from friends they have recently clicked on or messaged. * **Concept:** Algorithmic Auditing -> **Meaning:** The systematic process of externally testing an algorithm to detect hidden biases, price discrimination, or civil rights violations, typically by manipulating inputs and measuring outputs without needing access to the proprietary source code. -> **Application:** Researchers creating automated "sock puppet" accounts with different demographic profiles to test if a platform serves discriminatory housing advertisements. * **Concept:** Collective Auditing -> **Meaning:** A decentralized method where multiple users aggregate their individual, personalized algorithmic outcomes to reverse-engineer a system's behavior. -> **Application:** Consumers sharing their successful and unsuccessful bidding prices on a travel forum to map out the hidden pricing thresholds of a booking algorithm. ## 3. Evidence & Examples (Hyper-Specific Details) * **Twitter Trending Algorithm Manipulation:** When Twitter shifted its trending algorithm to highlight smaller "peaks and valleys" rather than sustained global trends, Justin Bieber dropped off the trending list. Unaware of the algorithm change, teenage users assumed the system was broken and successfully manipulated it by tweeting about "Jieber" to bypass the filters. * **The "Bic Cristal For Her" Anomaly:** Researcher Christian Sandvig posted a joke Amazon review about a pen on Facebook. Despite being a minor post, it stayed at the top of his colleagues' feeds for weeks. A group of computer scientists could not deduce why the algorithm prioritized it, prompting the creation of a formal study on feed algorithms. * **FeedVis User Study:** Researchers built "FeedVis," an API tool showing users two columns: everything their friends posted (left) and what the algorithm actually showed them (right). The study found that in 2013/2014, 62.5% of users were completely unaware that Facebook filtered their feed. When exposed to the hidden content (which was roughly 3 times larger than the shown content), users exhibited visceral anger and felt they were living in "The Matrix." * **Post-Audit Satisfaction Shift:** Despite initial anger, follow-up interviews 2 to 6 months after using FeedVis showed that roughly 80% of users had the same or higher satisfaction with Facebook. Knowing the algorithm existed reduced feelings of personal rejection (e.g., feeling "snubbed" when a post got no likes) and increased active use of feed control settings. * **We Meddle Custom Algorithm Tool:** Eric Gilbert created "We Meddle," a tool allowing users to define their own Twitter algorithms using sliders (e.g., adjusting "tie strength" or prioritizing "loud and quiet friends"). Users actively engaged with these controls, with one user noting the customized feed altered the romantic relationship ads they were subsequently targeted with. * **Algorithmic Political Bias on Twitter:** A 2015 data audit revealed that a standard query for political topics on Twitter yielded predominantly Democratic data and tweets. The algorithm's output was heavily skewed, indicating that the input training data natively possessed a strong political bias. * **Latanya Sweeney's Ad Discrimination Audit:** Sweeney manually searched names associated with different racial demographics. She found that searches for distinctly African-American names (e.g., Keisha Bentley) triggered ads implying criminal records ("Arrested?"), whereas searches for distinctly White names triggered neutral background check ads. * **Home Depot Price Discrimination:** An audit by Christo Wilson discovered that Home Depot's mobile application consistently presented prices that were, on average, $0.50 higher than the prices shown to desktop users for the exact same products. * **Orbitz OS-Based Discrimination:** Orbitz was audited and found to routinely prioritize and show more expensive flight and hotel options to users browsing on Mac OSX systems compared to those using Windows machines. * **Wall Street Journal "Blue Feed, Red Feed":** This seamful visualization tool used Facebook's API to show users side-by-side feeds of what a strictly liberal user versus a strictly conservative user would see regarding the same topic (e.g., Michelle Obama), explicitly highlighting algorithmic polarization. * **Forsyth and Fleck's "Finding Naked People" Code (1996):** Early computer vision code designed to detect skin tones successfully identified lighter skin but failed to detect darker skin. The developers had to manually hardcode new heuristic variables (adjusting Hue and Saturation thresholds) to eliminate the racial bias. ## 4. Actionable Takeaways (Implementation Rules) * **Rule 1: Expose the "Seams" of the System** -> Do not build completely seamless, "magic" interfaces. Introduce intentional friction or visual indicators that show the user exactly where data is coming from and how it is being processed (e.g., Kayak showing the specific airlines it is actively searching). * **Rule 2: Provide Explicit Categorization Controls** -> Instead of relying solely on implicit behavioral tracking, offer users transparent toggle switches and sliders to define their own feed priorities (e.g., sliders for tie-strength, topic preference, or chronological vs. top-rated sorting). * **Rule 3: Implement A/B Comparison Views for Education** -> To educate users about system behavior, temporarily present side-by-side views of unfiltered data versus algorithmically sorted data. This builds trust by proving the system is filtering out noise, rather than arbitrarily hiding vital information. * **Rule 4: Audit for Input Data Bias** -> Before deploying a ranking or machine learning algorithm, systematically audit the training data for demographic, political, or economic skew. An algorithm cannot output neutral results if it is fed predominantly skewed data. * **Rule 5: Conduct External Paired-Testing Audits** -> Regularly test platform outputs by deploying "sock puppet" accounts (automated profiles) that are identical in every way except for one protected variable (e.g., race, gender, OS type) to detect illegal or unethical algorithmic discrimination. * **Rule 6: Protect and Facilitate Collective Auditing** -> Do not use Terms of Service (like the CFAA) to criminalize users or researchers who are attempting to aggregate platform data to detect biases. Support mechanisms where users can share their algorithmic outcomes to hold platforms accountable. ## 5. Pitfalls & Limitations (Anti-Patterns) * **Pitfall:** Designing completely opaque, "black box" algorithms. -> **Why it fails:** Users lack the context to understand why content is hidden. When they inevitably notice missing information, they invent negative "folk theories," blame themselves, or experience visceral anger toward the platform. -> **Warning sign:** Users expressing feelings of being "snubbed" by friends or repeatedly stating they missed major life events of close contacts. * **Pitfall:** Reverting user-selected feed controls automatically. -> **Why it fails:** Platforms like Facebook often allow users to select "Most Recent" but silently revert the feed back to "Top Stories" upon the next login. This undermines user agency and creates a frustrating, adversarial relationship with the interface. -> **Warning sign:** High volumes of user complaints or support tickets asking how to permanently keep a feed in chronological order. * **Pitfall:** Relying on overly broad "Terms of Service" to prevent scraping. -> **Why it fails:** Blanket bans on automated scraping (enforced via laws like the Computer Fraud and Abuse Act) prevent crucial academic and civil rights audits, allowing hidden systemic discrimination to go unchecked. -> **Warning sign:** Security researchers and academics facing legal threats for conducting basic paired-testing on public-facing websites. * **Pitfall:** Assuming algorithms are inherently objective. -> **Why it fails:** Algorithms are optimization mechanisms that amplify the biases present in their training data. If historical data contains discriminatory patterns, the algorithm will scale that discrimination efficiently. -> **Warning sign:** Different demographic groups receiving consistently different search results, ad targeting (e.g., arrest ads), or pricing for identical queries. ## 6. Key Quote / Core Insight "The majority of users are entirely unaware that an invisible algorithm curates their digital reality. When you finally reveal this hidden filtering to them, the reaction is visceral—they feel as though they are waking up in 'The Matrix,' suddenly realizing that their view of the world has been selectively and artificially constructed." ## 7. Additional Resources & References * **Resource:** *Beyond Being There* by Jim Hollan and Scott Stornetta - **Type:** Academic Paper - **Relevance:** Foundational research on the human requirements for telepresence and connection, explaining why asynchronous social feeds effectively meet basic communication needs. * **Resource:** Folk Theories of Thermostats by Willett Kempton - **Type:** Academic Paper - **Relevance:** Provides the framework for understanding how everyday users construct mental models of complex systems (like feed algorithms) based on limited interaction. * **Resource:** *Finding Naked People* by David Forsyth and Margaret Fleck (1996) - **Type:** Academic Paper / Code - **Relevance:** An early, explicit example of computer vision bias where the algorithm failed to recognize darker skin tones, requiring manual heuristic adjustments. * **Resource:** *Big Data: A Report on Algorithmic Systems, Opportunity, and Civil Rights* (Executive Office of the President, May 2016) - **Type:** Government Report - **Relevance:** Establishes algorithmic auditing as a national civil rights priority to prevent automated discrimination in housing, employment, and lending. * **Resource:** Civil Rights Act of 1964 & Fair Housing Act (1968) - **Type:** Legal Frameworks - **Relevance:** The legal basis demonstrating that algorithmic discrimination based on protected classes (race, religion, sex) is illegal, regardless of the algorithm's intent.