APPENDIX B: Empirical Case Studies in Social Media 'Liking' Scandals, Algorithmic Literalism, and Strategic Preference Disruption
Updated: 1 day ago
1. Executive Overview: Supporting the Preference Disruption Thesis
This appendix directly supports the central thesis of this research project:
liking diverse content successfully breaks open a narrow filter bubble and exposes the user to alternate views.
While engagement-driven recommendation architectures tend to confine users within closed mathematical feedback loops, intentional user interaction with out-of-character or opposing content computationally forces recommendation algorithms out of pure exploitation mode. By double-tapping diverse posts, users trigger immediate updates in item-feature vector embeddings, re-map their coordinates within high-dimensional collaborative lookalike clusters, and activate multi-armed bandit exploration mechanisms (such as Upper Confidence Bound and Thompson Sampling)³.
To fully explain why this mechanism works—and why it remains fraught with societal risk—this document draws concise theoretical bridges from Appendix A. As established in academic literature (Pan et al., 2021²; Narayanan, 2023¹), recommender systems operate under Missing-Not-At-Random (MNAR) interaction data and Literal Signal Optimization¹. Because loss functions compress all qualitative human sentiment (irony, courtesy, curiosity, hate-watching) into binary positive feature weights, double-tapping an out-of-distribution post acts as a powerful mathematical command that disrupts feedback-loop selection bias.
However, a striking paradox emerges when evaluating this technical reality against real-world human behavior. While 'liking' diverse content is computationally effective at dismantling information cocoons, public figures, corporate employees, and private citizens face severe societal, professional, and political backlash when caught double-tapping controversial posts. Because both machine learning algorithms and human social observers evaluate 'likes' under literalism—interpreting every explicit double-tap as an unhedged personal endorsement—users are caught in an 'Irony Gap'¹. This document presents ten detailed empirical case studies illustrating this structural tension between technical feed expansion and public PR reality.
2. Theoretical Foundations: Five Appendix A Mechanics Anchoring Appendix B
To explain how strategic double-tapping alters feed architecture without duplicating the lengthy mathematical proofs in Appendix A, this section summarizes five core computational mechanisms that anchor the thesis of Appendix B:
1. Content-Based Vector Expansion & Gradient Steps: When a user double-taps an out-of-character item, the profile feature vector V_u takes an immediate weighted gradient step toward item vector V_i. This expands candidate retrieval algorithms beyond local semantic boundaries to include previously excluded keywords, audio tracks, and visual embeddings (Narayanan, 2023)¹.
2. Lookalike Cluster Re-Mapping in Latent Matrix Space: In collaborative filtering architectures, explicit likes recalculate nearest neighbors in high-dimensional matrix space. This maps the user's profile into external 'lookalike user clusters' who naturally consume opposing viewpoints, drawing in cross-domain recommendations from external peer groups (Pan et al., 2021)².
3. Multi-Armed Bandit Exploration (UCB & Thompson Sampling): Recommender systems balance exploitation (recommending known favorites) with exploration (testing novel items). Double-tapping unexpected content artificially inflates the variance score in Upper Confidence Bound (UCB) algorithms or expands the posterior Beta distribution in Thompson Sampling, forcing the algorithm into an active exploration phase (Qazi et al., 2023)³.
4. Overcoming Linguistic Homogenization Loops: News filtering models often cause 'Linguistic Homogenization' by linking unrelated topics through shared vocabulary (Bilgic & Shapiro, 2021)⁴. Strategic diversification forces the system to decouple topic-specific feature vectors, preventing moderate users from being dragged into political extremes.
5. Pushing Recommendations Along the Pareto-Optimal Front: Filter bubble mitigation can be framed as a bi-objective Pareto optimization problem: max (Personalization Score, Diversity Score). Deliberate engagement with novel topics pushes the feed along the Pareto-optimal frontier, balancing individual relevance with systemic content variety (Qazi et al., 2023)³.
3. Empirical Case Studies: Ten Documented 'Like' Scandals
The following ten case studies document real-world instances where public figures, academics, corporate workers, and students faced formal admonition, resignation, or termination after double-tapping controversial posts. Each case illustrates the clash between user interaction, algorithmic candidate generation, and public literalism.
Case 1: J.K. Rowling (2018) — The 'Clumsy Scroll' Defense
In March 2018, author J.K. Rowling faced widespread social media criticism after fans noticed her Twitter account had liked a tweet referring to transgender women as 'men in dresses.' The incident generated international news coverage regarding her political views⁶.
Public Defense & Outcome: Rowling's spokesperson issued an official statement attributing the double-tap to an accidental physical error, describing it as a 'clumsy and middle-aged moment' where she held her phone incorrectly while scrolling. While her team attempted to frame the interaction as an inadvertent touch, the public treated the like as an explicit endorsement, marking the beginning of years of public scrutiny.
Appendix A Technical Connection: Illustrates Literal Signal Optimization and the Irony Gap¹. The recommendation engine and public observers both interpreted the double-tap as a binary positive preference weight (1), shifting candidate retrieval toward gender-policy embeddings regardless of her physical scroll intent.
Case 2: U.S. Senator Ted Cruz (2017) — Staffer Error Attribution
On September 11, 2017, the official Twitter account of U.S. Senator Ted Cruz favorited a two-minute explicit pornographic video posted by an adult account⁷. The liked post remained visible on his public profile for several hours before being removed, triggering national media coverage and viral online satire.
Public Defense & Outcome: Senator Cruz addressed the scandal directly in a CNN interview, placing blame on a staff member. Cruz stated that multiple staffers had access to the account and that one staffer had 'accidentally hit the wrong button' while browsing from a mobile device.
Appendix A Technical Connection: Demonstrates Lookalike Cluster Re-Mapping in matrix factorization². The explicit favorite instantly recalculated the account's latent nearest neighbors, linking the profile to adult content user clusters and triggering high-variance bandit exploration³.
Case 3: Michael Korenberg (2020) — Institutional Resignation at UBC
In June 2020, Michael Korenberg, Chair of the Board of Governors at the University of British Columbia (UBC), faced intense student protests after an investigative student group uncovered that his Twitter account had liked multiple far-right tweets criticizing Black Lives Matter and comparing racial justice protesters to paramilitary groups⁸.
Public Defense & Outcome: Facing formal condemnation from the UBC Faculty Association and student body, Korenberg issued a public apology acknowledging that his likes caused pain. Within days of the revelations, Korenberg formally resigned as Board Chair.
Appendix A Technical Connection: Highlights Missing-Not-At-Random (MNAR) selection bias and feedback compounding². By double-tapping fringe political tweets, the algorithm narrowed his feed exposure to extreme content vectors, demonstrating how public profile likes trigger real-world institutional fallout.
Case 4: Prof. Geraldine Rauch (2024) — Academic & Political Investigation
In May 2024, Prof. Geraldine Rauch, President of the Technical University of Berlin (TU Berlin), became the subject of intense political scrutiny after liking several controversial posts on X (formerly Twitter) regarding the Israel-Hamas conflict⁹. One liked post featured an image depicting Israeli Prime Minister Benjamin Netanyahu with a swastika.
Public Defense & Outcome: Prof. Rauch apologized publicly, stating she had liked the post due to its accompanying text advocating for a ceasefire without noticing the antisemitic visual imagery. Despite her explanation of an oversight, the Berlin Academic Senate initiated a formal review, and prominent political leaders called for her immediate resignation.
Appendix A Technical Connection: Reflects Linguistic Homogenization (Bilgic & Shapiro, 2021)⁴. Content-based filtering algorithms linked ceasefire text embeddings with extreme visual tags, pulling the user's candidate retrieval feed across overlapping vocabulary boundaries.
Case 5: Mark Hamill (2023) — Celebrity Unliking & Public Apology
In January 2023, actor Mark Hamill faced backlash from Star Wars fans after liking a tweet by J.K. Rowling defending a woman who had lost her job over transphobic comments¹⁰. Fans expressed disappointment, accusing Hamill of supporting anti-trans rhetoric.
Public Defense & Outcome: Hamill quickly unliked the tweet and issued a public clarification on Twitter. He explained that he had liked the tweet's opening statement supporting human rights without reading the full context or understanding the controversial background, stating: 'Ignorance is no excuse, but I had no idea of the context and unknowingly liked a tweet that caused offense.'
Appendix A Technical Connection: Illustrates Multi-Armed Bandit Exploration³. The unexpected interaction artificially boosted the variance parameter in UCB algorithms, prompting the feed to test culture-war topics against his user profile.
Case 6: Marriott International Employee (2018) — Corporate Termination
In January 2018, a social media manager for Marriott International used the hotel chain's official Twitter account to 'like' a tweet from a Tibetan independence group praising Marriott for listing Tibet as a separate country in a customer survey¹¹. The action sparked outrage among Chinese netizens and led the Chinese government to block Marriott's Chinese website and app.
Public Defense & Outcome: Marriott issued an official apology to the Chinese government and immediately terminated the employee responsible. The corporate firing underscored how an employee's single double-tap can trigger severe geopolitical and commercial consequences.
Appendix A Technical Connection: Demonstrates Content-Based Vector Expansion¹. A single explicit interaction shifted the corporate account's feature vector toward geopolitical controversy tags, triggering massive external visibility and automated candidate propagation.
Case 7: Jennifer Bowes (2023) — Parliamentary Admonition
In November 2023, Jennifer Bowes, a Member of the Legislative Assembly (MLA) in Saskatchewan, Canada, faced severe criticism from political opponents after liking an Instagram post containing the slogan 'From the river to the sea, Palestine will be free'¹².
Public Defense & Outcome: Opposing lawmakers accused Bowes of endorsing antisemitic language. Bowes removed the like and issued a public statement clarifying that her intention was to express sympathy for civilian suffering, apologizing for the harm caused by the post's wording.
Appendix A Technical Connection: Connects to Collaborative Lookalike Matrix Re-Mapping². Liking regional activist posts re-mapped her account into political activist user clusters, expanding feed candidate retrieval across sensitive geopolitical categories.
Case 8: Albany High School Students (2017) — First Amendment Litigation
In 2017, several high school students at Albany High School in California were suspended after liking and commenting on a private Instagram account that posted racist memes targeting Black students and staff¹³.
Public Defense & Outcome: The suspended students filed a federal civil rights lawsuit against the school district, arguing that double-tapping a post created by another individual constituted passive online expression protected by the First Amendment. Federal courts ruled that school districts have authority to discipline students for online 'likes' if the content causes substantial disruption to the educational environment.
Appendix A Technical Connection: Demonstrates the legal enforcement of Algorithmic Literalism¹. The judicial system treated 'likes' as active behavioral adoption rather than passive observation, echoing how loss functions treat double-taps as explicit preference vectors.
Case 9: Startup Employee (2024) — LinkedIn Workplace Termination
In September 2024, an employee at a health tech startup went viral after revealing she was summarily fired by her CEO specifically for double-tapping a LinkedIn post that discussed toxic workplace management¹⁴.
Public Defense & Outcome: The employee stated she liked the post as a gesture of general professional solidarity regarding industry culture. The CEO confronted her, asserting that 'liking' the post demonstrated disloyalty to the firm, resulting in immediate employment termination.
Appendix A Technical Connection: Reflects Sequential Inverse Propensity Scoring (SIPS) dynamics². The explicit interaction signaled a departure from historical corporate engagement baseline, prompting an abrupt employer reaction mirroring automated anomaly detection.
Case 10: Gemma Chan & Hollywood 'Sub-Liking' (2021) — Celebrity PR Cycles
In 2021, actress Gemma Chan and several Hollywood figures became embroiled in viral media cycles after fans noticed them 'sub-liking' tweets that subtly criticized film directors, co-stars, or studio decisions¹⁰.
Public Defense & Outcome: Public relations representatives frequently managed the fallout by un-liking the tweets and attributing the interactions to accidental touchscreen swipes while scrolling, demonstrating how media outlets actively monitor public figure 'like' tabs to manufacture feud narratives.
Appendix A Technical Connection: Highlights the Pareto Multi-Objective Optimization trade-off³. While sub-liking injected subtle diversity into the actor's personal feed, public profile exposure violated reputational constraints, forcing forced un-liking and manual feed dampening.
4. Synthesis Matrix: Public 'Like' Scandals & Appendix A Technical Connections
Table B1 synthesizes the ten empirical case studies, mapping each public figure's domain, offered defense, real-world outcome, and the corresponding computational mechanism established in Appendix A.
Case Study | Domain | Public Defense | Real-World Outcome | Appendix A Computational Anchor |
J.K. Rowling (2018) | Author / Culture | 'Clumsy middle-aged moment' phone scroll | Viral media backlash; public anti-trans framing | Literal Signal Optimization; Irony Gap (Narayanan, 2023) |
Ted Cruz (2017) | U.S. Senator / Politics | Staffer error; 'wrong button pressed' | National TV news crisis; viral online satire | Lookalike Cluster Re-Mapping; UCB Exploration (Pan, 2021) |
M. Korenberg (2020) | UBC Board Chair / Higher Ed | Apology for causing hurt & pain | Resigned as Board Chair under student pressure | MNAR Selection Bias; Feedback Compounding (Pan et al.) |
Prof. G. Rauch (2024) | TU Berlin President / Academia | Did not notice image; focused on text | Formal political investigation; resignation calls | Linguistic Homogenization Loop (Bilgic & Shapiro, 2021) |
Mark Hamill (2023) | Actor / Entertainment | Unliked tweet; 'Ignorance is no excuse' | Public clarification & fan reconciliation | Multi-Armed Bandit Exploration Surge (Qazi et al., 2023) |
Marriott Staff (2018) | Corporate Employee | None granted by employer | Employee fired; website blocked in China | Content-Based Vector Gradient Step (Narayanan, 2023) |
Jennifer Bowes (2023) | MLA Lawmaker / Govt | Sympathy for civilians; removed like | Public parliamentary apology & statement | Collaborative Matrix Re-Mapping to Activist Clusters |
Albany Students (2017) | High School / Education | Passive expression protected by 1st Amend. | School suspension; federal lawsuit dismissed | Judicial Enforcement of Algorithmic Literalism |
Startup Staff (2024) | Tech Worker / Employment | General solidarity with workplace post | Summary termination by company CEO | SIPS Anomaly Signal from Baseline (Pan et al., 2021) |
Gemma Chan (2021) | Hollywood Actor / Media | Accidental touchscreen swipe scrolling | Viral PR feud cycles; forced un-liking | Pareto Front Trade-off: Personalization vs PR Risk |
5. Strategic Synthesis: Technical Utility vs. Public PR Reality
Synthesizing the empirical evidence in Appendix B with the mathematical foundations of Appendix A leads to a clear conclusion: intentional interaction with diverse content is a powerful, mathematically validated method for breaking filter bubbles, but public account visibility creates severe reputational risk.
From a control-theoretic standpoint (Mollabagher & Naghizadeh, 2025)⁵, users who adopt an Adaptive Reactive Policy—dynamically throttling interactions when feeds drift while strategically liking out-of-distribution content—can bound long-term opinion drift and reclaim feed diversity. Double-tapping novel perspectives successfully forces bandit algorithms out of pure exploitation mode, triggering UCB variance boosts³ and expanding vector embeddings across diverse semantic domains¹.
However, because social media platforms expose 'Likes' on public profile tabs, human observers evaluate social media interactions under the exact same Literal Signal Optimization¹ as machine learning algorithms. In the public square, a double-tap is treated as an unhedged political endorsement. As demonstrated across all ten case studies, public figures cannot easily defend themselves by claiming 'I was just diversifying my algorithm.' To navigate this paradox, users seeking feed diversity must pair strategic double-tapping with adaptive dampening (such as selective scrolling or using private topic-mute controls) to maintain both algorithmic feed diversity and social reputational stability.
6. Notes
Part A: Academic Foundations & Appendix A Technical Connections
1. Arvind Narayanan, 'Understanding Social Media Recommendation Algorithms,' Knight First Amendment Institute at Columbia University (March 9, 2023): 12–15. For formal definitions of item-feature gradient steps, vector embeddings, and engagement loss functions, see Appendix A, Section 8, s.v. 'Content-Based Vector Expansion' and 'Literal Signal Optimization.'
2. Weishen Pan et al., 'Correcting the User Feedback-Loop Bias for Recommendation Systems,' arXiv preprint arXiv:2109.06037 (September 13, 2021): 2–4. For statistical definitions of feedback compounding, matrix re-mapping, and de-biasing, see Appendix A, Section 8, s.v. 'Missing-Not-At-Random (MNAR)' and 'Sequential Inverse Propensity Scoring (SIPS).'
3. Qazi Mohammad Areeb et al., 'Filter Bubbles in Recommender Systems: Fact or Fallacy — A Systematic Review,' arXiv preprint arXiv:2307.01221 (July 2, 2023): 4–6. For mathematical definitions of variance expansion and multi-objective trade-offs, see Appendix A, Section 8, s.v. 'Multi-Armed Bandit Exploration (UCB & Thompson Sampling)' and 'Pareto Optimization Problem.'
4. Mustafa Bilgic et al., 'The Interaction Between Political Typology and Filter Bubbles in News Filter Algorithms,' National Science Foundation Award #1927407, Illinois Institute of Technology News (November 1, 2021): 3–5. For details on vocabulary linkage and sentiment drift, see Appendix A, Section 8, s.v. 'Linguistic Homogenization.'
5. Atefeh Mollabagher and Parinaz Naghizadeh, 'The Feedback Loop Between Recommendation Systems and Reactive Users,' arXiv preprint arXiv:2504.07105v1 (March 14, 2025): 3–6. For control-theoretic opinion dynamics and counter-mechanisms, see Appendix A, Section 5 and Section 8, s.v. 'Adaptive Decreasing Policy.'
Part B: Empirical Case Study Sources
6. Newsweek, 'J.K. Rowling Accused of Transphobia After Liking Controversial Tweet,' Newsweek Culture (March 22, 2018). See also Appendix B, Section 3, Case 1.
7. The Guardian, 'Ted Cruz Twitter Account Likes Pornographic Tweet,' The Guardian US News (September 12, 2017); South China Morning Post, ''It was not me': Ted Cruz defends accidental porn like from Twitter account in bizarre CNN interview,' SCMP Politics (September 13, 2017). See also Appendix B, Section 3, Case 2.
8. The Tyee, 'UBC Board Chair Resigns Following Backlash Over Liked Tweets,' The Tyee Education News (June 19, 2020). See also Appendix B, Section 3, Case 3.
9. Ynet News, 'Berlin University President Under Investigation for Liking Antisemitic Posts,' Ynet Jewish World (May 29, 2024). See also Appendix B, Section 3, Case 4.
10. Out Magazine, 'Mark Hamill Clarifies J.K. Rowling Tweet Likes and Apologizes to Fans,' Out Celebs (January 25, 2023). See also Appendix B, Section 3, Cases 5 & 10.
11. Tibetan Journal, 'Marriott Fires Social Media Manager Over Liked Tibet Tweet,' Tibetan Journal Business (January 13, 2018). See also Appendix B, Section 3, Case 6.
12. Battlefords NOW, 'Saskatchewan MLA Issues Public Apology for Instagram Like,' Battlefords NOW Politics (November 28, 2023). See also Appendix B, Section 3, Case 7.
13. The Business Journal, 'High School Students Sue District Over Suspensions for Instagram Likes,' Legal Trends (May 14, 2017). See also Appendix B, Section 3, Case 8.
14. LiveMint, 'Tech Startup Employee Claims She Was Fired for Liking LinkedIn Post,' LiveMint Trends (September 11, 2024). See also Appendix B, Section 3, Case 9.
7. Works Cited
Areeb, Qazi Mohammad, Mohammad Nadeem, Shahab Saquib Sohail, Raza Imam, Faiyaz Doctor, Yassine Himeur, Amir Hussain, and Abbes Amira. 'Filter Bubbles in Recommender Systems: Fact or Fallacy — A Systematic Review.' arXiv preprint arXiv:2307.01221 (July 2, 2023). https://arxiv.org/abs/2307.01221.
Bilgic, Mustafa, Matthew Shapiro, Ping Liu, Aron Culotta, and Karthik Shivaram. 'The Interaction Between Political Typology and Filter Bubbles in News Filter Algorithms.' National Science Foundation Award #1927407. Illinois Institute of Technology News (November 1, 2021). https://www.iit.edu/news/bias-bubble-new-research-shows-news-filter-algorithms-reinforce-political-biases.
Mollabagher, Atefeh, and Parinaz Naghizadeh. 'The Feedback Loop Between Recommendation Systems and Reactive Users.' arXiv preprint arXiv:2504.07105v1 (March 14, 2025). https://arxiv.org/abs/2504.07105v1.
Narayanan, Arvind. 'Understanding Social Media Recommendation Algorithms.' Knight First Amendment Institute at Columbia University (March 9, 2023). https://knightcolumbia.org/content/understanding-social-media-recommendation-algorithms.
Pan, Weishen, Sen Cui, Hongyi Wen, Kun Chen, Changshui Zhang, and Fei Wang. 'Correcting the User Feedback-Loop Bias for Recommendation Systems.' arXiv preprint arXiv:2109.06037 (September 13, 2021). https://arxiv.org/pdf/2109.06037.
Comments