Pith. sign in

REVIEW 4 major objections 8 minor 62 references

YouTube Recommendations Reinforce Negative Emotions: Auditing Algorithmic Bias with Emotionally-Agentic Sock Puppets

T0 review · 4 major / 8 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Sock-puppet audit shows YouTube recognizes and amplifies emotional preferences, especially negative ones.

desk verdict A worthwhile audit with a new personalized-vs-contextual comparison, but the causal claim about emotion recognition outruns the evidence and the protocol documentation has unresolved gaps. read the letter →

arxiv 2501.15048 v1 pith:BUPLYV2H submitted 2025-01-25 cs.SI cs.CY

Hussam Habib , Rishab Nithyanand This is my paper · ORCID
classification cs.SIcs.CY
keywords sock-puppetauditrecommendationalgorithmemotionalbiasYouTubealgorithmicamplificationfilterbubblerevealedpreferencessentimentanalysis
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tests whether YouTube's recommendation algorithm learns and reinforces users' emotional preferences. Using automated sock-puppet accounts, each assigned a fixed emotion, the authors let each bot pick videos from the Up Next list that the bot's transcript-based utility function says best matches its assigned emotion, then compare the recommendations the bot receives against control bots that pick randomly. They find that bots revealing negative emotions—anger, grievance, negativity, and group identification—receive recommendations with both a higher average emotional utility and a stronger rank-utility correlation than the controls, and that this reinforcement grows with continued engagement and spills over to unrelated videos. The authors conclude that YouTube recognizes emotional revealed preferences and amplifies emotional filter bubbles, particularly for negative content.

What carries the argument

The core mechanism is a sock-puppet audit: simulated YouTube accounts with a scripted emotional preference choose videos from the Up Next list that maximize a transcript-based utility function (using LIWC and VADER to score anger, grievance, group identification, negativity, and positivity), thereby revealing that preference to the algorithm. The paper operationalizes reinforcement as utility prevalence (the mean utility of all recommended videos) and utility prominence (the Spearman correlation between a recommendation's rank and its utility), and measures amplification through a moderation analysis of reinforcement over recommendation depth and through a pre-defined video set that isolates personalization from context.

What would settle it

Run the same audit but assign each bot pairs of seed videos matched on topic, channel, and rough topic niche yet differing sharply in transcript-predicted emotional intensity; if the utility premium for anger and grievance disappears when these attributes are held fixed, the algorithm is responding to correlated content signals rather than to emotional preference itself.

Watch

Extended reading notes

Core claim

The central claim is that YouTube's recommendation system recognizes emotional preferences revealed through viewing choices and reinforces them, so that users who engage with negative emotional content are served increasingly more of it, in both prevalence (average alignment) and prominence (ranking). The effect is largest for anger, grievance, and group identification, is absent for a meaningless control preference based on letter frequency, grows with recommendation depth, and persists on pre-selected videos the user never watched. A secondary, surprising finding is that contextual recommendations (computed without watch history) often outperform personalized recommendations in reinforcing the stated preference, suggesting that collective user behavior encoded in collaborative filtering is a strong driver even without personalization.

Load-bearing premise

The load-bearing premise is that the stronger recommendations given to preference-revealing bots come from the algorithm recognizing the assigned emotion itself, not from the emotion being correlated with some other attribute (like a topic, creator, or niche) of the videos the bots happened to choose.

Editorial extensions

If this is right

  • Users who engage with angry or aggrieved content will receive progressively more such content, with the effect intensifying over time.
  • This reinforcement persists even on videos the user has never watched, implying an emotional filter bubble that extends beyond the immediate viewing context.
  • Because contextual recommendations reinforce at least as strongly as personalized ones, even a signed-out user can be steered toward emotionally aligned content based on the first video they watch.
  • Reinforcement requires a collective user base that demonstrably clicks on such content, since meaningless or rare preferences are not amplified.
  • The algorithm's tilt toward exploitation over exploration means the recommender prioritizes short-term alignment over content diversity.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The asymmetry between negative and positive emotions suggests the same engagement-optimization mechanism that produces convenience may systematically increase exposure to anger and grievance for real users, potentially affecting mood and polarization over longer periods than the study observes.
  • Because contextual recommendations alone already produce high reinforcement, users may not need a long watch history to be pulled into emotional echo chambers; the platform's collective behavior does much of the work.
  • A direct probe of the latent-treatment confound would be to hold topic, creator, and niche fixed across high- and low-emotion videos; if the utility premium disappears under that control, the claim of emotional recognition would need to be weakened.
  • The findings should generalize to any platform that optimizes for watch time or similar engagement metrics, so similar audits could be run on TikTok, Instagram Reels, or Twitch to test whether the pattern is specific to YouTube.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 8 minor

Summary. This paper reports a sock-puppet audit of YouTube's recommendation algorithm. The authors create YouTube accounts assigned to one of seven 'preferences' (anger, grievance, group identification, negativity, positivity, a meaningless H-frequency preference, and a random-selection control), have each bot repeatedly select from Up Next recommendations the video maximizing a transcript-based utility function for its assigned preference, and record personalized and contextual recommendations along the way. They compare the utility prevalence (mean utility score) and utility prominence (rank-utility Spearman correlation) of recommendations received by preference-revealing bots versus control bots across four seed domains (News, Fitness, Gaming, Random). The paper claims that YouTube recognizes and reinforces emotional preferences, particularly negative ones; that reinforcement intensifies over time and persists in recommendations for unrelated videos; and that contextual recommendations are often more reinforcing than personalized ones. The central H1 finding is reportedly supported, H2 is rejected, and H3 receives mixed support.

Significance. The strength of the paper is its ambitious, controlled measurement design: large-scale sock-puppet creation via the InnerTube API, random seed sampling following prior methods, and a validation appendix showing API-desktop recommendation similarity. If the causal interpretation were established, the result that YouTube amplifies negative emotional content would be an important contribution to the algorithmic-audit and filter-bubble literature. However, the paper's core claim—that YouTube recognizes the assigned emotional preference rather than merely responding to the selected videos' non-emotional features—is not identified in the current analysis. The unresolved placeholders and internal inconsistencies in the sample size and collection interval, together with the significant reinforcement of the meaningless H-frequency control, mean that the present evidence does not yet support the strong conclusions in the abstract. The contribution is therefore potentially significant but currently contingent.

major comments (4)
  1. [§3.0, §3.1.2, §3.3.2] The experiment parameters are internally inconsistent. §3.1.2 states that '8 videos are randomly selected as seeds,' but §3.0 states '560 sock puppets (7 preferences × 4 seed categories × 10 seed videos from each category × 2 replications),' which is arithmetically consistent only with 10 seeds per category (7×4×10×2 = 560), not 8 (which would yield 448). Additionally, §3.0 contains unresolved placeholders 'TTT times, using XX different preferences, YY seed videos, and ZZ repetitions,' §3.1.2 repeats 'XX videos,' and §3.2 reports 'a YY % rate of transcripts.' Finally, the collection interval for predefined videos is M=10 in §3.0 and M=20 in §3.3.2. These inconsistencies prevent the reader from knowing the actual sample size, collection timing, and transcript coverage, which are essential for assessing statistical power and reproducibility.
  2. [§7.4.1, §5.2, Table 2] The central causal claim that YouTube recognizes and reinforces emotional preferences is not identified because of the latent treatment effect the authors themselves acknowledge. The observed utility difference between treatment and control can be explained by the algorithm recommending videos similar to the selected videos' sub-topics, creators, or niches; such similarity-based recommendations would also produce higher utility on the metric used for selection. This alternative is strongly supported by the paper's own H2 result: contextual recommendations (which have no user history) are usually more reinforcing than personalized ones (Table 2, e.g., Anger in News: -7.94%, Cohen's d = -0.18), showing that the currently viewed video alone, without any learned user embedding, drives the emotional alignment. To support the emotion-recognition interpretation, the authors would need to show, for example, that the treatment effect remains after including video identity or transcript-feature fixed effects, or that a control selecting videos matched on non-emotional features does not produce the same utility gain.
  3. [Table 1, §4.2.1, §4.3] The 'meaningless' H-Frequency control exhibits significant reinforcement in the Random domain (Mean % Diff = 20, Cohen's d = 0.48, with p < 0.05). The paper acknowledges this exception in §4.2.1 ('except in random selections'), but the exception directly undercuts the claim that reinforcement is specific to meaningful emotional preferences. A purely arbitrary selection rule—proportion of the letter 'h' in transcripts—can produce recommendation drift in at least one domain, so the data are equally consistent with the algorithm responding to any consistent selection pattern rather than to emotional meaning. The takeaway in §4.3 that reinforcement implies recognition ('preferences with insufficient representation in the broader user base may not be actively reinforced') does not explain why a nonsensical letter-frequency preference with presumably zero user-base representation is reinforced in Random while not in other domains.
  4. [§4.2.2, Table 1] The 'prominence' result relies on Spearman correlations that are extremely small in magnitude: most significant treatment correlations are between 0.03 and 0.13 (e.g., Anger in Fitness: r = 0.07; Group Identity in Gaming: r = 0.10), while H1.2 is tested as whether the treatment correlation is stronger than the control. Given the large number of recommendation observations, p < 0.05 is not informative about practical prominence, and no correction is reported across the 28 preference–domain comparisons in Table 1. The abstract's phrase 'increasing their prevalence and prominence' overstates the strength of the rank-utility relationship.
minor comments (8)
  1. [Abstract] The phrase 'Our findings reveal reveal that' contains a duplicated word and should be corrected.
  2. [§3.1.1] The word 'emptional' in 'six emptional and one control preference' should be 'emotional.'
  3. [§3.3.1] The word 'simultenously' should be 'simultaneously.'
  4. [Figure 3] The spelling 'Greivance' should be 'Grievance,' and the label '0 th' appears incomplete.
  5. [§2.3.1] The sentence 'experiment often traverse the recommendations tree on a predefined path' has subject-verb disagreement; it should be 'experiments often traverse the recommendation tree...'
  6. [Box 1] In the Positivity entry, 'similar to strong negative emotions and sentient' should likely be 'sentiment.'
  7. [§7.4.2] The phrase 'allowing us to commenting on the behavior' is ungrammatical; it should be 'allowing us to comment on the behavior.'
  8. [Appendix A.1] The 'TheirTube' project is referenced only by a URL in a footnote; a proper citation would improve reproducibility.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: utility functions define an operationalized preference, no fitted parameters are renamed as predictions, and the core result has independent counterfactual support, though the acknowledged latent treatment effect limits the causal interpretation.

full rationale

The paper's derivation chain is an experiment rather than a fitted model, so the main circularity patterns do not apply. Sock puppets are assigned preferences p operationalized as utility functions u_p(v) over transcripts (Box 1), treatment bots select v_{t+1} = argmax_{v in R(v_t)} u_p(v) (Section 3.2), and reinforcement is measured by comparing the u_p values of the resulting Up Next recommendations against random-selection controls (Section 4.1). Reusing the same u_p for selection and outcome is a consistent operationalization of 'preference,' not a fitted parameter reintroduced as a prediction, and no equation in the paper makes the reported differences true by construction. The central claim also has independent content: Section 6.2.2 tests H3.2 on recommendations for pre-defined videos that were not selected for emotional alignment, and the random-seed batches plus the H-Frequency control provide counterfactual comparisons. The paper candidly acknowledges the latent treatment effect in Section 7.4.1, noting that observed reinforcement may reflect correlations with sub-topics, creators, or niches rather than the assigned emotional preference; this is a real identification threat, and the surprising H2 result that contextual recommendations are often more reinforcing than personalized ones reinforces that concern. However, identification threats and overinterpretation are validity limitations, not circular reductions of the derivation to its own inputs. The self-citations present ([12], [26], [27]) are used as methodological precedent and interpretive background, and none is load-bearing for the main quantitative result. Accordingly, no significant circularity is found; the low score reflects minor self-citations and the unresolved selection–similarity confound rather than any construct-level circularity.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

The central claim rests on several domain assumptions: that transcript-based LIWC/VADER scores capture emotions, that the InnerTube API mirrors desktop recommendations, that random-selecting bots are a valid counterfactual, and that argmax video selection reveals a stable preference. These are not fitted parameters, but they are unverified or partially verified premises. The paper's own limitations section (7.4) acknowledges the latent treatment effect and the limited interaction signals, which further underdetermine the causal attribution to emotion.

free parameters (3)
  • Number of seed videos per category = 8 or 10 (text inconsistent)
    Section 3.1 states 'From each seed category, 8 videos are randomly selected as seeds' but the total 560 calculation uses 10 seed videos per category, which would give 448 with 8; this affects sample size and statistical power.
  • Recommendation collection interval M = 10 then 20
    Section 3 Overview says 'after every M = 10 steps', while Section 3.3.2 says 'After every M steps (M = 20)'; this changes how many predefined-video recommendation sets are collected.
  • Transcript coverage rate = YY% placeholder
    Section 3.2 reports a placeholder for the rate of videos with usable transcripts; the filtering of videos without English transcripts changes the recommendation universe and could bias utility comparisons.
assumptions (6)
  • domain assumption YouTube's recommendation algorithm optimizes engagement and infers preferences from watch history (Covington et al., 2016 model)
    Section 2.1 relies on this to argue that revealed preferences shape recommendations.
  • domain assumption Transcript text is a sufficient proxy for a video's emotional content
    Box 1 and Section 3.1.1 define all utility functions over transcripts; non-verbal emotional signals are ignored (limitation in 7.4.4).
  • domain assumption LIWC and VADER scores accurately measure the assigned emotional preferences
    Box 1 uses these tools to define anger, grievance, group identification, negativity, and positivity; misclassification is acknowledged in 7.4.4.
  • domain assumption The InnerTube API serves recommendations consistent with the desktop YouTube interface
    Appendix A.1 attempts to validate this, but the validation uses only three personas and small samples.
  • ad hoc to paper Random selection by control bots provides a valid counterfactual baseline
    Section 4.1 defines the baseline as random-selecting control bots; if random selection changes the content domain in ways that correlate with utility, the baseline is biased.
  • ad hoc to paper The sock puppet's argmax selection reveals a stable, singular emotional preference
    Section 3.2 assumes repeated selection of highest-utility videos communicates a clear preference; real users have mixed preferences.

how reviews work

0 comments
Cite this review

Pith. "Pith review of YouTube Recommendations Reinforce Negative Emotions: Auditing Algorithmic Bias with Emotionally-Agentic Sock Puppets." pith.science (2026). https://pith.science/paper/BUPLYV2H

@misc{pith2026250115048,
  author       = {Pith},
  title        = {Pith review of: YouTube Recommendations Reinforce Negative Emotions: Auditing Algorithmic Bias with Emotionally-Agentic Sock Puppets},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BUPLYV2H}},
  note         = {Machine review of arXiv:2501.15048}
}
read the original abstract

Personalized recommendation algorithms, like those on YouTube, significantly shape online content consumption. These systems aim to maximize engagement by learning users' preferences and aligning content accordingly but may unintentionally reinforce impulsive and emotional biases. Using a sock-puppet audit methodology, this study examines YouTube's capacity to recognize and reinforce emotional preferences. Simulated user accounts with assigned emotional preferences navigate the platform, selecting videos that align with their assigned preferences and recording subsequent recommendations. Our findings reveal reveal that YouTube amplifies negative emotions, such as anger and grievance, by increasing their prevalence and prominence in recommendations. This reinforcement intensifies over time and persists across contexts. Surprisingly, contextual recommendations often exceed personalized ones in reinforcing emotional alignment. These findings suggest the algorithm amplifies user biases, contributing to emotional filter bubbles and raising concerns about user well-being and societal impacts. The study emphasizes the need for balancing personalization with content diversity and user agency.

Figures

Figures reproduced from arXiv: 2501.15048 by the authors.

Figure 1
Figure 1. Distribution of utility values derived from Up Next recommendations for preference-aligned (treatment) and control sock puppets, grouped by seed domain. Note that some utility values have been scaled independently better visual representation within the graph. For exact values refer to table 1. recommendations to the reinforcement within contextual recom￾mendations for the same set of preference-aligned videos. Here… view at source ↗
Figure 2
Figure 2. Distribution of utility values from personalized and contextual Up Next recommendations. We aggregate recommendations across all domains. Note that some utility values have been scaled independently better visual representation within the graph. For exact values refer to table 2. summarized in [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Mean utility within Up Next recommendations of predefined videos for preference-revealing (treatment) and random-selecting (control) sock puppets. The values are shown as the percentage difference from the first control-observed recommendations. All values at the 0th essentially show utility within contextual recommendations. 6.2.2 Reinforcement persists across recommendations for predefined videos not selected to o… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

62 extracted references · 35 canonical work pages

  1. [1]

    [n. d.]. S&T TVTP The Application of Risk Assessment Tools in the Criminal Justice and Rehabilitation Process Literature Review | Homeland Security. https://www.dhs.gov/archive/science-and- technology/publication/st-tvtp-application-risk-assessment- tools-criminal-justice-and-rehabilitation-process

    work page

  2. [2]

    Adam Carter, Peter Clutton, and Colin Klein

    Mark Alfano, Amir Ebrahimi Fard, J. Adam Carter, Peter Clutton, and Colin Klein. 2021. Technologically scaffolded atypical cognition: the case of YouTube’s recommender system. Synthese 199, 1 (Dec. 2021), 835–858. https://doi.org/ 10.1007/s11229-020-02724-x

    doi:10.1007/s11229-020-02724-x 2021

  3. [3]

    I Don’t Even Remember What I Read

    Amanda Baughan, Mingrui Ray Zhang, Raveena Rao, Kai Lukoff, Anastasia Schaadhardt, Lisa D. Butler, and Alexis Hiniker. 2022. “I Don’t Even Remember What I Read”: How Design Influences Dissociation on Social Media. InProceedings of the 2022 CHI Conference on Human Factors in Computing Systems (CHI ’22) . Association for Computing Machinery, New York, NY, U...

    arXiv 2022

  4. [4]

    Baumeister, Ellen Bratslavsky, Catrin Finkenauer, and Kathleen D

    Roy F. Baumeister, Ellen Bratslavsky, Catrin Finkenauer, and Kathleen D. Vohs

    work page

  5. [5]

    Jonah Berger. 2011. Arousal Increases Social Transmission of Information. Psychological Science 22, 7 (July 2011), 891–893. https://doi.org/10.1177/ 0956797611413294 Publisher: SAGE Publications Inc

    work page 2011

  6. [6]

    Jonah Berger and Katherine L. Milkman. 2012. What Makes Online Content Viral? Journal of Marketing Research 49, 2 (April 2012), 192–205. https://doi. org/10.1509/jmr.10.0353 Publisher: SAGE Publications Inc

    doi:10.1509/jmr.10.0353 2012

  7. [7]

    Ryan Boyd, Ashwini Ashokkumar, Sarah Seraj, and James Pennebaker. 2022. The Development and Psychometric Properties of LIWC-22 . https://doi.org/10. 13140/RG.2.2.23890.43205

    arXiv 2022

  8. [8]

    Brady, M

    William J. Brady, M. J. Crockett, and Jay J. Van Bavel. 2020. The MAD Model of Moral Contagion: The Role of Motivation, Attention, and Design in the Spread of Moralized Content Online. Perspectives on Psychological Science 15, 4 (July 2020), 978–1010. https://doi.org/10.1177/1745691620917336 Publisher: SAGE Publications Inc

    doi:10.1177/1745691620917336 2020

Show all 62 references
  1. [9]

    Brady, Julian A

    William J. Brady, Julian A. Wills, John T. Jost, Joshua A. Tucker, and Jay J. Van Bavel. 2017. Emotion shapes the diffusion of moralized content in social networks. Proceedings of the National Academy of Sciences 114, 28 (July 2017), 7313–7318. https://doi.org/10.1073/pnas.161...

    2017 doi

  2. [10]

    Brown, James Bisbee, Angela Lai, Richard Bonneau, Jonathan Nagler, and Joshua A

    Megan A. Brown, James Bisbee, Angela Lai, Richard Bonneau, Jonathan Nagler, and Joshua A. Tucker. 2022. Echo Chambers, Rabbit Holes, and Algorithmic Bias: How YouTube Recommends Content to Real Users. https://doi.org/ 10.2139/ssrn.4114905

    2022 doi

  3. [11]

    Jordan Carpenter, William Brady, Molly Crockett, Rene Weber, and Walter Sinnott-Armstrong. 2020. Political Polarization and Moral Outrage on Social Media. Connecticut Law Review 52 (2020), 1107. https://heinonline.org/H OL/Page?handle=hein.journals/conlr52&id=1134&div=&collection=

    2020

  4. [12]

    Sarmad Chandio, Muhammad Daniyal Pirwani Dar, and Rishab Nithyanand. 2024. How Audit Methods Impact Our Understanding of YouTube’s Recommendation Systems. Proceedings of the International AAAI Conference on Web and Social Media 18 (May 2024), 241–253. https://doi.org/10.1609/i...

    2024 doi

  5. [13]

    Chen, Brendan Nyhan, Jason Reifler, Ronald E

    Annie Y. Chen, Brendan Nyhan, Jason Reifler, Ronald E. Robertson, and Christo Wilson. 2023. Subscriptions and external links help drive resentful users to alternative and extremist YouTube channels. Science Advances 9, 35 (Aug. 2023), eadd8080. https://doi.org/10.1126/sciadv.a...

    2023 doi

  6. [14]

    Nicoleta Corbu, Alina Bârgăoanu, Raluca Buturoiu, and Oana S, tefănit,ă. 2020. Does fake news lead to more engaging effects on social media? Evidence from Romania. Communications 45, s1 (Nov. 2020), 694–717. https://doi.org/10. 1515/commun-2019-0152 Publisher: De Gruyter Mouton

    2020

  7. [15]

    Emily Corner, Paul Gill, Ronald Schouten, and Frank Farnham. 2018. Men- tal Disorders, Personality Traits, and Grievance-Fueled Targeted Violence: The Evidence Base and Implications for Research and Practice. Jour- nal of Personality Assessment 100, 5 (Sept. 2018), 459–470. ht...

    2018

  8. [16]

    Paul Covington, Jay Adams, and Emre Sargin. 2016. Deep Neural Networks for YouTube Recommendations. In Proceedings of the 10th ACM Conference on Recommender Systems . ACM, Boston Massachusetts USA, 191–198. https: //doi.org/10.1145/2959100.2959190

    2016

  9. [17]

    Crockett

    Molly J. Crockett. 2017. Moral outrage in the digital age. Nature human be- haviour 1, 11 (2017), 769–771. https://www.nature.com/articles/s41562- 017-0213-3 Publisher: Nature Publishing Group UK London

    2017

  10. [18]

    Tom Cunningham. 2023. Ranking by Engagement | Tom Cunningham – Tom Cunningham. https://tecunningham.github.io/posts/2023-04-28-rank ing-by-engagement.html

    2023

  11. [19]

    Tom Cunningham, Sana Pandey, Leif Sigerson, Jonathan Stray, Jeff Allen, Bonnie Barrilleaux, Ravi Iyer, Smitha Milli, Mohit Kothari, and Behnam Rezaei. 2024. What We Know About Using Non-Engagement Signals in Content Ranking. https://doi.org/10.48550/arXiv.2402.06831 arXiv:2402...

    2024 doi arXiv

  12. [20]

    Ernesto de León and Damian Trilling. 2021. A Sadness Bias in Political News Sharing? The Role of Discrete Emotions in the Engagement and Dissemina- tion of Political News on Facebook. Social Media + Society 7, 4 (Oct. 2021), 20563051211059710. https://doi.org/10.1177/205630512...

    2021 doi

  13. [21]

    Michela Del Vicario, Gianna Vivaldo, Alessandro Bessi, Fabiana Zollo, Antonio Scala, Guido Caldarelli, and Walter Quattrociocchi. 2016. Echo Chambers: Emo- tional Contagion and Group Polarization on Facebook. Scientific Reports 6, 1 (Dec. 2016), 37825. https://doi.org/10.1038/...

    2016 doi

  14. [22]

    Ullrich K. H. Ecker, Stephan Lewandowsky, John Cook, Philipp Schmid, Lisa K. Fazio, Nadia Brashier, Panayiota Kendeou, Emily K. Vraga, and Michelle A. Amazeen. 2022. The psychological drivers of misinformation belief and its resistance to correction. Nature Reviews Psychology ...

    2022 doi

  15. [23]

    Rui Fan, Jichang Zhao, Yan Chen, and Ke Xu. 2014. Anger Is More Influential than Joy: Sentiment Correlation in Weibo. PLOS ONE 9, 10 (Oct. 2014), e110184. https://doi.org/10.1371/journal.pone.0110184 Publisher: Public Library of Science

    2014 doi

  16. [24]

    Amit Goldenberg and Robb Willer. 2023. Amplification of emotion on social media. Nature Human Behaviour 7, 6 (June 2023), 845–846. https://doi.org/ 10.1038/s41562-023-01604-x Publisher: Nature Publishing Group

    2023 doi

  17. [25]

    Raj Kumar Gupta and Yinping Yang. 2019. Predicting and Understanding News Social Popularity with Emotional Salience Features. In Proceedings of the 27th ACM International Conference on Multimedia (MM ’19) . Association for Com- puting Machinery, New York, NY, USA, 139–147. htt...

    2019

  18. [26]

    Hussam Habib, Padmini Srinivasan, and Rishab Nithyanand. 2022. Making a Radical Misogynist: How Online Social Engagement with the Manosphere Influences Traits of Radicalization. Proc. ACM Hum.-Comput. Interact. 6, CSCW2 (Nov. 2022), 450:1–450:28. https://doi.org/10.1145/3555551

    2022 doi

  19. [27]

    Hussam Habib, Ryan Stoldt, Raven Maragh-Lloyd, Brian Ekdale, and Rishab Nithyanand. 2024. Uncovering the Interaction Equation: Quantifying the Effect of User Interactions on Social Media Homepage Recommendations. https: //doi.org/10.48550/arXiv.2407.07227 arXiv:2407.07227 [cs]

    2024 doi arXiv

  20. [28]

    Muhammad Haroon, Anshuman Chhabra, Xin Liu, Prasant Mohapatra, Zubair Shafiq, and Magdalena Wojcieszak. 2022. YouTube, The Great Radicalizer? Au- diting and Mitigating Ideological Biases in YouTube Recommendations. http: //arxiv.org/abs/2203.10666 arXiv:2203.10666 [cs]

    2022 arXiv

  21. [29]

    Rothschild, and Duncan J

    Homa Hosseinmardi, Amir Ghasemian, Aaron Clauset, Markus Mobius, David M. Rothschild, and Duncan J. Watts. 2021. Examining the consumption of radical content on YouTube. Proceedings of the National Academy of Sciences 118, 32 (Aug. 2021), e2101967118. https://doi.org/10.1073/p...

    2021 doi

  22. [30]

    Homa Hosseinmardi, Amir Ghasemian, Miguel Rivera-Lanas, Manoel Horta Ribeiro, Robert West, and Duncan J. Watts. 2024. Causally estimating the effect of YouTube’s recommender system using counterfactual bots. Proceedings of the National Academy of Sciences 121, 8 (Feb. 2024), e...

    2024 doi

  23. [31]

    Hutto and Eric Gilbert

    C. Hutto and Eric Gilbert. 2014. VADER: A Parsimonious Rule-Based Model for Sentiment Analysis of Social Media Text. Proceedings of the International AAAI Conference on Web and Social Media 8, 1 (May 2014), 216–225. https: //doi.org/10.1609/icwsm.v8i1.14550 Number: 1

    2014 doi

  24. [32]

    Natalie Kitroeff. 2014. Why That Video Went Viral. The New York Times (May 2014). https://www.nytimes.com/2014/05/20/science/why-that-video- went-viral.html

    2014

  25. [33]

    Jon Kleinberg, Sendhil Mullainathan, and Manish Raghavan. 2024. The Challenge of Understanding What Users Want: Inconsistent Preferences and Engagement Optimization. Management Science 70, 9 (Sept. 2024), 6336–6355. https://doi. org/10.1287/mnsc.2022.03683 Publisher: INFORMS

    2024

  26. [34]

    Brian Knutson, Tiffany Hsu, Michael Ko, and Jeanne Tsai. 2024. News source bias and sentiment on social media. https://doi.org/10.31234/osf.io/jxu95

    2024 doi

  27. [35]

    Luan. 2025. LuanRT/YouTube.js. https://github.com/LuanRT/YouTube.js original-date: 2021-09-27T21:36:05Z

    2025

  28. [36]

    Vera Liao, James Choi, Kaiyue Fan, Sean A

    Kai Lukoff, Ulrik Lyngs, Himanshu Zade, J. Vera Liao, James Choi, Kaiyue Fan, Sean A. Munson, and Alexis Hiniker. 2021. How the Design of YouTube Influences User Sense of Agency. InProceedings of the 2021 CHI Conference on Human Factors in Computing Systems . ACM, Yokohama Jap...

    2021

  29. [37]

    Cameron Martel, Gordon Pennycook, and David G. Rand. 2020. Reliance on emo- tion promotes belief in fake news. Cognitive Research: Principles and Implications 5, 1 (Oct. 2020), 47. https://doi.org/10.1186/s41235-020-00252-3

    2020 doi

  30. [38]

    Adrienne Massanari. 2017. #Gamergate and The Fappening: How Reddit’s algo- rithm, governance, and culture support toxic technocultures.New Media & Society 19, 3 (March 2017), 329–346. https://doi.org/10.1177/1461444815608807 Publisher: SAGE Publications

    2017 doi

  31. [39]

    Clark McCauley and Sophia Moskalenko. 2008. Mechanisms of Political Radical- ization: Pathways Toward Terrorism. Terrorism and Political Violence 20, 3 (July 2008), 415–433. https://doi.org/10.1080/09546550802073367 Publisher: Routledge _eprint: https://doi.org/10.1080/0954655...

    2008 doi

  32. [40]

    Ryan McGrady, Kevin Zheng, Rebecca Curran, Jason Baumgartner, and Ethan Zuckerman. 2023. Dialing for Videos: A Random Sample of YouTube. Journal of Quantitative Description: Digital Media 3 (Dec. 2023). https://doi.org/10. 51685/jqd.2023.022

    2023

  33. [41]

    McLoughlin, William J

    Killian L. McLoughlin, William J. Brady, Aden Goolsbee, Ben Kaiser, Kate Klonick, and M. J. Crockett. 2024. Misinformation exploits outrage to spread online.Science 386, 6725 (Nov. 2024), 991–996. https://doi.org/10.1126/science.adl2829 Publisher: American Association for the ...

    2024 doi

  34. [42]

    Kevin Munger and Joseph Phillips. 2022. Right-wing YouTube: a supply and demand perspective. The International Journal of Press/Politics 27, 1 (2022), 186–

    2022

  35. [43]

    Fox News. 2024. What is brain rot? The science behind what obsessive scrolling does to our brains. https://www.foxnews.com/health/brain-rot-science- behind-too-much-scrolling-does-brains

    2024

  36. [44]

    Nicolas Pröllochs, Dominik Bär, and Stefan Feuerriegel. 2021. Emotions explain differences in the diffusion of true vs. false social media rumors. Scientific Reports 11, 1 (Nov. 2021), 22721. https://doi.org/10.1038/s41598-021-01813-2 Publisher: Nature Publishing Group

    2021 doi

  37. [45]

    Van Bavel, and Sander van der Linden

    Steve Rathje, Jay J. Van Bavel, and Sander van der Linden. 2021. Out-group ani- mosity drives engagement on social media. Proceedings of the National Academy of Sciences 118, 26 (June 2021), e2024292118. https://doi.org/10.1073/pnas .2024292118 Publisher: Proceedings of the Na...

    2021 doi

  38. [46]

    Manoel Horta Ribeiro, Raphael Ottoni, Robert West, Virgílio A. F. Almeida, and Wagner Meira. 2020. Auditing radicalization pathways on YouTube. In Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency (FAT* ’20). Association for Computing Machinery, ...

    2020

  39. [47]

    Manoel Horta Ribeiro, Veniamin Veselovsky, and Robert West. 2023. The Amplification Paradox in Recommender Systems. Proceedings of the Interna- tional AAAI Conference on Web and Social Media 17 (June 2023), 1138–1142. https://doi.org/10.1609/icwsm.v17i1.22223

    2023 doi

  40. [48]

    Robertson, Nicolas Pröllochs, Kaoru Schwarzenegger, Philip Pärnamets, Jay J

    Claire E. Robertson, Nicolas Pröllochs, Kaoru Schwarzenegger, Philip Pärnamets, Jay J. Van Bavel, and Stefan Feuerriegel. 2023. Negativity drives online news consumption. Nature Human Behaviour 7, 5 (2023), 812–822. https://doi.or g/10.1038/s41562-023-01538-4

    2023 doi

  41. [49]

    Ruggiero

    Thomas E. Ruggiero. 2000. Uses and Gratifications Theory in the 21st Century. Mass Communication and Society 3, 1 (Feb. 2000), 3–37. http s://doi.org/10.1207/S15327825MCS0301_02 Publisher: Routledge _eprint: https://doi.org/10.1207/S15327825MCS0301_02

    2000 doi

  42. [50]

    Laura Savolainen, Damian Trilling, and Dimitra Liotsiou. 2020. Delighting and Detesting Engagement: Emotional Politics of Junk News. Social Media + Society 6, 4 (Oct. 2020), 2056305120972037. https://doi.org/10.1177/ 2056305120972037 Publisher: SAGE Publications Ltd

    2020

  43. [51]

    Michael D. Slater. 2007. Reinforcing spirals: The mutual influence of media selectivity and media effects and their impact on individual behavior and social identity. Communication Theory 17, 3 (2007), 281–303. https://doi.org/10. 1111/j.1468-2885.2007.00296.x Place: United Ki...

    2007

  44. [52]

    Maddie Stone. 2015. How Project InnerTube Helped Pull YouTube Out of the Gutter. https://gizmodo.com/how-project-innertube-helped-pull-yo utube-out-of-the-gu-1704946491

    2015

  45. [53]

    Luke Thorburn, Jonathan Stray, and Priyanjana Bengani. 2022. What Does it Mean to Give Someone What They Want? The Nature of Prefer- ences in Recommender Systems. https://medium.com/understanding- recommenders/what-does-it-mean-to-give-someone-what-they-want- the-nature-of-pre...

    2022

  46. [54]

    Zeynep Tufekci. 2018. Opinion | YouTube, the Great Radicalizer. The New York Times (March 2018). https://www.nytimes.com/2018/03/10/opinion/sund ay/youtube-politics-radical.html

    2018

  47. [55]

    Valentino, Vincent L

    Nicholas A. Valentino, Vincent L. Hutchings, Antoine J. Banks, and Anne K. Davis. 2008. Is a Worried Citizen a Good Citizen? Emotions, Political Informa- tion Seeking, and Learning via the Internet. Political Psychology 29, 2 (2008), 247–273. https://doi.org/10.1111/j.1467-922...

    2008

  48. [56]

    Soroush Vosoughi, Deb Roy, and Sinan Aral. 2018. The spread of true and false news online. Science 359, 6380 (March 2018), 1146–1151. https://doi.org/10. 1126/science.aap9559 Publisher: American Association for the Advancement of Science

    2018

  49. [57]

    Sarita Yardi and Danah Boyd. 2010. Dynamic Debates: An Analysis of Group Polarization Over Time on Twitter. Bulletin of Science, Technology & Society 30, 5 (Oct. 2010), 316–327. https://doi.org/10.1177/0270467610380011 Publisher: SAGE Publications Inc

    2010 doi

  50. [58]

    Xudong Yu, Muhammad Haroon, Ericka Menchen-Trevino, and Magdalena Woj- cieszak. 2024. Nudging recommendation algorithms increases news consumption and diversity on YouTube.PNAS nexus 3, 12 (2024), pgae518. https://academic .oup.com/pnasnexus/article-abstract/3/12/pgae518/79047...

    2024

  51. [59]

    Jia Zhou, Yanhua Li, Vijay Kumar Adhikari, and Zhi-Li Zhang. 2011. Counting YouTube videos via random prefix sampling. In Proceedings of the 2011 ACM SIGCOMM conference on Internet measurement conference (IMC ’11) . Association for Computing Machinery, New York, NY, USA, 371–3...

    2011

  52. [219]

    Publisher: SAGE Publications Sage CA: Los Angeles, CA

  53. [370]

    https://doi.org/10.1037/1089-2680.5.4.323 Place: US Publisher: Educational Publishing Foundation

    doi

  54. [2001]

    Review of General Psychology 5, 4 (2001), 323–

    Bad is stronger than good. Review of General Psychology 5, 4 (2001), 323–

    2001

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.