Pith. sign in

REVIEW 3 major objections 4 minor 94 references

Compass: Continuously Aligning Social Media Feeds via In-Situ Reflections

T0 review · 3 major / 4 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read By moving preference reflection into the feed and automating realignment, Compass gets users to refine their preferences about 82 times more often and more than doubles the share of feed content they actually wanted.

desk verdict Worth a serious referee: a solid integrated system with a real effect, but the feed-alignment advantage is not cleanly attributed between simulated signals and direct injection. read the letter →

arxiv 2608.04274 v1 pith:HLTBODIH submitted 2026-08-04 cs.HC cs.AI

Aadit Barua , Leijie Wang , Amy X. Zhang This is my paper · ORCID
classification cs.HCcs.AI
keywords reflectivepreferencesfeedalignmentin-situreflectionrecommendersystemsYouTubeShortsbrowserextensionsimulatedengagementfieldstudy
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

This paper tries to establish that the gap between what people engage with in the moment and what they want after reflection can be narrowed by moving preference-setting out of configuration pages and into the feed itself. It presents Compass, a browser extension for YouTube Shorts that prompts users to confirm, reject, or refine their stated preferences while scrolling, and then keeps the feed aligned automatically by simulating watch and like events and, when needed, inserting or removing videos directly. In a 10-day between-subjects field study with 15 participants, Compass users refined their preferences about 82 times more often than users of a manual configuration baseline, saw about 49% of their feed match their 'see more' preferences versus 20%, and spent about 59% of their watch time on that content versus 21%, with comparable usability and diversity. The authors argue that sustained, low-effort reflection in context, rather than one-time setup, is what makes feed alignment actually work.

What carries the argument

Three components carry the system. First, an in-situ reflection layer intercepts YouTube Shorts' preloaded video queue, runs each upcoming video through a real-time pipeline that combines metadata and three thumbnails into an LLM-generated semantic description, and embeds that description. Second, each user preference is represented as an editable bullet-point rubric plus optional example videos, also embedded; videos whose similarity falls near a threshold are flagged as uncertain and offered to the user as annotation opportunities, following active-learning uncertainty sampling. Third, an alignment engine monitors each preference's match rate over a sliding 48-hour window; when a 'see more' preference falls below a 6% rate, it simulates watching and liking matching videos in a background tab, and while waiting for the recommender to respond it also injects unwatched matching videos into the feed via URL updates and removes 'avoid' videos via DOM deletion. This combination is what the paper credits for the large alignment and refinement differences observed.

What would settle it

A direct mechanism test would isolate simulation from DOM manipulation. Run the same Compass interface in two conditions, one with background-tab simulation enabled and one with it disabled while keeping video insertion and removal active; if 'see more' alignment stays near the reported 49.3% in both conditions, the recommender-steering component is not load-bearing, whereas a drop toward the 20.3% baseline would confirm that simulated engagement is what moves the feed.

Watch

Extended reading notes

Core claim

Compass's central claim is that continuous alignment between a recommendation feed and a user's reflective preferences is achievable without replacing the underlying recommender or sacrificing casual browsing. The system surfaces lightweight reflection prompts directly in the feed: it asks users to label videos whose match to a preference is uncertain, suggests new preferences when a user's behavior signals misalignment (for example, disliking a video after watching several like it), and periodically summarizes how preferences have shaped the feed. Each prompt produces a concrete preference revision that the system then acts on through two channels: simulated engagement behaviors meant to steer the recommender, and direct DOM manipulation that inserts matching videos or removes unwanted ones. In the field study, participants using Compass made 11.50 refinements on average compared with 0.14 in the baseline, saw 49.3% of watched videos match their 'see more' preferences compared with 20.3%, and directed 58.7% of watch time to that content compared with 21.0%, while SUS scores (84.1 vs 83.2) and feed diversity were comparable. The paper reads these results as evidence that in-situ reflection plus automated realignment closes the gap between stated and revealed preferences more effectively than requiring users to take initiative.

Load-bearing premise

The weakest load-bearing premise is that YouTube's recommendation algorithm treats simulated watch and like events performed in a background tab the same as real user engagement, and changes future recommendations accordingly. If the recommender ignores or heavily discounts those simulated signals, the measured feed-alignment advantage would have to be attributed to the client-side insertion and removal of videos, and the system's 'steering' mechanism would not be doing the work the paper assigns to it.

Editorial extensions

If this is right

  • Users who otherwise would not revisit configuration pages will refine preferences repeatedly if asked in the feed at the right moment: 73 of 92 refinements in the Compass condition happened directly in the feed.
  • The same system can raise the share of a feed that matches a 'see more' preference to roughly half, and direct more than half of watch time to it, without measurably reducing content diversity.
  • A baseline that offers the same preference language but requires manual triggering produces far less alignment, indicating that the interaction design, not the preference vocabulary, is the key bottleneck.
  • Adding this reflection layer does not hurt perceived usability: SUS scores were statistically indistinguishable from the baseline and both fell in the 'excellent' range.
  • The real-time matching pipeline reaches or exceeds the LLM-as-judge F1 once two example videos are added, which is what makes in-feed interventions fast enough for short-form video.

Reading between the lines

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

  • An editor's inference: the paper's alignment numbers may overstate the recommender-steering contribution, because the Compass condition also performed 107 direct DOM injections that the baseline never received; a controlled ablation would be needed before attributing the alignment gap to simulation.
  • The paper leaves implicit that the 6% match-rate threshold and 48-hour window were chosen through pilot testing; on other platforms with different recommender responsiveness, those parameters would likely need retuning, not just reuse.
  • The in-situ reflection pattern could generalize beyond video: any preloading feed with identifiable content units (articles, posts, audio) could prompt users at moments of uncertain relevance, but the latency constraint is stronger for short-form video, so the general mechanism may be easier to deploy in slower media.
  • If platforms harden against simulated engagement or block DOM-level rewriting, Compass's steering channel loses its reach; a mobile implementation would have to rely on accessibility APIs, which the paper flags as open work.
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

3 major / 4 minor

Summary. The paper presents Compass, a browser extension for YouTube Shorts that elicits users' reflective preferences through lightweight in-situ notifications, represents those preferences as evolving rubrics with example videos, and aligns the feed by simulating engagement signals in a background tab while also directly injecting and removing videos via DOM manipulation. The authors report a 10-day between-subjects field study (N=15) in which Compass participants made significantly more preference refinements (11.50 vs 0.14 per participant, p=.001), saw more content matching their See More preferences (49.3% vs 20.3%, p=.009), and devoted more watch time to that content (58.7% vs 21.0%, p=.001), with comparable SUS scores and feed diversity versus a baseline without continuous support. They also report a technical evaluation on 408 human-labeled preference-video pairs showing that embedding-based matching with LLM-synthesized video representations achieves strong ROC-AUC, with gains from adding example videos.

Significance. If the causal claims hold, Compass makes a valuable contribution to HCI and recommender-systems research: it demonstrates a practical way to embed continuous reflection into an existing short-form video platform, and it articulates two complementary feed-alignment mechanisms (behavioral-signal simulation and client-side feed manipulation) that could generalize beyond YouTube. The paper is unusually complete in its implementation detail: the full prompt suite is given in Appendix C, the field-study instrumentation and timelines are in the appendix, and the technical evaluation uses human labels rather than only system-internal scores. The statistical analysis is careful for the sample size, with participant-level bootstraps and effect sizes. However, the central empirical claim about the mechanism of feed alignment is not fully supported because the outcome metric shares its representational family with the content-selection mechanism, and because the simulation mechanism is never isolated from the direct DOM injections. The paper's own limitation section is honest about sample size and platform scope, but these two unresolved points affect the headline interpretation.

major comments (3)
  1. [Section 4.2.2 and Appendix A.3, Table 2] The feed-alignment and watch-time measures are not independent of the mechanism that produces the alignment. The measurement classifier is 'Name→LLM + Meta+Frames→LLM' at threshold 0.370, which is the same embedding family and the same LLM-description pipeline that Compass uses to retrieve and inject videos for See More preferences. For an injected video, a match to the preference is true by construction, because the video was selected by high similarity to the same preference embedding. The 49.3% vs 20.3% alignment gap and the 58.7% vs 21.0% watch-time gap therefore partly measure the system's own selection rule rather than an independent property of the feed. The human-labeled evaluation in Section 5 validates the classifier on general pairs, but it does not validate the specific injected videos. Please report these metrics separately for organic versus injected videos, or add a human-rating check on a sample of watched videos.
  2. [Section 4.2.2, Section 6.2, and Section 3.2.3] The paper's central mechanism claim—that Compass steers YouTube's recommender through simulated behavioral signals—is not isolated from direct client-side manipulation. The text in Section 4.2.2 says the alignment difference is explained by more background simulations 'as well as 107 total direct video injections exclusive to Compass by design,' but the abstract and introduction assert that the advantage was 'largely driven by' proactive simulation. There is no condition, probe, or post-hoc analysis that separates changes in the organic feed caused by simulated watch/like events from changes caused by 107 DOM injections. Given that Section 3.2.3 itself notes that negative feedback on YouTube has limited documented impact, the possibility that YouTube discounts simulated signals is not hypothetical. If the recommender ignores those signals, the contribution reduces to client-side filtering and insertion, and the claimed generality of 'steering an existing engagement-based algorithm with behavioral signals' in Section 6.2 is unsupported. Please add a decomposition analysis (e.g., alignment of organic videos before vs after simulation events within Compass) or a separate technical probe measuring whether simulated events shift subsequent recommendations.
  3. [Section 4.1.2 and Table 5] The between-subjects comparison relies on alternating assignment with only 15 participants, and the participant table reveals visible baseline differences in self-reported daily usage intensity (e.g., B1 reports 4+ hours/day; B2, B4, and B5 report 2–4 hours/day, while several Compass participants report under 1 hour). If baseline participants are heavier and more habitual users, their lower refinement and alignment rates may partly reflect different usage patterns rather than the absence of in-situ support. The paper discloses the small sample but does not report a baseline-equivalence check on usage frequency or duration. Please add a comparison of these pre-treatment variables and, if feasible, a sensitivity analysis that adjusts for or matches on them.
minor comments (4)
  1. [Appendix A.3] The phrase 'Name→LLM and Meta+Frames→LLM with zero examples provided' is ambiguous because it concatenates the preference representation and the video representation; please write 'the preference representation Name→LLM with the video representation Meta+Frames→LLM and zero examples.'
  2. [Table 2 and Section 4.2.2] The Avoid-preference alignment and watch-time comparisons are non-significant and, if anything, directionally favor the baseline (19.0% vs 6.5% alignment; 16.7% vs 6.9% watch-time share). The paper's headline wording 'stronger feed alignment' should be qualified as applying to See More preferences, since the Avoid results do not support a general alignment claim.
  3. [Section 5.1 and Table 1] The thresholds reported as 'best F1' are selected on the same 408-sample evaluation set used to report ROC-AUC, so the numerical operating points may be optimistic. This does not affect the field-study comparison between conditions, but a sentence clarifying that no held-out threshold selection was performed would improve transparency.
  4. [Section 4.2.1] The statement that added reflection did not come at the cost of 'the casual nature' relies mainly on comparable SUS scores and qualitative quotes; SUS is a usability measure, not a direct measure of casual browsing experience. Consider reporting a more direct indicator, such as session length or frequency of uninterrupted scrolling, if available.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the feed-alignment outcome is independently validated and not reducible to the system's own parameters.

full rationale

I find no circular step in the derivation chain. The central field-study claims are empirical between-subjects comparisons, not quantities derived from fitted parameters. The feed-alignment metric is deliberately constructed to be shared across conditions: Appendix A.3 states that the authors did not use Compass's deployed classifier and instead 'adopt the method from our technical evaluation that best represents both approaches: Name→LLM and Meta+Frames→LLM with zero examples provided,' with the threshold (0.370) chosen during the technical evaluation. That measurement configuration is externally validated on 408 human-labeled preference-video pairs (ROC-AUC 0.842 in Table 1), giving the outcome measure independent grounding despite sharing an embedding family with the system's selection mechanism. The difference in preference-refinement counts (11.50 vs. 0.14) is a direct consequence of the manipulated feature set rather than a fitted prediction, and the paper does not present it as a model output. Self-citations, such as [26] and [75], support related-work observations about example-based preference articulation but are not load-bearing for the core results, and no uniqueness theorem is imported from the authors' prior work. The unresolved attribution between simulated behavioral signals and 107 direct DOM injections is a real internal-validity and mechanism-isolation concern, but it is a confound rather than a circular reduction: the reported alignment advantage would still be an empirical outcome even if caused entirely by injection. Because no equation reduces to an input by construction and no fitted parameter is renamed as a prediction, the paper is not circular in the sense defined here.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The paper's central results depend on a small number of tuned thresholds, on the unverified responsiveness of YouTube's recommender to simulated signals, and on trusting the LLM-embedding matching as both intervention and measurement. Free parameters are modest but include the 6% trigger, the deployed and measurement similarity thresholds, and the near-boundary uncertainty band. No invented entities. External grounding comes from the human-labeled evaluation dataset, which grounds the matching model, but not the simulation mechanism.

free parameters (4)
  • underrepresentation threshold for triggering feed alignment = 6% of videos in a sliding 48-hour window
    Triggers simulation and injection; described as 'determined through pilot experiments' (Section 3.2.3).
  • deployed matching similarity threshold range = 0.4 to 0.46
    Chosen during the field study because the authors 'experimented and found a preference for precision' (Appendix A.3).
  • measurement classifier threshold = 0.370
    Selected to maximize F1 on the human-labeled evaluation dataset, then applied to label feed videos in the field study (Appendix A.3).
  • uncertain-match threshold near the decision boundary
    Uncertain matches are flagged near the threshold, with the threshold 'set more strictly for preferences with fewer user-provided examples' (Section 3.2.2). The exact rule is not specified.
assumptions (4)
  • domain assumption YouTube Shorts preloads upcoming videos and exposes interpretable API responses that a browser extension can intercept to extract video IDs before display.
    Used to give the backend processing time; stated in Section 3.2.1 without independent verification.
  • domain assumption Simulated watch and like signals executed in a background tab steer YouTube's recommender toward similar content.
    Core alignment mechanism; the paper cites prior work on user signaling [45] but does not validate that background-tab simulation affects the production recommender (Sections 3.2.3, 4.2.2).
  • domain assumption LLM-generated semantic descriptions plus text-embedding similarity adequately capture whether a short video matches a user's preference for feed-level measurement.
    Partially validated by the technical evaluation on 408 human-labeled pairs (Table 1), but the field measurement still relies on the model rather than human labels (Appendix A.3).
  • domain assumption Participants did not use YouTube outside study sessions except on a different account or incognito mode.
    Stated instruction in Section 4.1.2; self-reported compliance is not verified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Compass: Continuously Aligning Social Media Feeds via In-Situ Reflections." pith.science (2026). https://pith.science/paper/HLTBODIH

@misc{pith2026260804274,
  author       = {Pith},
  title        = {Pith review of: Compass: Continuously Aligning Social Media Feeds via In-Situ Reflections},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HLTBODIH}},
  note         = {Machine review of arXiv:2608.04274}
}
read the original abstract

Social media recommendation feeds often optimize for users' immediate impulses rather than preferences they would hold after deeper reflection. Some systems address this misalignment by incorporating users' explicit preferences via a configuration page or in-feed controls instead of just behavioral signals. However, users typically have evolving preferences, and their stated preferences and behavior naturally diverge, necessitating continuous reflection and feed realignment. But existing strategies require the user to take initiative and are often effortful; as a result, in practice they are rarely invoked. We present Compass, a system that aligns a user's feed with their reflective preferences by helping users reflect on and articulate their preferences given their behavior. To enable continuous reflection during everyday browsing, Compass surfaces in-situ reflections via lightweight notifications, while feed alignment is achieved by periodically simulating behavioral signals and directly manipulating feed content. We embedded Compass within YouTube Shorts and compared it against a baseline without continuous support through a 10-day field study (N=15). We found that Compass promoted more reflective and purposeful feed consumption, iterative preference adjustment, and stronger feed alignment, without sacrificing the casual nature of feed browsing.

Figures

Figures reproduced from arXiv: 2608.04274 by the authors.

Figure 1
Figure 1. Compass supports continuous alignment between social media recommendation feeds and users’ reflective preferences. Rather than treating reflection as a one-time configuration, Compass helps users articulate their reflective preferences through in-situ reflections during everyday browsing. Compass then automatically aligns their feed through simulation of behavioral signals and direct DOM-based insertion and removal … view at source ↗
Figure 2
Figure 2. Compass embeds lightweight interactions directly into the recommendation feed to support users to author new reflective preferences and refine them more accurately over time, accumulating into a central preferences page for ongoing management. (A-C) In-situ reflections prompt users to author new reflective preferences. (A) When users express explicit signals during browsing (e.g., disliking a video), Compass suggest… view at source ↗
Figure 3
Figure 3. Whenever it detects mismatches between the feed [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Compass semantically represents upcoming videos and users’ reflective preferences, then compares these representa￾tions to support feed reflection and alignment. For each upcoming video, Compass retrieves its metadata and keyframes to generate a semantic description an…
Figure 5
Figure 5. Figure 5: Session activity timelines for two Compass participants (C2, C6) and two Baseline participants (B1, B4) across up to five study sessions. Each gray bar represents the time a participant spent watching videos during a session. Colored marks on or above the bar indicate …
Figure 6
Figure 6. Figure 6: Compass participants rated Purposeful Scrolling, Feed Control, and Time Well-Spent higher than baseline participants, with no difference in Consumption Awareness. Feed Control reached statistical significance (∗𝑝 < .05). Importantly, we find that this added reflection …
Figure 7
Figure 7. Figure 7: Mean events per participant across both conditions. Com￾pass participants engaged in significantly more preference refine￾ment and like/dislike interactions, whereas preference creation, preference deletion, and manual alignment were comparable across conditions. (∗∗𝑝 …
Figure 8
Figure 8. Figure 8: breaks these scores down by item. Both conditions rated the systems highly on ease of use (Q3: 4.75 vs. 4.71), learnability (Q7: 4.62 vs. 4.43), and low need for technical support (Q4: 1.12 vs. 1.43), and similarly low on cumbersomeness (Q8: 1.50 vs. 1.57) and prior le…
Figure 9
Figure 9. Figure 9: Session activity timelines for all Compass (C1–C8) and Baseline (B1–B7) participants across up to five study sessions, showing the frequency and distribution of key interactions. Events: preference creation, refinement, and deletion; manual alignment; like or dislike v…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

94 extracted references · 61 canonical work pages

  1. [1]

    Amanda Y Agan, Diag Davenport, Jens Ludwig, and Sendhil Mullainathan. 2023. Automating automaticity: How the context of human choice affects the extent of algorithmic bias. Technical Report. National Bureau of Economic Research

    2023

  2. [2]

    Aaron Bangor, Philip T Kortum, and James T Miller. 2008. An empirical evaluation of the system usability scale.Intl. Journal of Human–Computer Interaction24, 6 (2008), 574–594

    2008

  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. 1–13

    2022

  4. [4]

    Rahul Bhargava, Anna Chung, Neil S Gaikwad, Alexis Hope, Dennis Jen, Jasmin Rubinovitz, Belén Saldías-Fuentes, and Ethan Zuckerman. 2019. Gobo: A system for exploring user control of invisible algorithms in social media. InCompanion publication of the 2019 conference on computer supported cooperative work and social computing. 151–155

    2019

  5. [5]

    Nattapat Boonprakong, Benjamin Tag, Jorge Goncalves, and Tilman Dingler

  6. [6]

    William J Brady, Joshua Conrad Jackson, Björn Lindström, and MJ Crockett. 2023. Algorithm-mediated social learning in online social networks.Trends in cognitive sciences27, 10 (2023), 947–960

    2023

  7. [7]

    William J Brady, Julian A Wills, John T Jost, Joshua A Tucker, and Jay J Van Bavel

  8. [8]

    Virginia Braun and Victoria Clarke. 2019. Reflecting on reflexive thematic analysis. Qualitative research in sport, exercise and health11, 4 (2019), 589–597

    2019

Show all 94 references
  1. [9]

    John Brooke et al. 1996. SUS-A quick and dirty usability scale.Usability evaluation in industry189, 194 (1996), 4–7

    1996

  2. [10]

    Micah Carroll, Adeline Foote, Kevin Feng, Marcus Williams, Anca Dragan, W Bradley Knox, and Smitha Milli. 2025. CTRL-Rec: Controlling recommender systems with natural language.arXiv preprint arXiv:2510.12742(2025)

    2025

  3. [11]

    Micah D Carroll, Anca Dragan, Stuart Russell, and Dylan Hadfield-Menell. 2022. Estimating and penalizing induced preference shifts in recommender systems. In International Conference on Machine Learning. PMLR, 2686–2708

    2022

  4. [12]

    Eun Kyoung Choe, Bongshin Lee, Matthew Kay, Wanda Pratt, and Julie A Kientz

  5. [13]

    Frederick Choi and Eshwar Chandrasekharan. 2025. Designing Usable Controls for Customizable Social Media Feeds.arXiv preprint arXiv:2509.19615(2025)

    2025

  6. [14]

    Ananta Chowdhury and Andrea Bunt. 2023. Co-Designing with Early Adoles- cents: Understanding Perceptions of and Design Considerations for Tech-Based Mediation Strategies that Promote Technology Disengagement. InProceedings of the 2023 CHI Conference on Human Factors in Computi...

    2023

  7. [15]

    Emily IM Collins, Anna L Cox, Jon Bird, and Daniel Harrison. 2014. Social networking use and RescueTime: the issue of engagement. InProceedings of the 2014 ACM International Joint Conference on Pervasive and Ubiquitous Computing: Adjunct Publication. 687–690

    2014

  8. [16]

    Sunny Consolvo, Predrag Klasnja, David W McDonald, Daniel Avrahami, Jon Froehlich, Louis LeGrand, Ryan Libby, Keith Mosher, and James A Landay. 2008. Flowers or a robot army? Encouraging awareness & activity with personal, mobile displays. InProceedings of the 10th internation...

    2008

  9. [17]

    Paul Covington, Jay Adams, and Emre Sargin. 2016. Deep neural networks for youtube recommendations. InProceedings of the 10th ACM conference on recommender systems. 191–198

    2016

  10. [18]

    Tom Cunningham, Sana Pandey, Leif Sigerson, Jonathan Stray, Jeff Allen, Bonnie Barrilleaux, Ravi Iyer, Mohit Kothari, Behnam Rezaei, Sanjay Kairam, et al. 2025. Ranking by engagement and non-engagement signals: Learnings from industry. Annals of the New York Academy of Science...

    2025

  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.arXiv preprint arXiv:2402.06831(2024)

    2024 arXiv

  12. [20]

    Algorithms ruin everything

    Michael Ann DeVito, Darren Gergle, and Jeremy Birnholtz. 2017. " Algorithms ruin everything" # RIPTwitter, Folk Theories, and Resistance to Algorithmic Change in Social Media. InProceedings of the 2017 CHI conference on human factors in computing systems. 3163–3174

    2017

  13. [21]

    Michael D Ekstrand and Martijn C Willemsen. 2016. Behaviorism is not enough: better recommendations through listening to users. InProceedings of the 10th ACM conference on recommender systems. 221–224

    2016

  14. [22]

    Úlfar Erlingsson, Vasyl Pihur, and Aleksandra Korolova. 2014. Rappor: Random- ized aggregatable privacy-preserving ordinal response. InProceedings of the 2014 ACM SIGSAC conference on computer and communications security. 1054–1067

    2014

  15. [23]

    Motahhare Eslami, Karrie Karahalios, Christian Sandvig, Kristen Vaccaro, Aimee Rickman, Kevin Hamilton, and Alex Kirlik. 2016. First I" like" it, then I hide it: Folk Theories of Social Feeds. InProceedings of the 2016 cHI conference on human factors in computing systems. 2371–2382

    2016

  16. [24]

    I always assumed that I wasn’t really that close to [her]

    Motahhare Eslami, Aimee Rickman, Kristen Vaccaro, Amirhossein Aleyasen, Andy Vuong, Karrie Karahalios, Kevin Hamilton, and Christian Sandvig. 2015. " I always assumed that I wasn’t really that close to [her]" Reasoning about invisible algorithms in news feeds. InProceedings of...

    2015

  17. [25]

    Motahhare Eslami, Kristen Vaccaro, Min Kyung Lee, Amit Elazari Bar On, Eric Gilbert, and Karrie Karahalios. 2019. User attitudes towards algorithmic opacity and transparency in online reviewing platforms. InProceedings of the 2019 CHI Conference on Human Factors in Computing S...

    2019

  18. [26]

    KJ Kevin Feng, Xander Koo, Lawrence Tan, Amy Bruckman, David W McDonald, and Amy X Zhang. 2024. Mapping the Design Space of Teachable Social Media Feed Experiences. InProceedings of the 2024 CHI Conference on Human Factors in Computing Systems. 1–20

    2024

  19. [27]

    Casey Fiesler and Brianna Dym. 2020. Moving across lands: Online platform migration in fandom communities.Proceedings of the ACM on Human-Computer Interaction4, CSCW1 (2020), 1–25

    2020

  20. [28]

    Dan Friedman and Adji Bousso Dieng. 2022. The vendi score: A diversity evalua- tion metric for machine learning.arXiv preprint arXiv:2210.02410(2022)

    2022 arXiv

  21. [29]

    Cristos Goodrow. 2021. On YouTube’s recommendation system. YouTube Blog. https://blog.youtube/inside-youtube/on-youtubes-recommendation-system/ Ac- cessed: 2026-03-26

    2021

  22. [30]

    Andrew M Guess, Neil Malhotra, Jennifer Pan, Pablo Barberá, Hunt Allcott, Taylor Brown, Adriana Crespo-Tenorio, Drew Dimmery, Deen Freelon, Matthew Gentzkow, et al. 2023. How do social media feed algorithms affect attitudes and behavior in an election campaign?Science381, 6656...

    2023

  23. [31]

    Aastha Gupta. 2021. Incorporating More Feedback Into News Feed Ranking. Meta Newsroom. https://about.fb.com/news/2021/04/incorporating-more-feedback- into-news-feed-ranking/ Accessed: 2026-03-26

    2021

  24. [32]

    Muhammad Haroon, Magdalena Wojcieszak, Anshuman Chhabra, Xin Liu, Pras- ant Mohapatra, and Zubair Shafiq. 2023. Auditing YouTube’s recommendation system for ideologically congenial, extreme, and problematic recommendations. Proceedings of the national academy of sciences120, 5...

    2023

  25. [33]

    Silas Hsu, Kristen Vaccaro, Yin Yue, Aimee Rickman, and Karrie Karahalios. 2020. Awareness, navigation, and use of feed control settings online. InProceedings of the 2020 CHI Conference on Human Factors in Computing Systems. 1–13

    2020

  26. [34]

    Jack Jamieson and Naomi Yamashita. 2023. Escaping the Walled Garden? User Perspectives of Control in Data Portability for Social Media.Proceedings of the ACM on Human-Computer Interaction7, CSCW2 (2023), 1–27

    2023

  27. [35]

    Shagun Jhaver, Alice Qian Zhang, Quan Ze Chen, Nikhila Natarajan, Ruotong Wang, and Amy X Zhang. 2023. Personalizing content moderation on social media: User perspectives on moderation choices, interface design, and labor. Proceedings of the ACM on Human-Computer Interaction7,...

    2023

  28. [36]

    Chenyan Jia, Michelle S Lam, Minh Chau Mai, Jeffrey T Hancock, and Michael S Bernstein. 2024. Embedding democratic values into social media AIs via societal objective functions.Proceedings of the ACM on Human-Computer Interaction8, CSCW1 (2024), 1–36

    2024

  29. [37]

    2011.Thinking, fast and slow

    Daniel Kahneman. 2011.Thinking, fast and slow. macmillan

    2011

  30. [38]

    Jaejeung Kim, Hayoung Jung, Minsam Ko, and Uichin Lee. 2019. Goalkeeper: Exploring interaction lockout mechanisms for regulating smartphone use.Pro- ceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies 3, 1 (2019), 1–29

    2019

  31. [39]

    Young-Ho Kim, Jae Ho Jeon, Eun Kyoung Choe, Bongshin Lee, KwonHyun Kim, and Jinwook Seo. 2016. TimeAware: Leveraging framing effects to enhance personal productivity. InProceedings of the 2016 CHI Conference on Human Factors in Computing Systems. 272–283

    2016

  32. [40]

    Jon Kleinberg, Sendhil Mullainathan, and Manish Raghavan. 2024. The challenge of understanding what users want: Inconsistent preferences and engagement optimization.Management science70, 9 (2024), 6336–6355

    2024

  33. [41]

    Akaash Kolluri, Renn Su, Farnaz Jahanbakhsh, Dora Zhao, Tiziano Piccardi, and Michael S Bernstein. 2025. Alexandria: A Library of Pluralistic Values for Realtime Re-Ranking of Social Media Feeds.arXiv preprint arXiv:2505.10839(2025)

    2025 arXiv

  34. [42]

    Deepak Kumar, Yousef AbuHashem, and Zakir Durumeric. 2023. Watch your language: large language models and content moderation.arXiv preprint arXiv:2309.14517(2023)

    2023 arXiv

  35. [43]

    David D Lewis. 1995. A sequential algorithm for training text classifiers: Corri- gendum and additional data. InAcm Sigir Forum, Vol. 29. ACM New York, NY, USA, 13–19

    1995

  36. [44]

    Toby Jia-Jun Li, Amos Azaria, and Brad A Myers. 2017. SUGILITE: creating multimodal smartphone automation by demonstration. InProceedings of the 2017 CHI conference on human factors in computing systems. 6038–6049

    2017

  37. [45]

    Wenqi Li, Jui-Ching Kuo, Manyu Sheng, Pengyi Zhang, and Qunfang Wu. 2025. Beyond Explicit and Implicit: How Users Provide Feedback to Shape Personalized Recommendation Content. InProceedings of the 2025 CHI Conference on Human Factors in Computing Systems. 1–17. Compass: Conti...

    2025

  38. [46]

    Yabing Liu, Krishna P Gummadi, Balachander Krishnamurthy, and Alan Mis- love. 2011. Analyzing facebook privacy settings: user expectations vs. reality. InProceedings of the 2011 ACM SIGCOMM conference on Internet measurement conference. 61–70

    2011

  39. [47]

    Yuhan Liu, Emmy Song, Owen Xingjian Zhang, Jewel Merriman, Lei Zhang, and Andrés Monroy-Hernández. 2025. Understanding decentralized social feed curation on mastodon.Proceedings of the ACM on Human-Computer Interaction 9, 7 (2025), 1–25

    2025

  40. [48]

    Tao Lu, Hongxiao Zheng, Tianying Zhang, Xuhai “Orson” Xu, and Anhong Guo

  41. [49]

    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. 1–17

    2021

  42. [50]

    Kai Lukoff, Cissy Yu, Julie Kientz, and Alexis Hiniker. 2018. What makes smart- phone use meaningful or meaningless?Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies2, 1 (2018), 1–26

    2018

  43. [51]

    Omar El Malki, Marianne Aubin Le Quéré, Andrés Monroy-Hernández, and Manoel Horta Ribeiro. 2025. Bonsai: Intentional and personalized social media feeds.arXiv preprint arXiv:2509.10776(2025)

    2025

  44. [52]

    Arunesh Mathur, Jessica Vitak, Arvind Narayanan, and Marshini Chetty. 2018. Characterizing the use of{Browser-Based} blocking extensions to prevent online tracking. InFourteenth symposium on usable privacy and security (SOUPS 2018). 103–116

    2018

  45. [53]

    Meta. 2022. New Ways to Customize Your Facebook Feed. Meta Newsroom. https: //about.fb.com/news/2022/10/new-ways-to-customize-your-facebook-feed/ Ac- cessed: 2026-03-26

    2022

  46. [54]

    Thomas Mildner and Gian-Luca Savino. 2021. Ethical user interfaces: Exploring the effects of dark patterns on facebook. InExtended Abstracts of the 2021 CHI Conference on Human Factors in Computing Systems. 1–7

    2021

  47. [55]

    Carey K Morewedge, Sendhil Mullainathan, Haaya F Naushan, Cass R Sunstein, Jon Kleinberg, Manish Raghavan, and Jens O Ludwig. 2023. Human bias in algorithm design.Nature Human Behaviour7, 11 (2023), 1822–1824

    2023

  48. [56]

    Rosemery O Nelson and Steven C Hayes. 1981. Theoretical explanations for reactivity in self-monitoring.Behavior Modification5, 1 (1981), 3–14

    1981

  49. [57]

    2025.Americans’ Social Media Use 2025

    Pew Research Center. 2025.Americans’ Social Media Use 2025. Technical Re- port. Pew Research Center. https://www.pewresearch.org/internet/2025/11/20/ americans-social-media-use-2025/

    2025

  50. [58]

    Tiziano Piccardi, Martin Saveski, Chenyan Jia, Jeffrey Hancock, Jeanne L Tsai, and Michael S Bernstein. 2024. Reranking social media feeds: A practical guide for field experiments.arXiv preprint arXiv:2406.19571(2024)

    2024

  51. [59]

    Lindsay Popowski, Xiyuan Wu, Charlotte Zhu, Tiziano Piccardi, and Michael S Bernstein. 2026. Social Media Feed Elicitation.arXiv preprint arXiv:2602.18594 (2026)

    2026

  52. [60]

    Dorian Quelle and Alexandre Bovet. 2025. Bluesky: Network topology, polariza- tion, and algorithmic curation.PloS one20, 2 (2025), e0318034

    2025

  53. [61]

    Emilee Rader, Kelley Cotter, and Janghee Cho. 2018. Explanations as mechanisms for supporting algorithmic transparency. InProceedings of the 2018 CHI conference on human factors in computing systems. 1–13

    2018

  54. [62]

    Emilee Rader and Rebecca Gray. 2015. Understanding user beliefs about algorith- mic curation in the Facebook news feed. InProceedings of the 33rd annual ACM conference on human factors in computing systems. 173–182

    2015

  55. [63]

    2022.Does This Button Work? Investigat- ing YouTube’s Ineffective User Controls

    Becca Ricks and Jesse McCrosky. 2022.Does This Button Work? Investigat- ing YouTube’s Ineffective User Controls. Technical Report. Mozilla Founda- tion. https://assets.mofoprod.net/network/documents/Mozilla-Report-YouTube- User-Controls.pdf

    2022

  56. [64]

    Nicolas Ruiz, Gabriela Molina León, and Hendrik Heuer. 2024. Design frictions on social media: Balancing reduced mindless scrolling and user satisfaction. In Proceedings of Mensch und Computer 2024. 442–447

    2024

  57. [65]

    Paul A Samuelson. 2024. A note on the pure theory of consumer’s behaviour. In The Foundations of Price Theory Vol 4. Routledge, 101–116

    2024

  58. [66]

    Christian Sandvig, Kevin Hamilton, Karrie Karahalios, and Cedric Langbort. 2014. Auditing algorithms: Research methods for detecting discrimination on internet platforms.Data and discrimination: converting critical concerns into productive inquiry22, 2014 (2014), 4349–4357

    2014

  59. [67]

    H Sebastian Seung, Manfred Opper, and Haim Sompolinsky. 1992. Query by committee. InProceedings of the fifth annual workshop on Computational learning theory. 287–294

    1992

  60. [68]

    Hong Shen, Alicia DeVos, Motahhare Eslami, and Kenneth Holstein. 2021. Every- day algorithm auditing: Understanding the power of everyday users in surfacing harmful algorithmic behaviors.Proceedings of the ACM on Human-Computer Interaction5, CSCW2 (2021), 1–29

    2021

  61. [69]

    Jonathan Stray. 2021. Beyond Engagement: Aligning Algorith- mic Recommendations with Prosocial Goals. Partnership on AI. https://partnershiponai.org/beyond-engagement-aligning-algorithmic- recommendations-with-prosocial-goals/ Accessed: 2026

    2021

  62. [70]

    Felicia Fang-Yi Tan, Ashwin Ram, Moritz Alexander Messerschmidt, Hasini Amanda Dissanayake, and Suranga Nanayakkara. 2025. Curious shorts: Curiosity-driven exploration and learning on short-form video platforms. InProceedings of the 2025 CHI Conference on Human Factors in Comp...

    2025

  63. [71]

    Jiakai Tang, Yujie Luo, Xunke Xi, Fei Sun, Xueyang Feng, Sunhao Dai, Chao Yi, Dian Chen, Zhujin Gao, Yang Li, et al. 2025. Interactive Recommendation Agent with Active User Commands.arXiv preprint arXiv:2509.21317(2025)

    2025

  64. [72]

    Kristen Vaccaro, Dylan Huang, Motahhare Eslami, Christian Sandvig, Kevin Hamilton, and Karrie Karahalios. 2018. The illusion of control: Placebo effects of control settings. InProceedings of the 2018 CHI Conference on Human Factors in Computing Systems. 1–13

    2018

  65. [73]

    They’ve Over-Emphasized That One Search

    Julie A Vera and Sourojit Ghosh. 2025. " They’ve Over-Emphasized That One Search": Controlling Unwanted Content on TikTok’s For You Page. InProceedings of the 2025 CHI Conference on Human Factors in Computing Systems. 1–8

    2025

  66. [74]

    Karan Vombatkere, Sepehr Mousavi, Savvas Zannettou, Franziska Roesner, and Krishna P Gummadi. 2024. Tiktok and the art of personalization: investigating exploration and exploitation on social media feeds. InProceedings of the ACM Web Conference 2024. 3789–3797

    2024

  67. [75]

    Leijie Wang, Kathryn Yurechko, Pranati Dani, Quan Ze Chen, and Amy X Zhang

  68. [76]

    Cai Yang, Sepehr Mousavi, Abhisek Dash, Krishna P Gummadi, and Ingmar Weber. 2025. Studying behavioral addiction by combining surveys and digital traces: A case study of TikTok. InProceedings of the International AAAI Conference on Web and Social Media, Vol. 19. 2106–2123

    2025

  69. [77]

    YouTube. 2026. YouTube for Press. https://blog.youtube/press/. Accessed: 2026-03-21

    2026

  70. [78]

    Excellent

    Savvas Zannettou, Olivia Nemes-Nemeth, Oshrat Ayalon, Angelica Goetzen, Krishna P Gummadi, Elissa M Redmiles, and Franziska Roesner. 2024. Analyzing user engagement with TikTok’s short format video recommendations using data donations. InProceedings of the 2024 CHI Conference ...

    2024

  71. [80]

    InProceedings of the 2025 CHI Conference on Human Factors in Computing Systems

    End user authoring of personalized content classifiers: comparing exam- ple labeling, rule writing, and LLM prompting. InProceedings of the 2025 CHI Conference on Human Factors in Computing Systems. 1–21

    2025

  72. [85]

    {metadata_text} Model: gpt-4.1-nano (metadata only) Based on the following information about a video, generate:

    Exactly {settings.KEYWORD_COUNT} keywords for matching this video to user interests. {metadata_text} Model: gpt-4.1-nano (metadata only) Based on the following information about a video, generate:

  73. [86]

    A concise description of what the video is about

  74. [87]

    {metadata_text} {metadata_text}expands to a block of Title, Description, Channel, Categories, andTagslines

    Exactly {settings.KEYWORD_COUNT} keywords for matching this video to user interests. {metadata_text} {metadata_text}expands to a block of Title, Description, Channel, Categories, andTagslines. Preference description bullets.Authors or revises the bullet-point description of a ...

  75. [91]

    Use IDs starting from 1

  76. [92]

    SYSTEM" for all bullets Return the bullet list as a JSON array with format: [{

    Set owner to "SYSTEM" for all bullets Return the bullet list as a JSON array with format: [{"id": 1, " owner": "SYSTEM", "text": "..."}] Model: gpt-5.2 (update / regeneration) Preference title provided by user: {preference_name} {examples_section}Update the bullet points to re...

  77. [93]

    Generate up to 5 bullet points describing this preference

  78. [94]

    Keep bullets concise (10 words max)

  79. [95]

    Each bullet must represent a unique aspect of content

  80. [96]

    You may ADD, EDIT, or DELETE bullets with owner "SYSTEM"

  81. [97]

    You may NOT modify or remove bullets with owner "USER"

  82. [98]

    Ads/Promotional

    For new bullets, use incrementing IDs starting from {next_id} Current Bullets: {current_bullets_json} Return the updated bullet list as a JSON array. When example videos exist,{examples_section}is a Example videos of preference provided by user:header followed byVideo 1: {desc...

    2026

  83. [2015]

    InProceedings of the 2015 ACM international joint conference on pervasive and ubiquitous computing

    SleepTight: low-burden, self-monitoring technology for capturing and reflecting on sleep behaviors. InProceedings of the 2015 ACM international joint conference on pervasive and ubiquitous computing. 121–132

    2015

  84. [2017]

    Proceedings of the National Academy of Sciences114, 28 (2017), 7313–7318

    Emotion shapes the diffusion of moralized content in social networks. Proceedings of the National Academy of Sciences114, 28 (2017), 7313–7318

    2017

  85. [2024]

    InProceedings of the 2024 CHI conference on human factors in computing systems

    InteractOut: leveraging interaction proxies as input manipulation strategies for reducing smartphone overuse. InProceedings of the 2024 CHI conference on human factors in computing systems. 1–19

    2024

  86. [2025]

    In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems

    How do HCI researchers study cognitive biases? A scoping review. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems. 1–20

    2025

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.