Platform architecture determines whether recommendation algorithms can shape information quality on social media

David A. Broniatowski; Erica Gralla; Giovanni Luca Ciampaglia; Joseph Simons; Manan Suri; Mohammad Hammas Saeed

arxiv: 2605.19204 · v1 · pith:5CLGD57Qnew · submitted 2026-05-19 · 💻 cs.SI · cs.HC

Platform architecture determines whether recommendation algorithms can shape information quality on social media

Mohammad Hammas Saeed , David A. Broniatowski , Joseph Simons , Erica Gralla , Manan Suri , Giovanni Luca Ciampaglia This is my paper

Pith reviewed 2026-05-20 03:06 UTC · model grok-4.3

classification 💻 cs.SI cs.HC

keywords platform architecturerecommendation algorithmsinformation qualitysocial mediaagent-based simulationinformation spreadpublic discourse

0 comments

The pith

Platform architecture decides whether recommendation algorithms can improve or harm information quality on social media.

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

The paper uses agent-based simulations to test how platform structures interact with recommendation algorithms in shaping information spread and quality. By varying four architectures ranked by flexibility and two algorithms, it shows negligible algorithmic effects on rigid tree structures, modest positive effects on hierarchies and networks, and strong negative winner-take-all effects on complete graphs. This matters because it indicates that fundamental platform design choices can override or neutralize the influence of algorithm selection on public discourse.

Core claim

Through agent-based simulation that orthogonally varies platform architecture and recommendation algorithm, the authors find that the popularity-based Hot algorithm produces no detectable effect on information spread or quality in tree architectures, modest positive effects in layered hierarchies and networks, and strong negative effects in complete graph architectures by inducing unpredictable winner-take-all dynamics unrelated to content quality.

What carries the argument

Orthogonal manipulation of four prototypical platform architectures (tree, layered hierarchy, network, complete graph) ranked by flexibility together with two recommendation algorithms (chronological LIFO and popularity Hot) to measure resulting changes in information spread and quality.

If this is right

On tree-like platforms the choice of recommendation algorithm has no measurable effect on information quality or spread.
On layered hierarchy and network platforms the popularity algorithm modestly improves both the reach and quality of information.
On complete graph platforms the popularity algorithm produces strong negative effects on quality and spread through winner-take-all dynamics.
Architectural constraints act as a stronger control on information quality than the specific algorithm in use.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Regulators targeting algorithm design alone may achieve little on highly flexible platforms and may be unnecessary on rigid ones.
Real platforms could test the simulations by comparing quality metrics before and after architecture changes such as adding or removing hierarchical layers.
The results point toward designing new platforms with deliberate flexibility limits rather than relying on post-hoc algorithm fixes.

Load-bearing premise

The ranking of platform architectures by flexibility from prior theory correctly predicts the magnitude of any algorithmic effects on information quality and spread.

What would settle it

Finding that the popularity algorithm changes information quality and spread by similar amounts across all four simulated architecture types would falsify the claim that architecture determines algorithmic impact.

Figures

Figures reproduced from arXiv: 2605.19204 by David A. Broniatowski, Erica Gralla, Giovanni Luca Ciampaglia, Joseph Simons, Manan Suri, Mohammad Hammas Saeed.

**Figure 1.** Figure 1: Hot ÷ LIFO ratio of mean number of agents reached or exposed per message, in log scale. The dashed line corresponds to no effect (1× ratio). Platforms are ordered by Moses’ flexibility metric: Reddit (tree) < Facebook (layered) < Twitter (network) < TikTok (complete graph). Panel A: Reach; the Hot algorithm reduces the total number of agents reached per message monotonically with architectural flexibility… view at source ↗

**Figure 2.** Figure 2: Panel A shows the breadth of exposure and engagement (broken down by engagement [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

**Figure 3.** Figure 3: Quality, breadth, and depth of information exposure by platform and algorithm. Each [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

**Figure 4.** Figure 4: Content quality and production order of exposed messages for all architectures under [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗

**Figure 5.** Figure 5: Seed-level mean information quality of agent exposures by platform [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗

**Figure 6.** Figure 6: Network structures for TikTok (left-most), Twitter, Facebook and Reddit (right-most) [PITH_FULL_IMAGE:figures/full_fig_p016_6.png] view at source ↗

read the original abstract

Social media platforms shape public discourse through two fundamental design choices that naturally co-occur in any field investigation: platform architecture, which defines what types of actors exist and how they interact, and recommendation algorithm, which determines what content is surfaced to users. Using agent-based simulation, we orthogonally manipulate both factors, exploring four prototypical architectures -- tree (e.g., Reddit), layered hierarchy (e.g., Facebook), network (e.g., Twitter), and complete graph (e.g., TikTok) -- and two algorithms: chronological (LIFO) and popularity-based (Hot). Drawing on prior theory that identifies and ranks canonical system architectures in terms of their flexibility we hypothesize that algorithmic effects on information spread and quality should be largest on the most flexible platforms and smallest on the most constrained ones. We find strong confirmation of this prediction. On tree-like platforms like Reddit, the algorithm has no detectable effect on information spread and quality. On layered hierarchies and networks like Facebook and Twitter, respectively, the Hot algorithm has modest positive effects on both the spread of information and its quality. On complete structures like TikTok, the Hot algorithm leads to a winner-take-all dynamics that has strong negative effects on both information spread and quality, making the relation between content quality and popularity unpredictable. These findings imply that architectural considerations are more powerful levers than algorithmic interventions for the design of healthy online spaces and public discourse. Platform reform efforts focused exclusively on algorithm choice may be insufficient on architecturally unconstrained platforms and unnecessary on architecturally constrained ones.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The simulation finds algorithmic effects on information quality scale with platform flexibility, but the graph topologies may embed that ranking by construction.

read the letter

The main takeaway is that this agent-based simulation finds recommendation algorithms have little detectable effect on information spread and quality in constrained architectures like trees, modest positive effects in hierarchies and networks, and strong negative winner-take-all dynamics in complete graphs. Architecture appears to set the scope for what algorithms can do. The orthogonal crossing of four architectures with two algorithms in one framework is the clearest new element. It lets them test the prediction from prior flexibility theory that effects should be largest where platforms are most flexible. The directional pattern matches the hypothesis and gives a practical distinction between architecture and algorithm as design levers. The paper does a reasonable job linking the structures to real platforms and keeping the setup simple enough to interpret. A soft spot is the risk that the results follow mechanically from the graph choices rather than testing flexibility as an independent factor. The architectures are implemented as distinct interaction graphs whose connectivity differences could produce the observed ordering without separate calibration to quantitative flexibility metrics from the cited theory. The abstract also leaves out parameter settings, the exact definition of information quality, and any statistical or sensitivity details, which makes it harder to judge how robust the winner-take-all finding is. This is aimed at researchers working on platform design, information diffusion, or computational models of social media. Readers thinking about regulatory focus or intervention priorities would get the most direct value from the architecture-versus-algorithm framing. The work is coherent enough on its own terms to merit a full referee process. I would send it to peer review with requests for explicit calibration checks on the graphs and more on parameter robustness.

Referee Report

2 major / 2 minor

Summary. The manuscript uses agent-based simulation to orthogonally vary platform architecture (tree, layered hierarchy, network, complete graph) and recommendation algorithm (LIFO chronological vs. Hot popularity-based). It tests the prediction, drawn from prior theory on system flexibility, that algorithmic effects on information spread and quality will be smallest on the most constrained architectures and largest on the most flexible ones. Results indicate null effects on tree structures, modest positive effects on hierarchies and networks, and strong negative winner-take-all effects on complete graphs, supporting the claim that architecture is the dominant factor and that algorithm-focused reforms may be insufficient or unnecessary depending on architecture.

Significance. If the central findings hold after addressing the issues below, the work offers a useful demonstration that simulation can isolate the interaction between structural and algorithmic design choices in ways observational studies cannot. It provides a concrete basis for prioritizing architectural constraints in platform governance discussions and could inform debates on whether regulating recommendation systems alone is adequate for improving discourse quality.

major comments (2)

[Methods] Methods (model specification section): The flexibility ranking is imported from prior theory, yet the four architectures are implemented directly as distinct interaction graphs whose connectivity properties differ by construction. No quantitative calibration or independent flexibility metric (e.g., a scalar derived from the cited theory and applied to each graph) is reported to confirm that the observed ordering of algorithmic effects tracks flexibility rather than raw differences in propagation potential. This leaves open the possibility that the pattern is tautological with the graph topologies chosen.
[Results] Results (simulation outcomes and statistical reporting): The abstract and results describe 'strong negative effects' and 'winner-take-all dynamics' on complete graphs, but the manuscript does not report effect sizes, number of simulation replications, confidence intervals, or robustness checks across parameter ranges. Without these, it is not possible to judge whether the null result on trees and the modest effects on hierarchies/networks are statistically distinguishable from noise or sensitive to specific parameter choices.

minor comments (2)

[Abstract] The abstract states that the Hot algorithm has 'modest positive effects' on hierarchies and networks; a brief parenthetical note on the direction and approximate magnitude of these effects would improve readability.
[Methods] Clarify in the methods how 'information quality' is operationalized (e.g., is it a binary label, a continuous score, or derived from an external benchmark?).

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their detailed and constructive report. We address each major comment below and indicate the revisions we will make to strengthen the manuscript.

read point-by-point responses

Referee: [Methods] Methods (model specification section): The flexibility ranking is imported from prior theory, yet the four architectures are implemented directly as distinct interaction graphs whose connectivity properties differ by construction. No quantitative calibration or independent flexibility metric (e.g., a scalar derived from the cited theory and applied to each graph) is reported to confirm that the observed ordering of algorithmic effects tracks flexibility rather than raw differences in propagation potential. This leaves open the possibility that the pattern is tautological with the graph topologies chosen.

Authors: The flexibility ordering is taken directly from the cited prior theory on system architectures, with each graph selected as a canonical instantiation of one level in that ranking. The differing connectivity properties are not incidental but are the structural features that define flexibility in the theory (limited propagation in trees versus unconstrained spread in complete graphs). To address the concern, we will add to the methods section a supplementary table of standard graph metrics (average degree, diameter, clustering coefficient) for each architecture and show how these align with the theoretical flexibility sequence. This provides an explicit, if post-hoc, calibration linking topology to the predicted ordering of algorithmic effects. revision: yes
Referee: [Results] Results (simulation outcomes and statistical reporting): The abstract and results describe 'strong negative effects' and 'winner-take-all dynamics' on complete graphs, but the manuscript does not report effect sizes, number of simulation replications, confidence intervals, or robustness checks across parameter ranges. Without these, it is not possible to judge whether the null result on trees and the modest effects on hierarchies/networks are statistically distinguishable from noise or sensitive to specific parameter choices.

Authors: We agree that the statistical reporting should be expanded. In the revised results section we will state the number of replications per condition, report standardized effect sizes and 95% confidence intervals for the key comparisons between algorithms within each architecture, and add a robustness subsection that varies core parameters (user attention threshold and content quality distribution) over plausible ranges. These additions will allow readers to assess whether the null finding on trees and the modest effects on hierarchies and networks are robust and distinguishable from the strong effects on complete graphs. revision: yes

Circularity Check

0 steps flagged

Simulation uses independent graph structures and external benchmarks; no reduction to self-defined or fitted quantities

full rationale

The paper orthogonally manipulates four distinct interaction graphs (tree, layered hierarchy, network, complete) and two algorithms in an agent-based simulation, measuring effects on spread and quality. The flexibility ranking is drawn from prior theory and mapped to real platform examples, but the observed pattern (null on trees, modest on hierarchies/networks, strong negative on complete graphs) emerges from explicit simulation runs rather than by construction from any parameter fit or self-referential definition within this manuscript. No equations or results reduce to inputs by definition, and the central claim retains independent content from the simulation outcomes against external benchmarks. This is a normal low-circularity finding for a simulation study.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that the four chosen graph topologies accurately represent the flexibility ordering from prior theory and that the two algorithms capture the main real-world contrast; no free parameters or invented entities are described in the abstract.

axioms (1)

domain assumption Prior theory correctly ranks canonical system architectures by flexibility (tree least flexible, complete graph most flexible).
This ranking is invoked to generate the hypothesis that algorithmic effects should be largest on the most flexible platforms.

pith-pipeline@v0.9.0 · 5824 in / 1226 out tokens · 24619 ms · 2026-05-20T03:06:46.339560+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Foundation/AlexanderDuality.lean alexander_duality_circle_linking unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

four prototypical architectures – tree (e.g., Reddit), layered hierarchy (e.g., Facebook), network (e.g., Twitter), and complete graph (e.g., TikTok) … Drawing on prior theory that identifies and ranks canonical system architectures in terms of their flexibility
IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Moses’ theory of generic architectures ranks architectures by flexibility … the effect of a recommendation algorithm on information spread and quality should be largest on the most architecturally flexible platforms

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

38 extracted references · 38 canonical work pages

[1]

J. B. Bak-Coleman, I. Kennedy, M. Wack, A. Beers, J. S. Schafer, E. S. Spiro, K. Starbird, and J. D. West. Combining interventions to reduce the spread of viral misinformation.Nature Human Behaviour, 6(10):1372–1380, 2022

work page 2022
[2]

Baumann, D

F. Baumann, D. Halpern, A. D. Procaccia, I. Rahwan, I. Shapira, and M. Wüthrich. Optimal engagement-diversity tradeoffs in social media. InProceedings of the ACM Web Conference 2024, pages 288–299, 2024. 17

work page 2024
[3]

W. J. Brady, J. C. Jackson, B. Lindström, and M. J. Crockett. Algorithm-mediated social learning in online social networks.Trends in Cognitive Sciences, 27(10):947–960, Oct. 2023

work page 2023
[4]

D. A. Broniatowski, K. M. Hilyard, and M. Dredze. Effective vaccine communication during the disneyland measles outbreak.Vaccine, 34(28):3225–3228, 2016

work page 2016
[5]

D. A. Broniatowski, P. Hosseini, E. V. Porter, and T. J. Wood. The role of mental repre- sentation in sharing misinformation online.Journal of Experimental Psychology: Applied, 30(4):511–538, Dec. 2024

work page 2024
[6]

D. A. Broniatowski, D. Kerchner, F. Farooq, X. Huang, A. M. Jamison, M. Dredze, S. C. Quinn, and J. W. Ayers. Twitter and facebook posts about covid-19 are less likely to spread misin- formation compared to other health topics.PLOS ONE, 17(1):e0261768, 2022

work page 2022
[7]

D. A. Broniatowski and J. Moses. Flexibility, complexity, and controllability in large scale systems, 2014

work page 2014
[8]

D. A. Broniatowski and J. Moses. Measuring flexibility, descriptive complexity, and rework potential in generic system architectures.Systems Engineering, 19(3):207–221, 2016

work page 2016
[9]

D. A. Broniatowski and V. F. Reyna. To illuminate and motivate: a fuzzy-trace model of the spread of information online.Computational and Mathematical Organization Theory, 26(4):431–464, Aug. 2019

work page 2019
[10]

D. A. Broniatowski, J. R. Simons, J. Gu, A. M. Jamison, and L. C. Abroms. The efficacy of facebook’s vaccine misinformation policies and architecture during the covid-19 pandemic. Science Advances, 9(37):eadh2132, 2023

work page 2023
[11]

D. A. Broniatowski, W. Zhong, J. R. Simons, A. M. Jamison, M. Dredze, and L. C. Abroms. Explaining twitter’s inability to effectively moderate content during the covid-19 pandemic. Scientific Reports, 15(1), oct 2025

work page 2025
[12]

Y. T. Cao, L.-F. Domingo, S. Gilbert, M. L. Mazurek, K. Shilton, and H. Daumé Iii. Toxi- city detection is NOT all you need: Measuring the gaps to supporting volunteer content moderators through a user-centric method. In Y. Al-Onaizan, M. Bansal, and Y.-N. Chen, editors,Proceedings of the 2024 Conference on Empirical Methods in Natural Language Pro- cessi...

work page 2024
[13]

G. L. Ciampaglia, A. Nematzadeh, F. Menczer, and A. Flammini. How algorithmic popularity bias hinders or promotes quality.Scientific reports, 8(1):15951, 2018

work page 2018
[14]

L. Cima, B. Tessa, S. Cresci, A. Trujillo, and M. Avvenuti. Investigating the heterogenous effects of a massive content moderation intervention via difference-in-differences, 2025

work page 2025
[15]

D. Eckles. Algorithmic transparency and assessing effects of algorithmic ranking. Technical report, SocArXiV, apr 2022. 18

work page 2022
[16]

Huszár, S

F. Huszár, S. I. Ktena, C. O’Brien, L. Belli, A. Schlaikjer, and M. Hardt. Algorithmic amplification of politics on Twitter.Proceedings of the National Academy of Sciences, 119(1):e2025334119, Jan. 2022

work page 2022
[17]

Jhaver, S

S. Jhaver, S. Frey, and A. X. Zhang. Decentralizing Platform Power: A Design Space of Multi-Level Governance in Online Social Platforms.Social Media + Society, 9(4):20563051231207857, Oct. 2023

work page 2023
[18]

Kleppmann, P

M. Kleppmann, P. Frazee, J. Gold, J. Graber, D. Holmgren, D. Ivy, J. Johnson, B. Newbold, and J. Volpert. Bluesky and the at protocol: Usable decentralized social media. InProceedings of the ACM Conext-2024 Workshop on the Decentralization of the Internet, DIN ’24, page 1–7, New York, NY, USA, 2024. Association for Computing Machinery

work page 2024
[19]

Leskovec

J. Leskovec. Twitter follower network. https://snap.stanford.edu/data/twitter-2010.html, 2010

work page 2010
[20]

Leskovec and C

J. Leskovec and C. Faloutsos. Sampling from large graphs. InProceedings of the 12th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’06, page 631–636, New York, NY, USA, 2006. Association for Computing Machinery

work page 2006
[21]

M. Li, J. Suk, Y. Zhang, J. C. Pevehouse, Y. Sun, H. Kwon, R. Lian, R. Wang, X. Dong, and D. V. Shah. Platform affordances, discursive opportunities, and social media activism: A cross-platform analysis of #MeToo on twitter, facebook, and reddit, 2017–2020.New Media & Society, 28(1):119–147, Oct. 2024

work page 2017
[22]

S. D. McCabe, D. Ferrari, J. Green, D. M. J. Lazer, and K. M. Esterling. Post-january 6th deplatforming reduced the reach of misinformation on twitter.Nature, 630(8015):132–140, 2024

work page 2024
[23]

Milli, M

S. Milli, M. Carroll, Y. Wang, S. Pandey, S. Zhao, and A. D. Dragan. Engagement, user satis- faction, and the amplification of divisive content on social media.PNAS Nexus, 4(3):pgaf062, 03 2025

work page 2025
[24]

Moehring, A

A. Moehring, A. Cooper, A. Narayanan, A. Ovadya, E. Redmiles, J. Allen, J. Stray, J. Kamin, L. Sigerson, L. Thorburn, M. Motyl, M. Eslami, N. F. Johnson, N. Lubin, R. Iyer, and Z. Arnao. Better feeds: Algorithms that put people first. resreport, Knight–Georgetown Institute, Mar. 2025

work page 2025
[25]

Orben and J

A. Orben and J. N. Matias. Fixing the science of digital technology harms.Science, 388(6743):152–155, 2025

work page 2025
[26]

Ovadya and L

A. Ovadya and L. Thorburn. Bridging systems: Open problems for countering destructive divisiveness across ranking, recommenders, and governance. Technical report, CoRR, jan

work page
[27]

arXiv:2301.09976 [cs]

work page arXiv
[28]

Piccardi, M

T. Piccardi, M. Saveski, C. Jia, J. Hancock, J. L. Tsai, and M. S. Bernstein. Reranking par- tisan animosity in algorithmic social media feeds alters affective polarization.Science, 390(6776):eadu5584, Nov. 2025. arXiv:2411.14652 [cs]. 19

work page arXiv 2025
[29]

Sasahara, W

K. Sasahara, W. Chen, H. Peng, G. L. Ciampaglia, A. Flammini, and F. Menczer. Social influ- ence and unfollowing accelerate the emergence of echo chambers.Journal of Computational Social Science, 4(1):381–402, 2021

work page 2021
[30]

P. E. Smaldino, A. Russell, M. R. Zefferman, J. Donath, J. G. Foster, D. Guilbeault, M. Hilbert, E. A. Hobson, K. Lerman, H. Miton, et al. Information architectures: a framework for un- derstanding socio-technical systems.npj Complexity, 2(1):13, 2025

work page 2025
[31]

van der Linden

S. van der Linden. How influencers and algorithms mobilize propaganda—and distort reality. Nature, 633(8029), 2024

work page 2024
[32]

S. Wang, S. Huang, A. Zhou, and D. Metaxa. Lower quantity, higher quality: Auditing news content and user perceptions on twitter/x algorithmic versus chronological timelines. Proceedings of the ACM on Human-Computer Interaction, 8(CSCW2):1–25, 2024

work page 2024
[33]

Wei and D

Z. Wei and D. A. Broniatowski. Characterizing system architectures using network data. Procedia Computer Science, 153:301–308, 2019

work page 2019
[34]

L. Weng, A. Flammini, A. Vespignani, and F. Menczer. Competition among memes in a world with limited attention.Scientific reports, 2(1):335, 2012

work page 2012
[35]

A. X. Zhang, M. S. Bernstein, D. R. Karger, and M. S. Ackerman. Form-from: A design space of social media systems.Proceedings of the ACM on Human-Computer Interaction, 8(CSCW1):1–47, 2024. 20 Supplementary Information A User characteristics A user/pagea∈Ais an agent described by the following traits: •Quality preferenceϕ a ∈[−1,1]: A scalar reflecting the...

work page 2024
[36]

Motivational valence (i.e., degree of appeal or clickbait content)k m ∈[0,1], indicating how attention-grabbing the message is

work page
[37]

Illuminating powerβ m ∈[0,1], indicating how meaningful or insightful the message is

work page
[38]

These quantities are sampled for each message

Informativeness orqualityα m ∈[−1,1], where a value of−1indicates a misleading mes- sage,0neutral, and1and informative and factual message. These quantities are sampled for each message. Bothk m andβ m are sampled from the same triangular distribution to model skewed preferences, such as most messages being moderately engaging but not highly insightful. T...

work page

[1] [1]

J. B. Bak-Coleman, I. Kennedy, M. Wack, A. Beers, J. S. Schafer, E. S. Spiro, K. Starbird, and J. D. West. Combining interventions to reduce the spread of viral misinformation.Nature Human Behaviour, 6(10):1372–1380, 2022

work page 2022

[2] [2]

Baumann, D

F. Baumann, D. Halpern, A. D. Procaccia, I. Rahwan, I. Shapira, and M. Wüthrich. Optimal engagement-diversity tradeoffs in social media. InProceedings of the ACM Web Conference 2024, pages 288–299, 2024. 17

work page 2024

[3] [3]

W. J. Brady, J. C. Jackson, B. Lindström, and M. J. Crockett. Algorithm-mediated social learning in online social networks.Trends in Cognitive Sciences, 27(10):947–960, Oct. 2023

work page 2023

[4] [4]

D. A. Broniatowski, K. M. Hilyard, and M. Dredze. Effective vaccine communication during the disneyland measles outbreak.Vaccine, 34(28):3225–3228, 2016

work page 2016

[5] [5]

D. A. Broniatowski, P. Hosseini, E. V. Porter, and T. J. Wood. The role of mental repre- sentation in sharing misinformation online.Journal of Experimental Psychology: Applied, 30(4):511–538, Dec. 2024

work page 2024

[6] [6]

D. A. Broniatowski, D. Kerchner, F. Farooq, X. Huang, A. M. Jamison, M. Dredze, S. C. Quinn, and J. W. Ayers. Twitter and facebook posts about covid-19 are less likely to spread misin- formation compared to other health topics.PLOS ONE, 17(1):e0261768, 2022

work page 2022

[7] [7]

D. A. Broniatowski and J. Moses. Flexibility, complexity, and controllability in large scale systems, 2014

work page 2014

[8] [8]

D. A. Broniatowski and J. Moses. Measuring flexibility, descriptive complexity, and rework potential in generic system architectures.Systems Engineering, 19(3):207–221, 2016

work page 2016

[9] [9]

D. A. Broniatowski and V. F. Reyna. To illuminate and motivate: a fuzzy-trace model of the spread of information online.Computational and Mathematical Organization Theory, 26(4):431–464, Aug. 2019

work page 2019

[10] [10]

D. A. Broniatowski, J. R. Simons, J. Gu, A. M. Jamison, and L. C. Abroms. The efficacy of facebook’s vaccine misinformation policies and architecture during the covid-19 pandemic. Science Advances, 9(37):eadh2132, 2023

work page 2023

[11] [11]

D. A. Broniatowski, W. Zhong, J. R. Simons, A. M. Jamison, M. Dredze, and L. C. Abroms. Explaining twitter’s inability to effectively moderate content during the covid-19 pandemic. Scientific Reports, 15(1), oct 2025

work page 2025

[12] [12]

Y. T. Cao, L.-F. Domingo, S. Gilbert, M. L. Mazurek, K. Shilton, and H. Daumé Iii. Toxi- city detection is NOT all you need: Measuring the gaps to supporting volunteer content moderators through a user-centric method. In Y. Al-Onaizan, M. Bansal, and Y.-N. Chen, editors,Proceedings of the 2024 Conference on Empirical Methods in Natural Language Pro- cessi...

work page 2024

[13] [13]

G. L. Ciampaglia, A. Nematzadeh, F. Menczer, and A. Flammini. How algorithmic popularity bias hinders or promotes quality.Scientific reports, 8(1):15951, 2018

work page 2018

[14] [14]

L. Cima, B. Tessa, S. Cresci, A. Trujillo, and M. Avvenuti. Investigating the heterogenous effects of a massive content moderation intervention via difference-in-differences, 2025

work page 2025

[15] [15]

D. Eckles. Algorithmic transparency and assessing effects of algorithmic ranking. Technical report, SocArXiV, apr 2022. 18

work page 2022

[16] [16]

Huszár, S

F. Huszár, S. I. Ktena, C. O’Brien, L. Belli, A. Schlaikjer, and M. Hardt. Algorithmic amplification of politics on Twitter.Proceedings of the National Academy of Sciences, 119(1):e2025334119, Jan. 2022

work page 2022

[17] [17]

Jhaver, S

S. Jhaver, S. Frey, and A. X. Zhang. Decentralizing Platform Power: A Design Space of Multi-Level Governance in Online Social Platforms.Social Media + Society, 9(4):20563051231207857, Oct. 2023

work page 2023

[18] [18]

Kleppmann, P

M. Kleppmann, P. Frazee, J. Gold, J. Graber, D. Holmgren, D. Ivy, J. Johnson, B. Newbold, and J. Volpert. Bluesky and the at protocol: Usable decentralized social media. InProceedings of the ACM Conext-2024 Workshop on the Decentralization of the Internet, DIN ’24, page 1–7, New York, NY, USA, 2024. Association for Computing Machinery

work page 2024

[19] [19]

Leskovec

J. Leskovec. Twitter follower network. https://snap.stanford.edu/data/twitter-2010.html, 2010

work page 2010

[20] [20]

Leskovec and C

J. Leskovec and C. Faloutsos. Sampling from large graphs. InProceedings of the 12th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’06, page 631–636, New York, NY, USA, 2006. Association for Computing Machinery

work page 2006

[21] [21]

M. Li, J. Suk, Y. Zhang, J. C. Pevehouse, Y. Sun, H. Kwon, R. Lian, R. Wang, X. Dong, and D. V. Shah. Platform affordances, discursive opportunities, and social media activism: A cross-platform analysis of #MeToo on twitter, facebook, and reddit, 2017–2020.New Media & Society, 28(1):119–147, Oct. 2024

work page 2017

[22] [22]

S. D. McCabe, D. Ferrari, J. Green, D. M. J. Lazer, and K. M. Esterling. Post-january 6th deplatforming reduced the reach of misinformation on twitter.Nature, 630(8015):132–140, 2024

work page 2024

[23] [23]

Milli, M

S. Milli, M. Carroll, Y. Wang, S. Pandey, S. Zhao, and A. D. Dragan. Engagement, user satis- faction, and the amplification of divisive content on social media.PNAS Nexus, 4(3):pgaf062, 03 2025

work page 2025

[24] [24]

Moehring, A

A. Moehring, A. Cooper, A. Narayanan, A. Ovadya, E. Redmiles, J. Allen, J. Stray, J. Kamin, L. Sigerson, L. Thorburn, M. Motyl, M. Eslami, N. F. Johnson, N. Lubin, R. Iyer, and Z. Arnao. Better feeds: Algorithms that put people first. resreport, Knight–Georgetown Institute, Mar. 2025

work page 2025

[25] [25]

Orben and J

A. Orben and J. N. Matias. Fixing the science of digital technology harms.Science, 388(6743):152–155, 2025

work page 2025

[26] [26]

Ovadya and L

A. Ovadya and L. Thorburn. Bridging systems: Open problems for countering destructive divisiveness across ranking, recommenders, and governance. Technical report, CoRR, jan

work page

[27] [27]

arXiv:2301.09976 [cs]

work page arXiv

[28] [28]

Piccardi, M

T. Piccardi, M. Saveski, C. Jia, J. Hancock, J. L. Tsai, and M. S. Bernstein. Reranking par- tisan animosity in algorithmic social media feeds alters affective polarization.Science, 390(6776):eadu5584, Nov. 2025. arXiv:2411.14652 [cs]. 19

work page arXiv 2025

[29] [29]

Sasahara, W

K. Sasahara, W. Chen, H. Peng, G. L. Ciampaglia, A. Flammini, and F. Menczer. Social influ- ence and unfollowing accelerate the emergence of echo chambers.Journal of Computational Social Science, 4(1):381–402, 2021

work page 2021

[30] [30]

P. E. Smaldino, A. Russell, M. R. Zefferman, J. Donath, J. G. Foster, D. Guilbeault, M. Hilbert, E. A. Hobson, K. Lerman, H. Miton, et al. Information architectures: a framework for un- derstanding socio-technical systems.npj Complexity, 2(1):13, 2025

work page 2025

[31] [31]

van der Linden

S. van der Linden. How influencers and algorithms mobilize propaganda—and distort reality. Nature, 633(8029), 2024

work page 2024

[32] [32]

S. Wang, S. Huang, A. Zhou, and D. Metaxa. Lower quantity, higher quality: Auditing news content and user perceptions on twitter/x algorithmic versus chronological timelines. Proceedings of the ACM on Human-Computer Interaction, 8(CSCW2):1–25, 2024

work page 2024

[33] [33]

Wei and D

Z. Wei and D. A. Broniatowski. Characterizing system architectures using network data. Procedia Computer Science, 153:301–308, 2019

work page 2019

[34] [34]

L. Weng, A. Flammini, A. Vespignani, and F. Menczer. Competition among memes in a world with limited attention.Scientific reports, 2(1):335, 2012

work page 2012

[35] [35]

A. X. Zhang, M. S. Bernstein, D. R. Karger, and M. S. Ackerman. Form-from: A design space of social media systems.Proceedings of the ACM on Human-Computer Interaction, 8(CSCW1):1–47, 2024. 20 Supplementary Information A User characteristics A user/pagea∈Ais an agent described by the following traits: •Quality preferenceϕ a ∈[−1,1]: A scalar reflecting the...

work page 2024

[36] [36]

Motivational valence (i.e., degree of appeal or clickbait content)k m ∈[0,1], indicating how attention-grabbing the message is

work page

[37] [37]

Illuminating powerβ m ∈[0,1], indicating how meaningful or insightful the message is

work page

[38] [38]

These quantities are sampled for each message

Informativeness orqualityα m ∈[−1,1], where a value of−1indicates a misleading mes- sage,0neutral, and1and informative and factual message. These quantities are sampled for each message. Bothk m andβ m are sampled from the same triangular distribution to model skewed preferences, such as most messages being moderately engaging but not highly insightful. T...

work page