pith. sign in

arxiv: 2606.08294 · v1 · pith:ISJYSFAJnew · submitted 2026-06-06 · 💰 econ.TH

Platform-Driven Hate Speech: An Epidemiological Model with Optimal Taxation

Nazaria Solferino This is my paper

Pith reviewed 2026-06-27 18:42 UTC · model grok-4.3

classification 💰 econ.TH
keywords hate speechepidemiological modeloptimal taxationStackelberg equilibriumalgorithmic amplificationplatform incentivesbistabilityvictim harm
0
0 comments X

The pith

A government tax on algorithmic amplification reduces hate speech prevalence, removes bistability, and lowers victim harm in an epidemiological model of platform incentives.

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

The paper constructs an epidemiological model in which hate speech spreads according to platform algorithms whose reactivity feeds back into the platform's profits. A welfare-maximizing government chooses a tax on amplification that the profit-maximizing platform takes as given, yielding a Stackelberg equilibrium solved analytically and numerically. If the tax is set optimally it lowers the steady-state level of hate speech, collapses the region of bistability into a single stable low-prevalence equilibrium, and reduces the harm suffered by victims.

Core claim

In the model the platform's profit rises with hate-speech prevalence and with its own algorithmic reactivity, closing a feedback loop with the epidemic dynamics; the government then selects the tax rate that maximises social welfare net of tax revenue and deadweight loss, and the resulting equilibrium exhibits strictly lower hate-speech prevalence, elimination of bistability, and lower victim harm than the untaxed equilibrium.

What carries the argument

The closed feedback loop between platform profit, algorithmic reactivity, and epidemiological prevalence, solved as a Stackelberg game between platform and government.

If this is right

  • Hate speech prevalence falls at the new equilibrium.
  • Bistability disappears, so small shocks no longer trigger large jumps in prevalence.
  • Victim harm, measured by exposure or incidents, declines.
  • Tax revenue is collected while deadweight loss from reduced platform activity is balanced against the externality reduction.

Where Pith is reading between the lines

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

  • If real platforms exhibit similar profit-prevalence feedback, the same tax logic could be applied to other engagement-driven harms such as misinformation.
  • The model suggests testing whether observed jumps in hate-speech metrics after algorithm changes are consistent with the predicted bistable thresholds.
  • Extending the framework to multiple platforms could reveal whether coordinated taxation or competition among platforms alters the optimal tax rate.

Load-bearing premise

The platform's profit is assumed to depend directly on hate speech prevalence and algorithmic reactivity in a manner that generates a closed feedback loop with the epidemic dynamics.

What would settle it

Numerical simulation of the untaxed system showing two stable equilibria for a given parameter set, followed by the same simulation after the optimal tax is imposed showing collapse to a single low-prevalence equilibrium.

Figures

Figures reproduced from arXiv: 2606.08294 by Nazaria Solferino.

Figure 1
Figure 1. Figure 1: displays the time evolution of the infected fraction I(t) under the two regimes. The dashed red curve corresponds to the laissez-faire equilibrium. The solid blue curve corresponds to the optimal tax; it declines monotonically from the same initial condition and converges to I ∗ ≈ 0.004. Therefore, tax reduces the endemic level and tends to eliminate the epidemic peak entirely in the long-run: the laissez-… view at source ↗
Figure 2
Figure 2. Figure 2: Time evolution of the victim fraction H(t) under the laissez-faire equilibrium (dashed red) and under the optimal tax (solid blue). Victims are reduced by approxi￾mately 93% [PITH_FULL_IMAGE:figures/full_fig_p017_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Endemic prevalence I ∗ as a function of the basic reproduction number R0. The S-shaped curve indicates the presence of a bistable region. The vertical dashed line marks the laissez-faire R0; the optimal tax shifts the system leftwards, out of the bistable region. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: shows how the platform’s reactivity γ and the endemic prevalence I ∗ respond to the tax rate τ . The top panel displays the reaction function γ(τ ), obtained from the platform’s profit maximisation. As expected, γ is strictly decreasing in τ : a higher tax makes amplification more costly, inducing the platform to reduce it. At the optimal tax τ ∗ = 0.122, the reactivity reaches zero. The bottom panel shows… view at source ↗
Figure 5
Figure 5. Figure 5: Sensitivity analysis. Each panel shows the laissez-faire (red) and optimal tax [PITH_FULL_IMAGE:figures/full_fig_p019_5.png] view at source ↗
read the original abstract

Online hate speech is a global challenge amplified by engagement8-driven social media algorithms. This paper develops an epidemiological model of hate speech propagation capturing the strategic interaction between a profit-maximizing platform and a welfare-maximizing government. The platform's profit depends on the prevalence of hate speech and on its own algorithmic reactivity, creating a feedback loop between the epidemic and economic incentives. The government sets an optimal tax on amplification to internalize the social costs, balancing the benefit of tax revenue against the deadweight loss of taxation. The Stackelberg equilibrium is characterised analytically and solved numerically. The optimal tax reduces hate speech prevalence, eliminates bistability and lowers victim harm.

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.

Referee Report

2 major / 2 minor

Summary. The paper develops an epidemiological model of hate speech propagation on social media that incorporates strategic interaction between a profit-maximizing platform and a welfare-maximizing government. Platform profit is specified to depend directly on hate-speech prevalence p and algorithmic reactivity r, generating a closed feedback loop. The government acts as Stackelberg leader and chooses an optimal tax on amplification; the resulting equilibrium is characterized analytically and solved numerically. The central results are that the optimal tax lowers steady-state prevalence, eliminates bistability, and reduces victim harm.

Significance. If the modeling assumptions hold, the work supplies a novel integration of compartmental epidemic dynamics with platform incentives and optimal taxation, yielding a concrete policy instrument (tax on amplification) whose comparative-static effects on prevalence and stability can be derived. The analytical characterization of the Stackelberg equilibrium and the numerical demonstration of bistability elimination are strengths that could inform regulatory design, provided the profit specification is robust.

major comments (2)
  1. [Model section (profit function definition)] The functional dependence of platform profit π on hate-speech prevalence p (explicitly stated in the abstract and used to close the feedback loop) is load-bearing for the claim that the optimal tax eliminates bistability. If instead π depends only on aggregate engagement or ad revenue independent of p, the platform’s best-response function changes and the government’s Stackelberg problem no longer produces the reported disappearance of multiple steady states. The manuscript should either provide empirical or micro-founded justification for π(p,r) or report robustness checks under alternative specifications.
  2. [Numerical results / comparative-statics section] The numerical results on elimination of bistability and reduction in victim harm rest on the specific parameterization of the profit function and the tax instrument. Without reported sensitivity analysis on the elasticity of π with respect to p (or on the functional form of the amplification tax), it is unclear whether the policy conclusions survive modest changes in the maintained assumptions.
minor comments (2)
  1. [Notation and model setup] Notation for the state variables (prevalence p, reactivity r) and the tax instrument should be introduced once and used consistently; currently the abstract and model description introduce overlapping symbols without a consolidated table.
  2. [Equilibrium characterization] The abstract claims an “analytical characterization” of the Stackelberg equilibrium; the main text should state the precise conditions (e.g., concavity assumptions or closed-form best-response functions) under which the characterization holds.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive comments on our manuscript. We address each of the major comments below and indicate the revisions we plan to make.

read point-by-point responses

  1. Referee: [Model section (profit function definition)] The functional dependence of platform profit π on hate-speech prevalence p (explicitly stated in the abstract and used to close the feedback loop) is load-bearing for the claim that the optimal tax eliminates bistability. If instead π depends only on aggregate engagement or ad revenue independent of p, the platform’s best-response function changes and the government’s Stackelberg problem no longer produces the reported disappearance of multiple steady states. The manuscript should either provide empirical or micro-founded justification for π(p,r) or report robustness checks under alternative specifications.

    Authors: The specification of platform profit depending on hate speech prevalence p is motivated by the economic incentive for platforms to amplify content that drives higher user engagement and thus ad revenue. This is supported by numerous studies on social media dynamics. To strengthen the paper, we will add a micro-founded derivation of the profit function based on a simple model of content virality and user time allocation. We will also include robustness checks under an alternative profit specification that depends only on aggregate engagement, showing that the optimal tax continues to reduce prevalence and harm, although the complete elimination of bistability holds under a broader set of conditions. revision: yes

  2. Referee: [Numerical results / comparative-statics section] The numerical results on elimination of bistability and reduction in victim harm rest on the specific parameterization of the profit function and the tax instrument. Without reported sensitivity analysis on the elasticity of π with respect to p (or on the functional form of the amplification tax), it is unclear whether the policy conclusions survive modest changes in the maintained assumptions.

    Authors: We agree that sensitivity analysis would enhance the credibility of the numerical findings. In the revised version of the manuscript, we will expand the numerical section to include sensitivity analyses with respect to the elasticity parameter in the profit function and alternative forms of the tax instrument. These additional results will demonstrate that the key conclusions regarding the reduction in prevalence, elimination of bistability, and lower victim harm are robust to variations in these parameters within empirically plausible ranges. revision: yes

Circularity Check

0 steps flagged

No circularity: explicit assumptions and analytical characterization stand independently

full rationale

The paper states an explicit modeling assumption that platform profit depends on hate-speech prevalence p and reactivity r, then derives the Stackelberg equilibrium analytically from the resulting coupled epidemic-economic system. No quoted equations reduce a claimed prediction or equilibrium outcome to a fitted parameter or self-citation by construction; the feedback loop is introduced as a premise rather than derived from the target results. The abstract and skeptic summary contain no self-definitional steps, no renaming of known results, and no load-bearing self-citations. The derivation chain therefore remains self-contained against external benchmarks once the functional form of π(p,r) is accepted as given.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review yields no explicit free parameters, axioms, or invented entities; the central claim rests on an unstated functional dependence between platform profit and hate-speech prevalence that is not detailed here.

pith-pipeline@v0.9.1-grok · 5627 in / 1010 out tokens · 28761 ms · 2026-06-27T18:42:11.698673+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

40 extracted references · 5 canonical work pages

  1. [1]

    Accountability Paradox Research Group. (2024). The accountability paradox: How platform API restrictions undermine AI transparency mandates.arXiv preprint, arXiv:2405.01234. https://doi.org/10.48550/arXiv.2405.01234

    work page doi:10.48550/arxiv.2405.01234 2024

  2. [2]

    (2021).Online hate and harassment: The American expe- rience 2021

    Anti-Defamation League. (2021).Online hate and harassment: The American expe- rience 2021. Center for technology and society, 10-23

    2021

  3. [3]

    ADL Center for Technology and Society, and Tech Transparency Project.From bad to worse: Algorithmic amplification of antisemitism and extremism on major platforms. ADL

  4. [4]

    Applied Network Science. (2021). Estimation of the reproduction num- ber of hate speech on Twitter.Applied Network Science,6(1), 45. https://doi.org/10.1007/s41109-021-00389-0

    work page doi:10.1007/s41109-021-00389-0 2021

  5. [5]

    (2025).Hate Speech Index – Q1 2025

    Areto Labs. (2025).Hate Speech Index – Q1 2025. Areto Labs. https://www.aretolabs.com/hate-speech-index

    2025

  6. [6]

    Badr, P., Carminati, B., & Ferrari, E. (2021). A SIR model for the diffusion of ideas in social networks.Physica A: Statistical Mechanics and its Applications,568, 125721. https://doi.org/10.1016/j.physa.2020.125721

    work page doi:10.1016/j.physa.2020.125721 2021

  7. [7]

    J., Wills, J

    Brady, W. J., Wills, J. A., Jost, J. T., Tucker, J. A., Van Bavel, J. J. (2017). Emotion shapes the diffusion of moralized content in social networks.Nature Human Behaviour,1(7), 0116

    2017

  8. [8]

    Castillo-Chavez, C., & Song, B. (2004). Dynamical models of tuberculosis and their applications.Mathematical Biosciences and Engineering,1(2), 361–404. https://doi.org/10.3934/mbe.2004.1.361

    work page doi:10.3934/mbe.2004.1.361 2004

  9. [9]

    Chadwick, A., Vaccari, C., Kaiser, J. (2025). The amplification of exaggerated and false news on social media: The roles of platform use, motivations, affect, and ideol- ogy.American Behavioral Scientists,69(2), 113-130

    2025

  10. [10]

    Del Vicario, M., Vivaldo, G., Bessi, A., Zollo, F., Scala, A., Caldarelli, G., & Quat- trociocchi, W. (2016). Echo chambers: Emotional contagion and group polarization on Facebook.Scientific Reports,6(1), 37825

    2016

  11. [11]

    R., Aiyappa, R., Pacheco, D., Bryden, J., & Menczer, F

    DeVerna, M. R., Aiyappa, R., Pacheco, D., Bryden, J., & Menczer, F. (2024). Identi- fying and characterizing superspreaders of low-credibility content on Twitter.PLOS ONE,19(5), e0302201

    2024

  12. [12]

    (2025).Revised Code of Conduct on Countering Illegal Hate Speech Online integrated into the Digital Services Act

    European Commission. (2025).Revised Code of Conduct on Countering Illegal Hate Speech Online integrated into the Digital Services Act. Brussels: European Commis- sion

    2025

  13. [13]

    K., Teklu, S

    Hailu, G. K., Teklu, S. W (2024). Investigation into hate speech dissemination dy- namics in a community using fractional order modeling approach.Research in Math- ematics,178, 114321. 24

    2024

  14. [14]

    (2026).Online hate surge threatens public safety: 2025–2026 global report

    Global Safety Alliance. (2026).Online hate surge threatens public safety: 2025–2026 global report. International Security and Digital Rights Press

    2026

  15. [15]

    Kramer, A. D. I., Guillory, J. E., & Hancock, J. T. (2014). Exper- imental evidence of massive-scale emotional contagion through social net- works.Proceedings of the National Academy of Sciences,111(24), 8788–8790. https://doi.org/10.1073/pnas.1320040111

    work page doi:10.1073/pnas.1320040111 2014

  16. [16]

    (2025, May).Magnifica Humanitas: Encyclical letter on the digital age and human dignity

    Pope Leo XIV. (2025, May).Magnifica Humanitas: Encyclical letter on the digital age and human dignity. Vatican City: Libreria Editrice Vaticana

    2025

  17. [17]

    Effects of Prebunking Interventions on Misinformation: A Systematic Review and Meta- Analysis

    Lopez, R., Molina, A., Iglesias-Parro, S., Soriano, M., Ordu˜ na, A., Arbol, J. Effects of Prebunking Interventions on Misinformation: A Systematic Review and Meta- Analysis

  18. [18]

    Maarouf, P., Prollochs, J., & Ferrieguel, S. (2024). The virality of hate speech on social media.Proceedings of the ACM on Human-Computer Interaction,8(CSCW1), 1-22

    2024

  19. [19]

    and Paille S.(2025)

    Madriaza, P., Hassan, G., Brouillette-Alarie, S., Mounchingam, A.N., Durocher- Corfa, L., Borokhovski, E., Pickup, D. and Paille S.(2025). Exposure to hate in online and traditional media: A systematic review and meta-analysis of the impact of this exposure on individuals and communities.Campbell Systematic Reviews,21(1), cl2-70018

    2025

  20. [20]

    (2023).Economic cost of hate crimes

    Martell, M. (2023).Economic cost of hate crimes. Bard Center for the Study of Hate, Bard College, Annandale-on-Hudson

    2023

  21. [21]

    Mazzarisi, P., Muscillo, A., Pacati, C., Pin, P. (2026). The rise and fall of ideas’ popularity.Journal of Economic Behavior and Organization,245(1), 107507

    2026

  22. [22]

    (2025).Stop Hiding Hate Act(Senate Bill S1234)

    New York State Senate. (2025).Stop Hiding Hate Act(Senate Bill S1234). Albany, NY: New York State Legislature

    2025

  23. [23]

    (2025).Investigation into social media platforms over illegal terror and hate content

    Ofcom. (2025).Investigation into social media platforms over illegal terror and hate content. London: Office of Communications

    2025

  24. [24]

    Park, T.J., Rohatagi A. (2024). Balancing the platform responsibility paradox: A case for amplification regulation to mitigate the spread of harmful but legal content online.Computer Law and Security Review,52, 105960

    2024

  25. [25]

    Peterson-Salahuddin, C. (2024). Repairing the harm: Toward an algorithmic repa- rations approach to hate speech content moderation.New Media & Society,11(2), 20539517241245333

    2024

  26. [26]

    Public Health Institute

    Public Health Institute and UCLA Center for Health Policy Research (2025).The impacts of hate in California: Victim costs and unmet needs. Public Health Institute

    2025

  27. [27]

    Pigou, A. C. (1920).The economics of welfare. London: Macmillan

    1920

  28. [28]

    Online hate: Is hate an infectious disease? Is social media a promoter?.Journal of Applied Philosophy,40(5), 788-812

    Popa-Wyatt, M.(2023). Online hate: Is hate an infectious disease? Is social media a promoter?.Journal of Applied Philosophy,40(5), 788-812. 25

    2023

  29. [29]

    Ramponi, A., & Tessitore, P. (2023). Optimal social and vaccination control in the SVIR epidemic model.Mathematics,12(7), 933

    2023

  30. [30]

    (2006).The diffusion of innovations in social networks

    Reich, B. (2006).The diffusion of innovations in social networks. London: University College London

    2006

  31. [31]

    Reynolds, C., Hallinan, B. (2024). User-generated accountability: Public participa- tion in algorithmic governance on YouTube.New Media and Society,26(9), 5107- 5129

    2024

  32. [32]

    H., Ottoni, R., West, R., Almeida, V

    Ribeiro, M. H., Ottoni, R., West, R., Almeida, V. A. F., & Meira, W. (2020). Auditing radicalization pathways on YouTube. InProceedings of the 2020 Conference on Fairness, Accountability, and Transparency (FAccT ’20)(pp. 131–141)

    2020

  33. [33]

    and Van Der Linden, S.(2020)

    Roozenbeek, J., Schneider, C.R., Dryhurst, S., Kerr, J., Freeman, A.L., Recchia, G., Van Der Bles, A.M. and Van Der Linden, S.(2020). Susceptibility to misinformation about COVID-19 around the world.Royal Society open science,7(10)

    2020

  34. [34]

    G., Naik, S

    Amballoor, R. G., Naik, S. B.(2024). SIR model for understanding the spread of fake news and hate speech. InText and Social Media Analytics for Fake News and Hate Speech Detection(pp. 166-180). Chapman and Hall/CRC

    2024

  35. [35]

    Solferino, N. (2026). Fighting Misinformation: Super-Spreaders, Learning, and Op- timal Education Policies.Working Paper SSRN. May

    2026

  36. [36]

    W., Abebaw,Y.F

    Teklu, S. W., Abebaw,Y.F. (2024). Analysis of the hate speech and racism co- existence dissemination model with optimal control strategies.Chaos, Solitons and Fractals: X,12, 100109

    2024

  37. [37]

    Weimann, G., Masri, N. (2021). TikTok’s spiral of antisemitism.Journalism and Media,2(4), 697-708

    2021

  38. [38]

    Vaidya, A., Nagar, S., Nanavati, A. A. (2024). Analyzing the spread of toxicity on Twitter: An SEIR approach.Proceedings of the 7th Joint International Confer- ence on Data Science and Management of Data (11th ACM IKDD CODS and 29th COMAD), 118-126

    2024

  39. [39]

    van den Driessche, P., & Watmough, J. (2002). Reproduction numbers and sub- threshold endemic equilibria for compartmental models of disease transmission. Mathematical Biosciences,180(1–2), 29-48

    2002

  40. [40]

    Walther, J. B. (2024). The effects of social approval signals on the production of on- line hate: A theoretical explication.Communication Research, 00936502241278944. 26

    2024