A Study of Cyber Hate on Twitter with Implications for Social Media Governance Strategies

Adam Edwards; Helena Webb; Marina Jirotka; Matt Williams; Pete Burnap; Rob Procter; William Housley

REVIEW 3 major objections 6 minor 1 cited by

A Study of Cyber Hate on Twitter with Implications for Social Media Governance Strategies

T0 review · 3 major / 6 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read When more distinct Twitter users post counter-speech against hate, the hate threads end sooner, offering a measurable lever for social media self-governance.

desk verdict The headline finding about unique counter-speech contributors shortening threads is a compositional artifact, not a marginal effect; the paper is still worth reviewing for its dataset and honest reporting. read the letter →

arxiv 1908.11732 v1 pith:YKA762XW submitted 2019-08-30 cs.CY cs.SI

Rob Procter , Helena Webb , Marina Jirotka , Pete Burnap , William Housley , Adam Edwards , Matt Williams This is my paper · ORCID

classification cs.CYcs.SI

keywords cyberhatecounter-speechTwitterthreadlengthself-governancemachinelearningclassificationsocialmediagovernanceonlinespeech

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 self-governance of social media through counter-speech works best when many distinct people contribute the counter-speech. Studying 300 Twitter threads triggered by hateful posts, the authors model thread length as a proxy for a hate thread's potential for harm. They find that the raw number of counter-speech posts lengthens a thread, but the number of unique counter-speech contributors shortens it, with significant negative coefficients for sexist, racist, and homophobic threads. The authors interpret this as evidence that diffuse, many-voiced challenge ends hate threads sooner, and they argue this has concrete implications for how sentinel accounts and platforms should encourage counter-speech. A second contribution is a machine classifier intended to detect counter-speech in real time, though the paper acknowledges its accuracy on support and counter-speech classes is currently weak.

What carries the argument

The key mechanism is a linear regression model of thread length—the number of posts in a reply-linked Twitter thread—regressed on counts of hateful posts, supportive posts, disagreeing posts, insults, unique contributors, original-poster contributions, unique hateful contributors, and unique counter-speech contributors. The load-bearing term is the unique-counter-speech-contributor count, whose negative and significant coefficient across all three bias strands carries the paper's argument that diffuse, many-voiced counter-speech curtails hate threads.

What would settle it

Collect a random sample of cyber hate threads that do not originate from sentinel accounts, fit the paper's regression with the same covariates, and test whether the coefficient on unique counter-speech contributors is still negative and statistically significant for sexist, racist, and homophobic threads; if it is not, the central claim fails to generalize.

Watch

Extended reading notes

Core claim

The central claim is that the number of unique individuals who contribute counter-speech to a Twitter thread is negatively associated with the length of that thread. In a linear regression of thread length on response-type counts, the coefficient for unique counter-speech contributors is statistically significant and negative for sexist ($-7.51$), racist ($-2.31$), and homophobic ($-1.42$) threads, while the raw count of counter-speech posts is positively associated with thread length. The authors interpret this as mass self-governance: when many different people join in to challenge hate, the thread ends sooner; when a small number of people volley counter-speech back and forth, the thread grows.

Load-bearing premise

The dataset is drawn from three sentinel accounts that exist specifically to publicize hateful posts and provoke counter-speech, so the relationship between the number of unique counter-speech contributors and thread length may be a property of those accounts' assembled audiences rather than a general feature of cyber hate threads on Twitter.

Editorial extensions

If this is right

Platforms and monitoring organizations can use a rising count of distinct counter-speech contributors as a leading indicator that a hate thread is nearing its end.
Sentinel accounts will get more curtailment by recruiting many different followers to respond than by concentrating on a few highly active respondents.
Thread length becomes a usable outcome metric for evaluating the real-world impact of counter-speech campaigns, since it is responsive to the structure of participation.
A reliable real-time classifier for support and counter-speech, once improved, would let moderators direct human review to threads where counter-speech is not yet diffuse.

Reading between the lines

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

The many-voices effect may generalize to other platforms: in any forum where a hostile post draws responses from a large, diverse set of users, the social pressure on the original poster may grow and the interaction may terminate sooner.
The regression design leaves open a selection confound: threads that attract many distinct counter-speakers may already be widely condemned, so the shorter length could reflect the audience's prior disposition rather than the counter-speech itself.
A direct test would compare reply-thread lengths when the same hateful content is posted from an anonymous account versus a public persona, or when the number of visible responses is artificially capped.
The paper's use of thread length as the harm proxy assumes shorter threads are less harmful; if a short thread simply reduces scrutiny, a hateful post could evade detection, so harm measurement should be validated against follow-on behaviors.

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.

Desk Editor's Note

read the letter

Here's the one thing to know: the paper's headline result—that more unique counter-speech contributors shorten cyber hate threads—is not actually supported by the regression. The negative coefficient on uniqCScontributors only says that among threads with the same total unique participants and the same counts of disagreeing/insulting posts, a larger share of one-off counter-speech accounts is associated with shorter threads. That is a compositional effect. Adding a new counter-speech user would also increase the total unique contributor count and the disagree count, and with the sexist strand coefficients those increases roughly cancel the negative coefficient (2.875 - 7.508 + 4.638 ≈ 0). So the conclusion that 'engagement in counter-speech by more unique individual posters reduces the thread length' overstates what the model identifies.

What is genuinely useful: the paper is honest and transparent. The mixed-methods design—qualitative conversation analysis paired with a coded dataset of 300 threads and a regression—is a serious way to explore the space. The ML classifier experiments are a routine extension of Burnap and Williams, but the authors clearly flag the poor performance on the counter-speech class, which is refreshing. The annotation scheme and the dataset are contributions others can build on.

Soft spots, in order: (1) the compositional issue is the big one; the practical policy claim about mobilizing many distinct responders is not identified. (2) Section 4.1 contains a direct contradiction: the text says unique hateful contributors was omitted for sexist/racist threads and not significant for homophobic, then a few sentences later says it was 'statistically significant across all strands'. That needs fixing. (3) The sentinel-account sampling means threads come from accounts that deliberately provoke counter-speech; results may not generalize to organic threads. The authors acknowledge the source but don't discuss this limitation. (4) Thread length as a proxy for harm is asserted, not defended. (5) Minor: no inter-annotator agreement metric reported, just a 75% removal rule.

Who is this for? Researchers on counter-speech, online hate, and platform governance. The ML people will find the classifier results modest but the dataset valuable; the social scientists will appreciate the qualitative grounding. It deserves peer review—not desk rejection—but it needs major revision: reframe the regression result as a conditional association, resolve the contradiction, and add a clear limitations section. If the authors do that, the paper becomes a useful empirical contribution rather than a misleading policy hook.

Referee Report

3 major / 6 minor

Summary. The paper investigates whether counter-speech on Twitter can shorten cyber-hate threads, treating thread length as a proxy for harmful impact. The authors collected 300 threads from three sentinel accounts (@Homophobes, @YesYoureRacist, @YesYoureSexist), annotated the replies with a six-category scheme, and fitted a linear regression of thread length on counts of hateful posts, supporting posts, disagreeing posts, insults, unique contributors, original-poster contributions, unique hateful contributors, and unique counter-speech contributors. They report that the number of unique counter-speech contributors has a statistically significant negative coefficient across sexist, racist, and homophobic threads, and interpret this as evidence that mass self-governance through counter-speech curtails thread length. They also train SVM classifiers to detect counter-speech, reporting reasonable overall F-scores but poor performance on the counter-speech class in the confusion matrices.

Significance. The paper's strength is that it formulates a falsifiable quantitative claim, tests it on three protected-characteristic strands using manually annotated data, and transparently reports the limitations of the machine classifier. If the negative coefficient on unique counter-speech contributors were a genuine effect of mobilizing additional counter-speech posters, the finding would be practically important for social media governance. However, as detailed in the major comments, the reported regression does not identify the marginal effect asserted in the conclusions, and the sentinel-account sampling frame restricts generalizability. The paper is therefore a useful descriptive study of interaction patterns within sentinel-account threads, but its headline governance recommendation requires re-analysis and re-framing.

major comments (3)

[§4.1/Table 1 and §6] The central interpretation of Table 1 is not identified by the reported model. Because uniqCScontributors is a subset of uniqcontributors, and every counter-speech post is also counted in one of the post-type variables (e.g., disagree), the regression cannot estimate the effect of adding one new unique counter-speech poster while holding the other predictors fixed. A one-unit increase in uniqCScontributors must also increase uniqcontributors and at least one post-count variable. Using the sexist point estimates, the net predicted change in thread length from one additional unique counter-speech contributor whose post is coded as 'disagree' is 2.874905 - 7.508296 + 4.638163 ≈ 0.005, not -7.51; for racist and homophobic strands the analogous net changes are positive (≈0.99 and ≈0.65). Thus the negative coefficient is a conditional composition effect (among threads with equal total unique participants and equal counts, a larger share of one-off CS accounts is associated with shorter threads), not the marginal effect of mobilizing additional unique counter-speech users that the Discussion and Conclusions describe. The authors should either estimate a model that explicitly separates composition from mobilization (for example, by including the share of unique contributors who are counter-speech-only) or clearly restate the claim as a conditional association.
[§4.1/Table 1] The text states that 'the number of unique hateful contributors ... was statistically significant across all strands and positively correlated with thread length.' This is contradicted by Table 1, where uniqhatefulcontributors is omitted for sexist and racist threads and has a non-significant negative coefficient (-1.152764) for homophobic threads. Please correct the text or the table; as written, the paragraph appears to confuse uniqcontributors with uniqhatefulcontributors.
[§3.1] The sample consists of the first 100 tweets from each of three sentinel accounts that explicitly seek out hateful posts and provoke counter-speech. The observed negative association between unique counter-speech contributors and thread length may reflect the audience and posting practices of these accounts rather than a general property of cyber-hate threads on Twitter. The paper should either re-analyze the claim on a broader or random sample of cyber-hate threads, or substantially soften the governance implications, which currently generalize beyond the sampling frame. At a minimum, the authors should report the date range, follower counts, and the exact selection procedure for the 'first 100' tweets.

minor comments (6)

[§4.1] The sentence 'support for the original hateful remark is only significant for racism' is inconsistent with the following claim that 'neither higher volume of cyber hate, nor increased support for cyber hate, influence the thread length'; the racist support coefficient is significant (p<0.01) and negative, so the text should state that support is significant for one strand and associated with shorter threads.
[§3.2] The annotation quality description is unclear: 'removed all tweets with less than 75 percent agreement and also those upon which the annotators could reach an absolute decision (i.e., the undecided class)' appears to mean 'could not reach an absolute decision.' Please report how many tweets were removed per strand and whether removal differed systematically by class, since this affects both the regression and the classifier inputs.
[§4.2/Table 5] For the homophobic strand, the classifier assigns 0 of 20 counter-speech (class 2) posts correctly; the Discussion acknowledges this, but the Conclusions claim that the classifier 'will enable the closer study of self-governance' and provide 'real-time input into the statistical model' should be explicitly conditioned on this poor class-level performance.
[§6] The sentence 'has a a role to play' contains a duplicated article; please proofread.
[§4.2] The phrase 'improves classification over the baseline Bag of Words approach for two of the tree classes' should read 'three classes' or 'two of the three classes' depending on the intended claim.
[Table 1] The variable label 'uniqhatefulcontributors0' appears to contain a placeholder '0'; the table notes should also report sample sizes (N=100 per strand) and the number of threads per strand.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the regression analysis estimates associations from annotated data rather than deriving its conclusions from its own inputs.

full rationale

The paper's central quantitative claim is an empirical regression result: thread length is regressed on counts of hateful posts, supportive posts, disagreeing posts, insults, unique contributors, original poster contributions, unique hateful contributors, and unique counter-speech contributors. The dependent variable (thread length) is defined as the number of posts in a thread including the original post, and the independent variables are separate counts annotated from the same threads; thread length is not defined in terms of the predictors, and no parameter is fitted to a subset of data and then used to predict the same subset. The negative coefficient on unique counter-speech contributors is an estimated association, not a quantity forced by construction. The paper's self-citations appear as background methodology or prior classifier work, and the load-bearing regression result does not reduce to those citations. A reviewer could raise a compositional or identifiability concern because every unique counter-speech contributor must also contribute a post, so the marginal interpretation of the coefficient is contestable, but that is a statistical inference issue, not circularity: the model does not define uniqCScontributors in terms of thread length or rename an input as a prediction. The machine classifier section similarly reports measured precision, recall, and F-measure against held-out annotations, not a self-fulfilling derivation. Therefore no circular step meeting the required evidentiary standard is present.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The core statistical result depends on several domain assumptions: the thread-length proxy, the representativeness of sentinel-account threads, the validity of annotations, and the choice of linear regression for count data. No new entities are postulated. The SVM hyperparameters are minor tuned parameters in the secondary classifier.

free parameters (2)

SVM gamma = 0.1
Chosen through experimentation in Section 3.4; affects classifier performance but not the thread-length model.
SVM C = 1.0
Chosen through experimentation in Section 3.4; regularisation parameter for the SVM classifier.

assumptions (4)

domain assumption Thread length is a proxy for potential harmful impact of a cyber hate thread.
Stated in Section 1; the interpretation that shorter threads mean less harm depends on this proxy.
domain assumption Sentinel accounts provide a suitable and representative source of cyber hate threads.
Section 3.1 justifies selecting @Homophobes, @YesYoureRacist, and @YesYoureSexist because they provoke counter-speech; this may not generalize to all cyber hate threads.
domain assumption Annotation validity can be established by removing tweets with less than 75 percent agreement instead of reporting an inter-rater reliability statistic.
Section 3.2; no kappa or alpha is reported, so label reliability is unquantified.
domain assumption Linear regression is an appropriate model for thread length as a count variable.
Section 3.3 uses ordinary linear regression on a count outcome without justification, for example a Poisson or negative binomial alternative.

how reviews work

0 comments

Cite this review

Pith. "Pith review of A Study of Cyber Hate on Twitter with Implications for Social Media Governance Strategies." pith.science (2026). https://pith.science/paper/YKA762XW

@misc{pith2026190811732,
  author       = {Pith},
  title        = {Pith review of: A Study of Cyber Hate on Twitter with Implications for Social Media Governance Strategies},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YKA762XW}},
  note         = {Machine review of arXiv:1908.11732}
}

read the original abstract

This paper explores ways in which the harmful effects of cyber hate may be mitigated through mechanisms for enhancing the self governance of new digital spaces. We report findings from a mixed methods study of responses to cyber hate posts, which aimed to: (i) understand how people interact in this context by undertaking qualitative interaction analysis and developing a statistical model to explain the volume of responses to cyber hate posted to Twitter, and (ii) explore use of machine learning techniques to assist in identifying cyber hate counter-speech.

Figures

Figures reproduced from arXiv: 1908.11732 by the authors.

**Figure 3.** counter-speech tweet example 1. Tweet 3 The first example of a counter-speech tweet ( [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

**Figure 2.** Provocative tweet example 2. This brief examination demonstrates that tweets that might be considered antagonistic or hateful do not necessarily contain individual words that are offensive. Instead, inflammatory opinions are expressed by constructing negative references and drawing on emotive and/or divisive contexts. These observations help us towards a more nuanced understanding of the expression of cyber ha… view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: counter-speech tweet example 2. Tweet 4 The second example ( [PITH_FULL_IMAGE:figures/full_fig_p003_4.png]

**Figure 5.** Figure 5: Annotation scheme for cyber hate threads. [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Echoes of Discord: Forecasting Hater Reactions to Counterspeech
cs.CL 2025-01 conditional novelty 6.0 of 10

A three-way classifier trained on Reddit hate speech/counterspeech pairs predicts hater reentry and reentry type more accurately than a two-stage predictor, with linguistic features of counterspeech signaling differen...

Reference graph

Works this paper leans on

24 extracted references · 21 canonical work pages · cited by 1 Pith paper

[1]

URL: " 'urlintro :=

ENTRY address author booktitle chapter edition editor howpublished institution journal key month note number organization pages publisher school series title type volume year eprint doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRINGS urlintro eprinturl eprintpr...
[2]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...
[3]

write newline

" write newline "" before.all 'output.state := FUNCTION fin.entry add.period write newline FUNCTION new.block output.state before.all = 'skip after.block 'output.state := if FUNCTION new.sentence output.state after.block = 'skip output.state before.all = 'skip after.sentence 'output.state := if if FUNCTION not #0 #1 if FUNCTION and 'skip pop #0 if FUNCTIO...

work page
[4]

Imran Awan. 2014. Islamophobia and twitter: A typology of online hate against muslims on social media. Policy & Internet, 6(2):133--150

work page 2014
[5]

Jamie Bartlett and Alex Krasodomski-Jones. 2015. Counter-speech examining content that challenges extremism online. DEMOS, October

work page 2015
[6]

Susan Benesch, Derek Ruths, Kelly Dillon, Haji Saleem, and Lucas Wright. 2016. Counterspeech on twitter: A field study. dangerous speech project

work page 2016
[7]

Pete Burnap and Matthew L Williams. 2016. Us and them: identifying cyber hate on twitter across multiple protected characteristics. EPJ Data Science, 5(1):1

work page 2016
[8]

Robert Faris, Amar Ashar, Urs Gasser, and Daisy Joo. 2016. Understanding harmful speech online. Berkman Klein Center Research Publication, (2016-21)

work page 2016

Show all 24 references

[9]

Iginio Gagliardone, Danit Gal, Thiago Alves, and Gabriela Martinez. 2015. Countering online hate speech. UNESCO Publishing

2015
[10]

Stephen Hester and Peter Eglin. 1997. Culture in action: Studies in membership categorization analysis. Number 4 in Studies in Ethnomethodology and Conversation Analysis. University Press of America

1997
[11]

William Housley, Helena Webb, Adam Edwards, Rob Procter, and Marina Jirotka. 2017 a . Digitizing S acks? approaching social media as data. Qualitative Research

2017
[12]

William Housley, Helena Webb, Adam Edwards, Rob Procter, and Marina Jirotka. 2017 b . Membership categorisation and antagonistic twitter formulations. Discourse & Communication

2017
[13]

Marie-Catherine de Marneffe, Bill MacCartney, and Christopher D. Manning. 2006. http://nlp.stanford.edu/pubs/LREC06_dependencies.pdf Generating typed dependency parses from phrase structure trees . In LREC

2006
[14]

Corien Prins. 2011. Digital tools: Risks and opportunities for victims: Explorations in e-victimology. In The New Faces of Victimhood, pages 215--230. Springer

2011
[15]

Harvey Sacks, Emanuel A Schegloff, and Gail Jefferson. 1974. A simplest systematics for the organization of turn-taking for conversation. language, pages 696--735

1974
[16]

Carla Schieb and Mike Preuss. 2016. Governing hate speech by means of counterspeech on facebook. In 66th ica annual conference, at fukuoka, japan, pages 1--23

2016
[17]

Mike Thelwall, David Wilkinson, and Sukhvinder Uppal. 2010. https://doi.org/10.1002/asi.21180 Data mining emotion in social network communication: Gender differences in myspace . Journal of the American Society for Information Science and Technology, 61(1):190--199

2010 doi
[18]

Gavan Titley, Ellie Keen, and L \'a szl \'o F \"o ldi. 2014. Starting points for combating hate speech online. Council of Europe, October 2014

2014
[19]

Peter Tolmie, Rob Procter, Mark Rouncefield, Maria Liakata, and Arkaitz Zubiaga. 2018. Microblog analysis as a program of work. ACM Transactions on Social Computing, 1(1):2

2018
[20]

Hanna M Wallach. 2006. Topic modeling: beyond bag-of-words. In Proceedings of the 23rd international conference on Machine learning, pages 977--984. ACM

2006
[21]

Helena Webb, Pete Burnap, Rob Procter, Omer Rana, Bernd Carsten Stahl, Matthew Williams, William Housley, Adam Edwards, and Marina Jirotka. 2016. D igital W ildfires: Propagation, verification, regulation, and responsible innovation. ACM Transactions on Information Systems (TO...

2016
[22]

Helena Webb, Marina Jirotka, Bernd Stahl, William Housley, Rob Procter, Adam Edwards, Matt Williams, Omer Rana, and Pete Burnap. 2017. The ethical challenges of publishing twitter data for research dissemination. In ACM Web Science. ACM Press

2017
[23]

Helena Webb, Marina Jirotka, Bernd Carsten Stahl, William Housley, Adam Edwards, Matthew Williams, Rob Procter, Omer Rana, and Pete Burnap. 2015. ' D igital W ildfires': a challenge to the governance of social media? In Proceedings of the ACM web science conference, page 64. ACM

2015
[24]

Lucas Wright, Derek Ruths, Kelly P Dillon, Haji Mohammad Saleem, and Susan Benesch. 2017. Vectors for counterspeech on twitter. In Proceedings of the First Workshop on Abusive Language Online, pages 57--62

2017

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [1]

URL: " 'urlintro :=

ENTRY address author booktitle chapter edition editor howpublished institution journal key month note number organization pages publisher school series title type volume year eprint doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRINGS urlintro eprinturl eprintpr...

[2] [2]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

[3] [3]

write newline

" write newline "" before.all 'output.state := FUNCTION fin.entry add.period write newline FUNCTION new.block output.state before.all = 'skip after.block 'output.state := if FUNCTION new.sentence output.state after.block = 'skip output.state before.all = 'skip after.sentence 'output.state := if if FUNCTION not #0 #1 if FUNCTION and 'skip pop #0 if FUNCTIO...

work page

[4] [4]

Imran Awan. 2014. Islamophobia and twitter: A typology of online hate against muslims on social media. Policy & Internet, 6(2):133--150

work page 2014

[5] [5]

Jamie Bartlett and Alex Krasodomski-Jones. 2015. Counter-speech examining content that challenges extremism online. DEMOS, October

work page 2015

[6] [6]

Susan Benesch, Derek Ruths, Kelly Dillon, Haji Saleem, and Lucas Wright. 2016. Counterspeech on twitter: A field study. dangerous speech project

work page 2016

[7] [7]

Pete Burnap and Matthew L Williams. 2016. Us and them: identifying cyber hate on twitter across multiple protected characteristics. EPJ Data Science, 5(1):1

work page 2016

[8] [8]

Robert Faris, Amar Ashar, Urs Gasser, and Daisy Joo. 2016. Understanding harmful speech online. Berkman Klein Center Research Publication, (2016-21)

work page 2016