PaperlayerView exact revision on arXiv

Source-linked AI summary

Attentional DoS: Repeat Reposting, Collective Attention, and Information Diffusion on X

Genta Toya, Bruno T. Sugano, Kei Ichikawa, Qianyun Wu, Shuhei Saigusa, Yasuhiro Hashimoto, Masashi Toyoda, Naoki Yoshinaga, Kazutoshi Sasahara

arXiv:2609.35385v1cs.SI

TL;DR

Repeat reposting lets an X user put the same post back into followers’ timelines, placing repeated demand on finite collective attention. The paper measures this behavior and its organization in large cascades, finding that later audience growth is associated less with repeat count than with the estimated reach of repeating accounts.

Abstract

from arXiv · show

Collective attention is a finite shared resource, and social media posts compete for limited opportunities to be seen. On X, users can undo a repost and repost it again. By repeating this cycle, a user can put the same post back into followers' timelines any number of times without making new content. We call this procedure repeat reposting, and we read it as placing repeated demand on this shared resource (Attentional DoS). We formalize this idea and explore repeat reposting in a large-scale dataset of cascades with at least 1,000 reactions, originating from posts classified as Japanese on X, covering April 2025 to March 2026. Repeat reposts are rare, appearing in a small share of all (user, post) pairs. Even so, close to a million posts have at least one repeater, and most repeats come from a small group of habitual accounts. The central result is that subsequent audience growth is associated less with the number of repeats than with the estimated reach of the repeating accounts. Through the lens of seriality, how habitually the same groups of accounts repeat reposts across many posts, a distinct distributed form emerges: several serial amplifiers converge on the same post (Attentional DDoS). Its synchrony, how closely their actions are timed together, forms a continuum from bursts on the scale of minutes to a daily clock. A check of the content of 99.7% of amplified posts shows that most ADDoS repeat events are directed at Chinese-script content, which our vocabulary matching and sample inspection indicate is predominantly commercial spam. Repeat reposting thus lets a post re-enter the competition for visibility. The observed pattern is better characterized as repeated temporal coverage of an existing audience and the self-reinforcement of posts that are already growing, rather than as evidence that repeat reposting takes reach from other content.

Introduction

The paper treats repeat reposting as repeated demand on finite collective attention and studies it alongside the organization of its sources. It asks how common the practice is, how it relates to later cascade growth, and how repeated amplification is organized and what content it targets.

Introduction

  • Repeat reposting places repeated demand on finite collective attention, which the paper names Attentional DoS by analogy with network denial-of-service.The analogy concerns excess requests on a finite shared resource; unlike informational DoS, repeat reposting hides nothing behind the repeats.Source: Introduction · p. 1
  • The paper crosses two axes—repeat reposting as the mechanism and the organization of sources—rather than treating either alone.Prior work on coordinated amplification measured source organization while taking the repeated operation as given; DoS describes a mechanism but not source organization.Source: Single Source and Distributed Sources:… · pp. 1–2
  • Attentional DDoS denotes the distributed form identified through seriality and synchrony, distinguishing habitual, closely timed repeaters from many sources alone.Seriality captures repeated participation across posts, while synchrony measures how closely actions are timed; many sources alone do not establish organization.Source: Single Source and Distributed Sources:… · p. 2

Research Questions

  • The study is exploratory: its questions and RQ2 design were fixed while examining the full year of data, so findings are hypotheses for confirmatory studies.The RQ2 design choices included dose, controls, and stratification by repeater reach and source structure.Source: Research Questions · p. 2
  • The questions address repeat reposting’s distribution, its association with subsequent cascade growth, and the organization and content of repeated amplification.RQ2 also asks how the association varies with repeater reach and whether one or several accounts supply the repeats.Source: Research Questions · p. 2

Related Work

Related work connects repeat reposting to research on scarce attention, reversible engagement operations, coordinated amplification, and information operations. It also motivates separating behavioral patterns from judgments about the intent or authenticity behind them.

Scarce Collective Attention and the Measurement of Diffusion

  • Prior work links limited attention and network structure to uneven information spread, and examines how overloaded timelines and early-adopter communities relate to diffusion.Studies also estimate modest causal effects of trending exposure, motivating a question about whether added visibility is associated with later audience growth.Source: Scarce Collective Attention and the… · p. 2

Abuse of Reversible Operations and the Inflation of Engagement

  • Earlier studies describe reversible engagement operations such as deleting posts to evade limits, repeating likes and unlikes, and ephemeral astroturfing.These inflate apparent engagement through different operations; the paper identifies repost undoing and redoing as an unmeasured form.Source: Abuse of Reversible Operations and the… · p. 2

Coordinated Amplification and Its Detection

  • Coordinated-amplification research infers organization from source multiplicity, co-action, repeated participation, and temporal or structural behavior patterns.Some approaches estimate coordination as a degree rather than making a binary decision.Source: Coordinated Amplification and Its… · p. 3

Flooding, Information Operations, and the Problem of Intent

  • Research on information operations documents flooding and repeated amplification, but behavioral signatures alone cannot distinguish spam, promotion, political manipulation, and enthusiastic individual behavior.The paper therefore pairs behavioral classification with content checks.Source: Flooding, Information Operations, and… · p. 3

Materials and Methods

The study analyzes X repost cascades classified as Japanese and uses behavioral thresholds to define repeaters and describe source organization. It compares matched cascades and measures later growth, while treating reach and content through specified proxies and checks.

Data

  • The corpus covers April 2025–March 2026 X cascades with at least 1,000 reactions; “Japanese” follows X’s per-post language classification.The collection contains about 4.43 billion reposts; smaller cascades are not observed, and the language-detection rules cannot be independently verified.Source: Data · p. 3

Definition of Repeat Reposting and Its Measures

Measures of Source Organization: Seriality and Synchrony

  • Seriality measures repeat participation across targets, while synchrony measures excess temporal co-occurrence of account pairs relative to independent activity.The default serial-amplifier threshold is 100 repeated targets; synchrony uses a 60-second window as one minute-scale instance and is kept continuous.Source: Measures of Source Organization:… · p. 4

Outcome Measures

  • Outcomes include direct reaction counts and unique reactors, plus cascade growth measured by arrivals and unique-reactor size through seven days.Reaction counts include repeats and are an activity measure, while unique reactors proxy engaged reach rather than views or impressions; about 0.1% of cascades receive reactions after day seven.Source: Outcome Measures · p. 4

Analysis Procedure

  • Treated cascades with repeats are paired to unrelated controls matched on timing, size, and arrival-curve shape; growth differences are estimated with cascade-cluster bootstrap intervals.The repeating account’s prior arrival volume after single reposts serves as a reach proxy, and supplementary controls match on its reach stratum.Source: Analysis Procedure · p. 4
  • The analysis also compares post-window growth speed, characterizes synchrony’s daily-clock scale, and checks amplified-post text against vocabularies and scripts.Root text was obtained for 892,506 posts (99.71% coverage); daily-clock accounts are identified from regular repeat intervals near 24 hours.Source: Analysis Procedure · pp. 4–5

Results

Repeat reposts were uncommon but concentrated among habitual accounts, and subsequent growth aligned more with repeater reach than repeat count. Serial amplifiers formed a distributed pattern whose repeat events were predominantly directed at Chinese-script content identified as commercial spam, while political repeats also appeared in other forms.

RQ1: Scale and Distribution of Repeat Reposting

  • 0.20% of user–target pairs had repeat reposts, yet 895,086 targets had a repeater; 8,685 habitual accounts (2.43%) held 88.0% of repeat pairs.Repeat intervals had a median of about 9 hours; only about 6% were under 60 seconds.Source: RQ1: Scale and Distribution of Repeat… · p. 5

RQ3: Organized Repeat Reposting

  • ADDoS requires at least three serial amplifiers and s ≥0.5; this 9.5% of targets accounts for 83.3% of repeat events.Seriality distinguishes habitual convergence from multi-source activity alone; synchrony remains descriptive.Source: RQ3: Organized Repeat Reposting · p. 6

The Time Scale of Synchrony and Its Shift

  • Synchrony had no clear threshold: 55.5% of ADDoS targets had σ=0, while 17.8% of c ≥5 repeat pairs were clock-like.The minute-scale share fell from 19.2% in April 2025 to 3.1–4.8% from October onward as median intervals moved toward 24 hours; the proposed infrastructure shift remains a hypothesis.Source: The Time Scale of Synchrony and Its… · pp. 6–7

RQ2: the number of repeats and subsequent growth

  • Matched-cascade growth aligned more with main-repeater reach than repeat dose: reach elasticity was positive, while unique-reactor dose elasticity was negative.When matched to one-time reposts by equal-reach accounts, the overall 7-day size difference stayed near zero.Source: RQ2: the number of repeats and… · pp. 5–6, Table 2 · p. 7
  • Low-reach repeaters fell below controls after the window, while high-reach multi-source cascades rose above them; these comparisons are descriptive, not causal.The gap for high-reach multi-source cascades widened to +0.6 to +1.15 after 24 hours.Source: RQ2: the number of repeats and… · p. 6, Figure 2 · p. 7

Commercial Spam and Political Repeats

  • 98.2% of ADDoS repeat events with retrievable text targeted Chinese-character-only posts, predominantly commercial spam by vocabulary matching and sample inspection.The authors caution that vocabulary matching may undercount political content and that the spam characterization of unmatched posts rests on sample inspection.Source: Commercial Spam and Political Repeats · pp. 7, 9
  • Political repeats appeared mainly in single-source ADoS, while Japanese political posts also occurred in light ADDoS; multi-source non-serial repeats included giveaways.Most political targets were single-source (80.9%); the authors do not infer an absence of political coordinated inauthentic behavior.Source: Commercial Spam and Political Repeats · p. 9

Discussion

The observed growth association tracks the repeating accounts’ estimated reach more than repeat counts, while seriality distinguishes distributed repeat amplification and synchrony varies continuously. The paper cautions that these observational findings do not establish causality or account intent.

Amplification Is Not Diffusion

  • Subsequent growth tracked repeaters’ estimated reach more than repeat counts: low-reach repeats grew more slowly than matched controls, while high-reach or multiple-account repeats showed more sustained growth.With reach held approximately equal, repeated and single-repost cascades differed near zero; the data do not show whether added readers were the repeaters’ followers.Source: Amplification Is Not Diffusion · p. 9

Implications for Coordination Detection

  • Coordination detection should separate seriality from continuous synchrony, report synchrony’s time scale, and account for commercial spam alongside or instead of political operations.The paper also notes that political repeats exist at large scale, often in single-source form, which between-account co-occurrence methods cannot detect by construction.Source: Implications for Coordination Detection · pp. 9–10

Limitations

  • The study is exploratory, observational causality is unidentified, small cascades are outside the data, and repeat signatures do not establish account intent or authenticity.The data also cover one platform and language sphere; political single-source repeats cannot be distinguished from an enthusiastic individual using these data.Source: Limitations · p. 10

Conclusion

  • Repeat reposting returns the same post to the visibility competition without new content; the observed pattern is repeated exposure within a bounded audience, not demonstrated loss of reach by other content.Self-reinforcement among posts already growing remains inseparable from selection.Source: Amplification Is Not Diffusion · p. 9, Conclusion · pp. 10–11
  • Across twelve months, the corpus contained 895,086 repeat instances, ADDoS held 83% of repeat events, and synchrony shifted from minutes to a day.The conclusion presents these exploratory findings as hypotheses for confirmatory studies.Source: Conclusion · p. 11

A Thresholds and Sensitivity

The appendix tests how repeat-reposting classifications change across thresholds for repeaters, serial amplifiers, and seriality.

A Thresholds and Sensitivity

  • With c ≥2/3/4/5, ADoS instances were 2,899,085/895,086/462,634/295,768 and ADDoS counts were 452,974/84,804/40,259/25,388.The ADDoS share changes little as the repeater floor rises; c ≥3 is used to exclude mistakes.Source: A Thresholds and Sensitivity · p. 12
  • With T = 20/50/100/200/500, ADDoS counts were 101,807/92,613/84,804/74,864/62,490; the T = 100 floor is conventional because account repetition has no natural break.The account-level distribution is power-law-like, with exponent about 1.55, 95th percentile 19, 99th percentile 574, and maximum 14,600.Source: A Thresholds and Sensitivity · p. 12
  • With k = 2/3/5/10 serial amplifiers, ADDoS counts were 148,158/84,804/55,615/41,965; k = 2 adds 63,354 targets with exactly two serial amplifiers.At k = 2, 59% of the extracted set are light targets with 2 to 4 repeaters.Source: A Thresholds and Sensitivity · p. 12
  • With seriality cuts of 0.3/0.5/0.7/0.85, ADDoS counts were 86,939/84,804/78,821/69,468; counts were monotone across all 144 threshold-grid points.The 0.5 cut sits just above the 0.37 to 0.44 valley in the near-continuous band of targets with 20 or more repeaters.Source: A Thresholds and Sensitivity · p. 12

B Details of the Matching

The matching procedure selects comparable cascades by agreement on window characteristics and proximity in their arrival curves.

B Details of the Matching

  • Among cascades agreeing on window size and shape, matching uses the smallest Mahalanobis distance on cumulative arrivals and time since last arrival, with a 0.25 SD caliper.Matching is one-to-one without replacement; standardized differences for matching variables are at or below 0.05.Source: B Details of the Matching · p. 12

C Time Signature

ADDoS synchrony spans a continuum rather than distinct timing categories, and its observed time signature shifts toward a daily clock. Account sets at the minute-scale and clock ends overlap substantially.

C Time Signature

  • After controlling for unique reactors, the time-to-peak regression reports a unique-reactor coefficient of +0.047 and form coefficients of +0.45.The passage identifies the result as evidence that slowness is not a by-product of size; it does not specify the form coefficients' comparison in the supplied text.Source: C Time Signature · p. 12
  • For ADDoS synchrony (n = 84,804), more than half of values are zero, and the positive part is continuous with no break; every cutoff therefore divides a continuum.Source: Figure 4 · p. 13
  • Among ADDoS targets, the minute-scale-dominated share collapses from 19.2% to 3 to 5%, while median inter-repeat intervals of pairs approach 24 hours (n = 84,804).83.9% of accounts repeating at the minute-scale end also repeat at the clock end.Source: Figure 4 · p. 13
  • Figure 5 reports inter-repeat-interval spikes at 24, 48, and 72 hours; 18% of pairs are clockwork, while 68% drift from a fixed-time habit.Clock accounts span targets numbering up to 8,787.Source: Figure 5 · p. 13

Related papers

Loading 2609.35385v1…