Crude, Commercial, and Self-Referential: Chinese-Language Coordinated Activity in Japanese-Language X

1. Introduction

• Prior research has focused on detecting bots and creating labeled datasets, while the reach of coordinated accounts and factors associated with their activity remain poorly understood.The authors argue that understanding which accounts and posts are influential could help prioritize detection and mitigation.Source: §1 · pp. 1–2
• Few studies examine how coordination influences user behavior, and event-centered political studies may miss activity that is non-political or changes over longer periods.The introduction notes that one study of a Russian campaign found no effect on political attitudes [Bail et al., 2020].Source: §1 · p. 2
• The study examines crude coordination—repeated identical or near-identical text across accounts—and asks what coordinated accounts post, whether activity surges align with political events, and whether their posts elicit external reactions.The dataset contains 734,173 Chinese-language coordinated accounts and 495 million coordinated posts from May 2024 to March 2026.Source: §1 · p. 2

2.1. Detecting Coordinated Behavior.

• Repeated wording can coordinate dissemination without shared URLs, a distinctive activity type, or synchronized timing, so the paper treats textual identity as an additional shared trace.This complements schemes focused on shared objects, behavioral patterns, and timing; prior work describes repeated-wording coordination [Pacheco et al. 2020].Source: §2.1 · pp. 2–3

2.2. Coordinated Activity in the Chinese-language space.

• The study targets predominantly Chinese-language accounts surfacing in X’s Japanese stream, where prior event-focused reports leave the broader activity and its external reactions unclear.Reports describe some accounts becoming active in Japanese around political events despite usually posting in Chinese, and suggest their direct influence has been considered limited.Source: §2.2 · p. 3

2.3. Research Questions.

• The research questions ask how coordinated accounts behave and what they post during surges, whether surges align with political events, and which content draws reactions from outside the coordinated population.The Chinese-language space is presented as a useful setting for examining crude coordination and its reach beyond its own population.Source: §2.3 · p. 3

3.1. Data Collection.

• The archive covers X posts classified as Japanese from May 2024 through March 2026, and includes full cascade reactions for roots with at least 1,000 aggregate reactions.Reactive posts are linked to antecedent posts to form cascade trees; reactions within included cascades are retained regardless of their own counts.Source: §3.1 · pp. 3–4

3.2. Selecting Accounts That Post Predominantly in Chinese.

• Accounts with a kana share at or below 0.2 and at least 13 body posts were treated as predominantly Chinese-language accounts.The kana ratio compares kana characters with the combined kana and Han-character count; 13 posts reduced misclassification to one in the reported check.Source: §3.2 · p. 4, §3.3 · p. 4

3.4. Collecting Coordinated Texts.

• Coordination was identified by normalizing posts to capture superficial template differences, then requiring a wording to recur across accounts and repeatedly within an account.Texts had to exceed three on both distinct-account count and maximum repetitions by one account; qualifying texts made up 96.02% of Chinese-language body posts.Source: §3.4 · p. 4, §3.5 · p. 4

3.5. Identifying Coordinated Texts and Accounts.

• An account was classified as coordinated when at least half its body posts were coordinated texts; the threshold falls in a bimodal distribution valley and yielded 734,173 accounts.Using the same procedure for one month, the coordinated-account share was 10.7% for Japanese-language accounts versus 63.0% for Chinese-language accounts.Source: §3.5 · pp. 4–5

3.6. Identifying Suspended Coordinated Accounts.

• X suspensions externally corroborated the account identification, while amplifier and reactor accounts were separately identified using chance-adjusted reaction tests and majority-action requirements.Of identified coordinated accounts, 96.79% were suspended; the authors caution that suspension is not ground truth because X’s reasons are not public. Amplifiers repost, while reactors reply, mention, or quote coordinated accounts.Source: §3.6 · p. 5, §3.7 · p. 5

3.8.1. Embedding.

• Distinct normalized texts from the full Chinese-language population were embedded and clustered to survey topics, but topic labels were propagated to peripheral texts without a similarity threshold.Because topic peripheries are attached by proximity alone, topics were used to build the codebook and survey text types, while composition was estimated at the wording level.Source: §3.8.1 · p. 5, §3.8.2 · pp. 5–6

3.8.7. Codebook Refinement.

• The LLM induced categories from topic summaries, then assigned individual wordings to a frozen codebook; refinement raised topic-level inter-run agreement from κ = 0.528 to κ = 0.660.The codebook used observable textual criteria, explicit inclusion and exclusion rules, and a fixed priority order for overlapping categories; category induction and assignment both used Claude Haiku 4.5.Source: §3.8.3 · p. 6, §3.8.4 · p. 6, §3.8.5 · p. 6

3.8.8. Units of Analysis.

• Composition estimates were weighted by post volume and based on 5,000 high-volume wordings plus a random sample of 1,000 tail wordings, with each wording coded in three runs.Two researchers independently coded a stratified subsample; agreement was κ = 0.543 between coders and κ = 0.421 between the agreed human label and LLM label. Replacing LLM labels with human labels shifted political share from 0.038% to 0.035%, without changing category rank order.Source: §3.8.8 · p. 7, §3.8.9 · p. 7, §3.8.10 · pp. 7–8

4.1. Characteristics of Coordinated Accounts.

• Median account lifespan was 0.06 months (roughly 1.8 days), versus 10.18 months for Chinese general users; median account age was 237 versus 824 days.Coordinated accounts also concentrated posting hours more than the comparison groups, though the difference was under one bit; account creation gaps had a median of 8.9 seconds. Their mentions were narrow: the median account with at least 1,000 mentions reached 9 distinct recipients, versus 428 for Chinese general users.Source: §4.1 · p. 8
• Political claims made up just 0.038% of coordinated posts; content was concentrated in unrecoverable strings, everyday topics, money solicitation, and praise or good wishes.The four largest categories accounted for 29.88%, 23.57%, 22.62%, and 18.26%, respectively.Source: §4.1 · p. 9, Figure 2 · p. 9

4.2. Surges in Coordinated Activity.

• In November 2024, political claims were 0.60% of surge-window posts, and coordinated text posts contained attack keywords far less often than Chinese general users’ posts.The attack-keyword shares were 0.001% versus 0.316%; the broad-list shares were 0.074% versus 0.475%. Posts containing payout-scheme words were 2.379% versus 0.330%, but the analysis does not establish what caused the surge.Source: §4.2 · pp. 10–11, Figure 3 · p. 10
• During the May–July 2025 surge, coordinated posting rose from 85,000 on May 24 to 953,000 on May 25, while Chinese general users remained at baseline levels.The surge’s content was consistent with a money-solicitation family rather than political claims: “opportunity” texts accounted for 37.2% of the window, and political claims for 0.2%; the authors could not determine what drove the surge.Source: Figure 4 · p. 10, §4.2 · p. 11

4.3. Cascade Analysis.

• 99.9% of reactions in large coordinated-root cascades stayed within two levels, and coordinated-root cascades were shallower than Chinese-general-user-root cascades.Mean maximum depth was 1.494 versus 4.034; 64.5% versus 32.1% ended at depth one. Within these observable cascades, coordinated posts primarily circulated within the coordinated population.Source: §4.3 · p. 12, Figure 5 · p. 12
• Illegal-material roots received 40.7% of reactions from outside the Chinese-speaking population, a comparatively high external share; explicit vocabulary was more closely associated with external reach than the category label.Within that category, roots with terms denoting minors had 44.1% external reach and those with adult sexual terms had 41.2%; adult sexual terms raised reach in sexual solicitation from 10.7% to 36.2%.Source: §4.3 · pp. 12–13

5. Discussion

• Most reactions to crude coordinated accounts come from within the network, although a substantial share of their mentions is addressed outside it.Amplifiers followed the accounts only weakly and with delay, while reactors did not follow them.Source: §5 · p. 13
• Within the dataset, political claims account for 0.038% of coordinated posts, and the two major activity surges show no clear relationship with external political events.Only the first surge coincided with a change to X’s payout rules; monetization is a plausible interpretation, but the analysis does not establish causation.Source: §5 · p. 13
• Explicit sexual or illicit-trade vocabulary reaches outside the Chinese-speaking population at around 40%, whether classified as illegal material or sexual solicitation.The authors favor vocabulary over category labels because human validation found the category labeling imprecise.Source: §5 · p. 13
• Because this vocabulary reaches furthest outside the coordinated population and its main category is suspended less often than most others, vocabulary-based detection may inform moderation research.This is a proposed research direction, not a demonstrated moderation outcome.Source: §5 · pp. 13–14
• The findings cover X’s Japanese-classified stream and relatively crude coordination; phrasing-variable coordination is not captured, and the November 2024 surge remains tentative.The authors also note that the excluded population includes Chinese-language accounts below the posting threshold, so it is an upper bound on ordinary-user engagement.Source: §5 · p. 14

A.1. Amplifier and reactor accounts.

• 1.83% of reposts, 0.50% of replies and mentions, and 0.005% of quotes directed at coordinated accounts define the action-specific chance baselines.For each interaction type, the account-level test compares observed targeting against a binomial null using the corresponding baseline.Source: §A.1 · p. 18
• Accounts must have a majority of their actions of that type directed at coordinated accounts (s_i ≥ 0.5), in addition to a statistically significant excess.The significance test uses a binomial upper tail and Benjamini–Hochberg adjustment at α = 0.05.Source: §A.1 · p. 18
• The majority-share requirement prevents accounts with many interactions from qualifying solely because a negligible targeting bias is statistically significant.Source: §A.1 · p. 18

B.1. What the clusters can and cannot support.

• HDBSCAN clustered a uniform 1,200,000-text sample, then every text was assigned to its nearest centroid without a similarity threshold, including noise points.The sample size was set by memory constraints; 57.7% of sampled texts were marked as noise.Source: §B.1 · p. 18
• A topic’s label describes its sampled dense core, while its periphery is attached by proximity alone and is not guaranteed to fit that label.Source: §B.1 · p. 18
• Only 10.4% of post volume in topics labelled sexual solicitation was classified as sexual solicitation text by the individual-wording method.Everyday topics accounted for 35.5% and no recoverable message for 34.9%; the gap affects counting, not which kinds of text entered the codebook.Source: §B.1 · p. 18

B.2. The frozen codebook.

• The frozen codebook and its brief category decision rules are provided in an anonymized repository and Table A1.Source: §B.2 · p. 18

B.3. A fixed classification hierarchy eliminates this ambiguity.

• Table A1 lists nine categories and their decision rules in checking order; the fixed hierarchy resolves overlap by prioritizing rare categories over generic ones.The draft-to-frozen revisions renamed and redefined six categories, kept two, removed one, and added one, rewriting intent-based definitions as observable criteria.Source: §B.3 · pp. 18–19, Table A1 · p. 19, Table A2 · p. 19

B.4. Codebook development and refinement.

• 36.1% of post volume received different categories in two runs of the nine-category draft, with every large disagreement involving low-quality spam.That category asked whether a text was a mass-produced template, whereas the other categories asked what it was about.Source: §B.4 · p. 19
• Separating content categories from a mechanically computed repetition attribute raised agreement to κ = 0.629 (67.7% raw).A topic was templated if its most frequent text accounted for at least 50% of posts or its eight most frequent texts accounted for at least 80%.Source: §B.4 · p. 20
• Removing the unsupported “mixed” option raised agreement to κ = 0.660 (70.8% raw), while differing labels fell from 19.59% to 15.11% of post volume.Source: §B.4 · p. 20

C.1. Cascade depth by root group.

• Figure A1 compares maximum cascade depth and size for coordinated-account roots versus Chinese general-user roots.Panels e–f also show the share of reactions at each depth by reactor group.Source: §C.1 · p. 20

C.2. Who shares at each depth.

• Figure A2 decomposes cascades by depth within each root category, but seven of eleven categories have fewer than 25 nodes at depth five or beyond.Node counts are printed above each bar.Source: §C.2 · p. 20

C.3. Most frequent coordinated texts.

• Table A3 lists the 30 most frequent normalized coordinated texts, their English glosses, post totals, and most frequent unnormalized forms.Source: §C.3 · p. 20

• 96.79% of the coordinated accounts—710,595 of 734,173—had been suspended by X at the time of the API query.Source: §C.4 · p. 20
• Across all topics, at least 84% of coordinated accounts had been suspended by X.Source: §C.4 · p. 20
• The reported suspension figures were obtained using the X API and describe account status at the time of the query.Source: §C.4 · p. 20

• 44.1% of roots containing terms denoting minors reached non-coordinated accounts, versus 41.2% of roots containing adult sexual terms within the illegal-material category.The respective root counts were 824 and 3,684.Source: §C.5 · p. 20
• 36.2% of sexual-solicitation roots with adult sexual terms reached outside accounts, compared with 10.7% for the category overall.The respective root counts were 9,994 and 69,029.Source: §C.5 · p. 20
• Figure A2 reports who reacted at each cascade depth by the root post’s content; the reach analysis uses vocabulary-term matching rather than assigned categories.In the illegal-material category, 2,299 roots (33.8%) matched none of the three term lists, and the partition was not manually re-annotated.Source: §C.5 · p. 20, Figure A2 · p. 22

• 63.0% of Chinese-language accounts met the criterion in March 2025, compared with 10.7% of Japanese-language accounts.The counts were 58,331 of 92,548 Chinese-language accounts and 61,206 of 573,058 Japanese-language accounts.Source: §C.6 · p. 22
• The March 2025 comparison applied the same kana split, minimum of 13 body posts, normalization, and thresholds of three accounts and three repetitions.The month fell between the two surges, and the criterion counts repetitions within a period.Source: §C.6 · p. 22
• The Japanese-language population posted 42.8 million body posts versus 11.6 million for the Chinese-language population, so corpus size does not explain the share difference.Source: §C.6 · p. 22