TweeXster: Analyzing Duplication Campaigns on X

• TweeXster is a framework that identifies duplication campaigns—accounts repeatedly sharing near-identical tweets—to amplify misinformation on X/Twitter.
• It employs SBERT embeddings and DBSCAN clustering to detect near-duplicate content, ensuring high precision in mapping coordinated behaviors.
• Applied to AltNews-verified data, TweeXster reveals that most duplicators are human-driven and politically clustered, underpinning persistent toxic discourse.

TweeXster is a framework introduced to study inauthentic duplication on X/Twitter: the practice of many accounts posting the same or near-identical tweet content to amplify misinformation, abuse, or otherwise manipulative narratives. Its central analytical unit is the duplication campaign—sets of accounts repeatedly sharing highly similar posts over time—rather than the isolated post or the binary question of whether an account is a bot. In the underlying study, TweeXster is used to ask what kinds of accounts engage in duplicate misinformation posting and whether prior duplication behavior predicts future inauthentic activity. The framework is developed in the context of Indian misinformation verified by AltNews and is used to show that harmful amplification on X/Twitter is often persistent, clustered, and human-driven rather than primarily a bot problem (Shafin et al., 18 Jul 2025).

1. Conceptual focus and problem setting

TweeXster is designed around the claim that coordinated misinformation and abuse are hard to stop when moderation and fact-checking only react to individual posts. The framework therefore looks for duplication campaigns: repeated, campaign-level patterns in which multiple accounts share identical or near-identical text over time. The paper explicitly distinguishes this from approaches centered on exact string matching or on bot detection alone, arguing that duplicate tweets are not always byte-for-byte identical because they may differ in URLs, minor edits, or formatting (Shafin et al., 18 Jul 2025).

This orientation matters because the study’s main empirical argument is that the main drivers of duplicated misinformation are largely human-controlled or non-bot accounts, not the kind of automated bot farms often assumed to dominate misinformation diffusion. A plausible implication is that campaign-level analysis can reveal manipulative behavior that remains invisible to systems tuned primarily for automation detection. In that sense, TweeXster treats duplication history itself as a signal of coordinated or recurrent inauthentic activity.

2. Corpus construction and analytical scope

The study is built around misinformation verified by AltNews, described as a prominent Indian fact-checking organization. The authors collected 433 AltNews reports published between April 2020 and April 2022, spanning categories such as politics, religion, news, society, media, and technology. From these reports they extracted 622 tweets directly linked in the reports, then used Twitter search to find an additional 6,431 tweets duplicating those reported examples. This produced metadata for 7,053 tweets posted by 5,747 unique accounts. After excluding suspended/deleted accounts, private accounts, and accounts that were actively debunking the misinformation, the final set contained 5,493 duplicator accounts (Shafin et al., 18 Jul 2025).

To study broader behavior beyond the original fact-checked examples, the authors collected the latest available tweets from each account in February 2023 using the Twitter API. This yielded over 12.7 million tweets. After removing retweets, they retained 6,879,220 original tweets. They then filtered out non-Hindi/English posts and posts with fewer than four unique words to reduce noise, producing the final dataset called Dup’23, consisting of 5,070,548 tweets. The study also curated a list of 1,166 specious news websites from prior work and Indian fact-checking organizations in order to examine whether duplicator accounts were also spreading links to unreliable sources (Shafin et al., 18 Jul 2025).

Methodologically, this corpus definition anchors TweeXster in a fact-checking workflow rather than a purely platform-wide crawl. The paper also notes an important boundary condition: the account set comes from AltNews-verified misinformation, so the study is anchored in a specific fact-checking corpus and may not capture all misinformation ecosystems.

3. Detection pipeline, clustering logic, and validation

TweeXster’s core task is to identify clusters of near-duplicate posts rather than exact matches. The preprocessing step removes URLs, mentions, and similar noise from tweet text. The framework then generates embeddings with SBert, specifically the paraphrase multilingual MiniLM-L12-v2 model, which supports 50+ languages including English and Hindi and produces 384-dimensional vectors. Clustering is performed with DBSCAN using a maximum Euclidean distance of 1 (Shafin et al., 18 Jul 2025).

The threshold of 1 was chosen empirically. The authors tested DBSCAN on the 7,053 misinformation tweets and varied the distance from 0.1 to 2. A threshold of 1 produced 427 clusters, which closely matched the 433 AltNews reports; 7,044 of 7,053 tweets were correctly labeled, with only 9 misclassified due to identical tweets appearing in different contexts. The paper emphasizes that density-based alternatives like HDBSCAN were less suitable because they grouped semantically related tweets too broadly, whereas the task here is to identify strict textual duplication or near-duplication (Shafin et al., 18 Jul 2025).

Applied to Dup’23, this procedure found 172,589 clusters, 514,958 duplicate tweets, 1,295,785 tweet pairs, and 736,147 account pairs, involving 4,938 accounts. The clusters averaged 2.98 tweets each, with a standard deviation of 3.01, and the largest cluster contained 398 duplicates. To validate cluster quality, the authors computed cosine similarity among tweet embeddings inside clusters and found a lowest observed similarity: 0.91 and an average similarity: 0.99. They also manually inspected 200 randomly selected clusters, and all of them were indeed duplicate-post clusters with no anomalies (Shafin et al., 18 Jul 2025).

The paper compares TweeXster against a Ratcliff/Obershelp pattern-matching baseline called ROPM, in two configurations. ROPM-10 compares each tweet to the next 10 tweets chronologically and found 42,820 tweets, 25,934 pairs, 2,613 accounts. ROPM-100 compares each tweet to the next 100 tweets chronologically and found 172,635 tweets, 147,555 pairs, 4,348 accounts. By contrast, TweeXster found 514,958 tweets, 1,295,785 pairs, 4,938 accounts. The comparison shows that TweeXster finds many more duplicate tweets, more pairs, and more involved accounts than the baseline methods.

4. Account characterization and the bot question

A major result of the paper is that these misinformation duplicators are not mostly bots. Using BotometerX, the authors marked an account as bot-like only when both its Complete Automation Probability (CAP) and Raw Bot Score (RBS) exceeded 0.9. By this strict criterion, only 44 of the 5,493 active accounts were bots, that is 0.8%, summarized by the paper as fewer than 1% (Shafin et al., 18 Jul 2025).

The dataset also contained 213 Twitter-verified accounts under the old verification system, and all of these had bot scores below the bot threshold. After excluding bots and verified accounts, the authors identify 5,236 regular misinformation duplicator accounts. This is one of the paper’s key empirical claims: the dominant actors in duplicate misinformation posting are largely not captured by a narrow automation frame (Shafin et al., 18 Jul 2025).

This directly addresses a common misconception in platform discourse: that duplicated misinformation is primarily the work of bot networks. The findings do not deny the existence of bots; rather, they show that a high-precision bot threshold captures only a very small fraction of the accounts engaged in this behavior. The paper also notes a methodological caveat: bot detection uses a high-precision threshold, so some automated accounts may be missed by design.

5. Campaign structure, super-duplicators, and political communities

To understand whether duplication is random or organized, the authors examined pairs of accounts that duplicated at least 10 tweets in common. This produced 8,118 account pairs involving 1,366 unique accounts; these are called super-duplicators. Among them, only 3 were bots, while 86 were verified accounts. Political affiliations were mapped using the NivaDuck database, yielding 163 political accounts, including 79 BJP-affiliated and 73 INC-affiliated (Shafin et al., 18 Jul 2025).

The paper then constructs a graph in which nodes are accounts and edges connect accounts that duplicated at least 10 tweets in common, and runs Louvain community detection. This produced 62 communities with modularity = 0.74. The study reports that 85.6% of the 1,366 accounts belonged to just 9 communities. Several of the largest communities were dominated by BJP-affiliated accounts, others by INC-affiliated accounts, and one had no clear political leaning. Importantly, the paper reports that no community mixed opposing political affiliations (Shafin et al., 18 Jul 2025).

In network-analytic terms, the reported modularity indicates strong within-community cohesion and relatively weak cross-community mixing. This suggests that duplication behavior tends to align along ideological lines rather than cross-cutting political groups. The result is significant because it places duplication campaigns within a structure of repeated co-participation rather than treating duplicate content as a flat collection of independent reposts.

6. Repeated misinformation, toxic discourse, and specious websites

A central finding is that duplication is not limited to the original misinformation verified by AltNews. The paper shows that the same accounts repeatedly spread new harmful content, including misinformation, abusive language, and links to unreliable websites. To detect new misinformation, the authors created a list of 20 keywords tied to contentious political topics and abusive narratives, including Russia-Ukraine War, Trump, Putin, Biden, Rahul Gandhi, Narendra Modi, Kashmir files, boycott, Adani, Hindutva, bulldozer, stone pelters, and love jihad. In Dup’23, 79,703 tweets from 32,900 clusters contained at least one keyword. The authors then selected the five longest-active clusters per keyword, yielding 200 clusters, and had two researchers independently fact-check them using Alt News methodology. Agreement was moderate, with Cohen’s Kappa = 0.62 (Shafin et al., 18 Jul 2025).

This procedure found 53 clusters containing previously unreported misinformation and 34 additional clusters containing abusive speech. The misinformation clusters included 209 tweets with 128,658 likes and 65,998 retweets. Examples included false claims about Pfizer, U.S. troops discarding medals, changes to UN Kashmir status, and Israeli death sentences for rapists. The paper also uses Google’s Perspective API to score toxicity, severe toxicity, identity attack, threat, insult, and profanity, treating scores above 0.5 as positive for a category. Because cluster members are near duplicates, the authors scored one randomly selected tweet from each cluster. The paper reports thousands of toxic clusters, including 4,013 labeled toxicity, 1,088 severe toxicity, 4,226 identity attack, 1,454 threat, 5,391 insult, and 931 profanity (Shafin et al., 18 Jul 2025).

The analysis of unreliable-source amplification uses the curated list of 1,166 specious news websites. The authors found 41,421 tweets in Dup’23 containing links to such websites, spanning 4,249 duplication clusters. Among the 1,366 super-duplicators, 531 accounts posted at least one tweet linking to such sites, and they were responsible for 12,822 of those tweets. The paper further reports that BJP-affiliated accounts posted 12,440 tweets linking to specious sites, while INC-affiliated accounts posted 343. A timeline for the first six weeks of 2023 shows 3,245 tweets linking to specious websites, averaging more than 72 per day (Shafin et al., 18 Jul 2025).

Taken together, these results show that duplication campaigns can function as continuing infrastructures for revived misinformation, abusive speech, and unreliable-source circulation rather than as isolated episodes tied only to the original fact-checked tweet.

7. Implications, limitations, and relation to adjacent research tools

The paper’s broader message is that misinformation amplification on X/Twitter is often persistent, clustered, and human-driven, not simply a bot problem. The authors conclude that duplication history can help prioritize fact-checking, detect campaigns earlier, identify accounts likely to spread future false or abusive content, and support moderation strategies that focus on coordinated duplicators rather than only automated bots (Shafin et al., 18 Jul 2025).

The study also states several limitations. The account set comes from AltNews-verified misinformation; the analysis uses a subset of accounts and tweets, and the authors explicitly note that many campaigns likely extend beyond the sampled population. Clustering relies on thresholds and embedding similarity, which can miss some duplicates or merge some near-duplicates depending on wording and URL variation. Bot detection uses a high-precision threshold, so some automated accounts may be missed by design. The study also focuses on Indian misinformation and may not generalize identically to other countries or language environments (Shafin et al., 18 Jul 2025).

In the broader tooling landscape, SocioXplorer is presented as an interactive social-media exploration system that extends a TwiXplorer / TweeXster-style workflow from Twitter/X-only archival analysis to a broader platform for Twitter/X and YouTube, adding batch-based incremental updates, community filters, labels, topics-per-community analysis, and network maintenance based on Force Atlas 2 and Louvain (Chausson et al., 23 Jun 2025). This suggests a wider methodological trajectory in which campaign-level or network-level analysis is integrated with interactive exploratory systems for longitudinal social-data research. Within that trajectory, TweeXster’s distinctive contribution is to operationalize campaign-level analysis of inauthentic duplication on X/Twitter and to show that a relatively small set of accounts can sustain large, repeated ecosystems of duplicated misinformation, toxic speech, and unreliable-source amplification—often without obvious bot behavior (Shafin et al., 18 Jul 2025).