Fanchuan Behavior in Livestream Chats

• Fanchuan behavior is a covert form of online antisocial conduct defined by feigned support followed by targeted offensive messaging to harm reputations.
• It employs strategic deception and contextual camouflage by mimicking genuine fan support, making detection challenging for both human moderators and automated systems.
• Empirical studies reveal recurring patterns, such as toxicity spikes and high repeat offender rates, underscoring the need for hybrid moderation systems.

Fanchuan behavior is a covert, indirect form of online antisocial conduct that has gained prominence in livestreaming chat environments, particularly on large-scale Chinese platforms such as Bilibili. Unlike traditional trolling or harassment, fanchuan involves users who masquerade as supporters of a target entity—such as a celebrity, game, or brand—only to subsequently engage in offensive or irritating messaging that damages the entity’s reputation by association. This two-step, deceptive process exploits contextual ambiguity, making it uniquely difficult for both human and automated moderation systems to distinguish between genuine and malicious interactions. The strategic intent is long-term reputational harm, achieved through context-specific infiltration and exploitation of trust.

1. Definition and Distinguishing Characteristics

Fanchuan behavior is defined as a “novel, covert and indirect form of social attack” occurring primarily in real-time online contexts, such as livestreaming chats (Wei et al., 31 Aug 2025). Its core characteristics are as follows:

• Feigning Support: The perpetrator exhibits outward allegiance or positive sentiment toward the target, often adopting linguistic markers or symbols typical of genuine fans.
• Subsequent Offensive Messaging: Immediately following support, the perpetrator introduces remarks that are offensive, irritating, or reputationally damaging, with the intent to undermine both the entity and its surrounding fan community.
• Contextual Camouflage: Harmful messages are embedded within otherwise normal chat flows, creating a “gray area” that obscures intent.
• Indirect Attack Vector: Unlike overt insults or harassment, fanchuan’s effect is mediated by manipulation of perceived community identity.
• Location: Attacks frequently occur outside the target’s immediate community (e.g., in related or larger platforms), leveraging bystanders as amplifiers of reputational harm.

An illustrative example includes messages of apparent solidarity immediately followed by derogatory remarks: “We players of $G$ will never play your trash game,” where $G$ identifies the target.

2. Mechanisms and Strategic Dynamics

Fanchuan behavior leverages multi-layered deception tactics:

Step	Description	Contextual Implication
Feigned Allegiance	Using community symbols or positive sentiment to blend in	Authenticity camouflage
Offensive Messaging	Provocative, insulting, or irritating language deployed immediately after positive signaling	Reputational undermining by mimicry
Environmental Mixing	Intermixing with non-harmful chat traffic (benign, neutral, or genuinely supportive messages)	Detection avoidance

The strategic objective tends toward slow, cumulative damage rather than immediate disruption. Attackers rely on contextual ambiguity to induce bystander confusion, diluting the target’s credible fan base and amplifying negative sentiment through association and repetition.

A plausible implication is that this behavior artificially inflates toxicity metrics within community sentiment analyses, misleading both automated systems and platform governance efforts regarding the actual support structure of a target entity.

3. Quantitative Impact and Empirical Findings

Empirical assessment of fanchuan behavior was conducted via a large-scale dataset encompassing 2.7 million livestreaming sessions and 3.6 billion chat messages. Key findings include:

• Prevalence: 130,000 distinct fanchuan instances identified across 37,400 livestreaming sessions, with highest concentration in gaming streams (especially major esports titles such as League of Legends).
• User Behavior: Approximately 88% of individuals engaging in fanchuan are repeat offenders; >93–99% focus on a single entity or category.
• Temporal Patterns: Only 30% of incidents occur within the first 300 seconds after a user enters a session, highlighting delayed engagement and strategic timing.
• Chat Dynamics: Fanchuan events trigger a surge in gross message volume followed by a rapid drop in unique messages once duplicates are filtered, indicating mass mimicry.
• Sentiment Shifts: Sentiment and toxicity scores spike in proximity to flagged fanchuan episodes (from –10s to +20s of the event), confirming immediate chat atmosphere disruption.
• Topic Distribution: BERTopic analysis captures a rise in both directly and indirectly related topics, further disrupting chat coherence.

For timestamp estimation of fanchuan behavior, the following formula was applied:

$T_f = T - 11.96$

where $T$ is the flag time; $T_f$ estimates the true event occurrence.

4. Challenges in Detection and Moderation

Traditional moderation frameworks prove insufficient for fanchuan detection due to its covert, camouflaged nature:

• Human Moderators: Highly context-dependent; require historical, situational, and linguistic background to distinguish genuine support from sabotage.
• Automated Systems: Keyword and toxicity matching underperform due to abbreviation, emoji use, and context-specific cues; real-time reaction constraints limit remediation before harm occurs.
• Reputational Persistence: Even rapid flagging and message removal leave potential for lasting community perception damage, as bystanders may have already internalized the deceptive messages.

The recommended approach is hybrid moderation: leveraging crowd-sourced flagging (key terms “fanchuan,” “chuanzi,” “biechuan” are notable markers) in tandem with machine learning models that analyze aggregated user histories and behavioral features, instead of only individual messages.

5. Machine Learning Solutions for Perpetrator Identification

To address the identification challenge, robust feature engineering and learning algorithms have been applied:

• Features: Account age proxies (log-transformed user ID), historical chat activity, entity targeting frequency, sentiment and toxicity profiles.
• Models: Random Forest and Histogram-Based Gradient Boosting achieved promising ranking precision—over 34% of fanchuan users are identified in the top rank, and 53% within top five positions.
• Practical Integration: Such models can be embedded within platform moderation pipelines to relieve human moderators, enabling scalable real-time detection.

The feature selection and temporal modeling reflect a methodological advance, as traditional one-message classifiers are replaced with aggregate behavioral models capable of capturing covert attack patterns.

6. Theoretical Context and Broader Implications

Fanchuan behavior represents a unique evolution of online antisocial strategies, distinct from direct trolling, flaming, or cyberbullying. It is formally defined as both “covert” and “strategic,” operating through reputation manipulation rather than direct interpersonal harm. Empirical studies on platforms such as Bilibili have quantified its disruptive effects and developed methods for profiling users, identifying high-frequency offenders and characterizing their long-term targeting tendencies.

A plausible implication is that as antisocial behavior migrates toward indirect modalities, moderation tools and platform governance must increasingly incorporate historical, behavioral, and community-contextual analytics rather than relying on message-level toxicity classifiers.

7. Comparison With Analogous Collective Fan Dynamics

The emergence of coherent layered attitudes in fan populations, as detailed in “Onset of coherent attitude layers in a population of sports fans” (Hristovski et al., 2011), demonstrates related group-level phase transitions in response to global coupling with environmental stimuli. Empirical findings show that multiple coherent layers—differing across opinion, emotion, and action strata—arise under repeated exogenous provocation. Fanchuan attackers exploit these intrinsic susceptibilities: by manipulating community identity from outside, they can seed proto-group formation and reputation stratification, even in highly integrated environments.

This suggests that the layered coherence in traditional fan populations provides fertile ground for indirect, reputation-targeted attacks such as fanchuan—where external stimuli are leveraged to guide community segmentation and attitude polarization.

---

In summary, fanchuan behavior is an empirically characterized, indirect form of reputational attack via feigned support and subsequent harm, highly prevalent in gaming livestreams and strategically deployed for long-term impact. Its detection and moderation require advanced, hybrid analytic frameworks that consider historical user context and aggregate behavior, given its limitation of message-level cues and rapid real-time dynamics. The phenomenon is theoretically linked to broader mechanisms of collective attitude formation and group layering within populations subject to global informational coupling.