Examining user migration intention from social Q&A communities to generative AI

The study investigates whether and why users intend to migrate from traditional social Q&A communities (e.g., Quora, Zhihu) to generative AI for knowledge Q&A. Motivated by the growing capabilities of generative AI (e.g., ChatGPT) to provide real-time, accurate, and engaging answers, and the potential negative experiences in Q&A communities (e.g., information overload, fatigue), the research adopts the push-pull-mooring (PPM) framework to model migration intention. It posits that push factors (low content quality, information overload) increase community fatigue, pull factors (perceived anthropomorphism, accuracy, trustworthiness) enhance flow experience with AI, and the mooring factor (social influence) affects intention directly. The purpose is to quantify these effects and identify combinations of conditions driving migration intention, thereby informing strategies for Q&A platforms to retain users.

User migration in IS research extends physical migration concepts to cyberspace, including shifting across media (offline↔online), between similar platforms, and to platforms within the same medium but with different characteristics. Migration to generative AI for knowledge services fits the third category, driven by efficiency and novelty. The PPM framework posits migration is shaped by push (negative attributes of origin platform), pull (attractiveness of destination), and mooring (contextual constraints/facilitators such as social influence, inertia, costs). Prior work links dissatisfaction, overload, and unwanted interactions to push; ease of use and security to pull; and social influence, inertia, and switching/continuance costs to mooring. This study applies PPM to the Q&A-to-AI context and specifies hypothesized links: H1 low content quality→community fatigue; H2 information overload→community fatigue; H3 community fatigue→migration intention; H4 anthropomorphism→flow; H5 accuracy→flow; H6 trustworthiness→flow; H7 flow→migration intention; H8 social influence→migration intention.

Design: Mixed-method analysis combining structural equation modeling (SEM) and fuzzy-set qualitative comparative analysis (fsQCA). Measures: Nine latent constructs with 3–4 items each measured on 5-point Likert scales, adapted from prior literature and contextualized to Q&A/generative AI (see Table 1 in paper for items and sources). Constructs: low content quality, information overload, community fatigue, perceived anthropomorphism, perceived accuracy, perceived trustworthiness, flow experience, social influence, migration intention. Pretest with 20 dual-experience users informed item refinement. Data collection: Online survey built on Credamo; link disseminated via WeChat and Weibo to users with experience using both social Q&A communities and generative AI; snowballing encouraged. Two-week collection yielded 532 responses; after screening (time <1 min, failed attention check, straightlining), 483 valid cases remained. Sample: 44.7% male, 55.3% female; 92.8% aged 20–29; education: 63.8% bachelor’s, 32.5% master’s+; commonly used Q&A communities: Zhihu (91.3%), Baidu Knows (82.0%), Xiaohongshu/Weibo/Douban (82.0%), education-related (65.8%), tech-related (37.5%); generative AI used: ChatGPT (66.9%), Baidu ERNIE Bot (24.6%), New Bing (24.2%). SEM analysis: Reliability and validity assessed with SPSS 26 and CFA. All Cronbach’s alpha >0.80; factor loadings mostly >0.70; CR >0.80; AVE >0.50; discriminant validity satisfied (square roots of AVE exceed inter-construct correlations). Multicollinearity not an issue (tolerance >0.5; VIF <2). Model fit: χ²/df=1.951, GFI=0.904, AGFI=0.883, NFI=0.902, IFI=0.950, CFI=0.949, RMSEA=0.044. fsQCA: Antecedents calibrated using 5%, 50%, 95% thresholds; necessity analysis (consistency threshold 0.90) revealed no single necessary condition. Truth table analysis with frequency threshold=5, consistency=0.80, PRI=0.70 identified six sufficient configurations for high migration intention; three had higher raw coverage and were emphasized.

SEM results: H1 not supported—low content quality did not significantly affect community fatigue. H2 supported—information overload significantly increased community fatigue (large effect; reported coefficient around 0.66). H3 supported—community fatigue positively affected migration intention (significant, smaller effect). H4–H6 supported—perceived anthropomorphism (β≈0.11), perceived accuracy (β≈0.31), and perceived trustworthiness (β≈0.28) positively influenced flow experience. H7 supported—flow experience strongly increased migration intention (β≈0.49). H8 supported—social influence strongly increased migration intention (β≈0.54). Explained variance: R²(flow)=0.57; R²(community fatigue)=0.39; R²(migration intention)=0.69. Measurement quality: All constructs showed good reliability (alpha >0.80), convergent and discriminant validity (CR >0.80, AVE >0.50), and acceptable global model fit (χ²/df=1.951; CFI=0.949; RMSEA=0.044). fsQCA results: Six sufficient configurations for high migration intention; three principal paths (higher raw coverage) were:

• S1: Information overload (core) × Community fatigue (peripheral) × Perceived accuracy (core) × Perceived trustworthiness (peripheral) × Flow experience (core) × Social influence (peripheral).
• S2: Information overload (core) × Perceived anthropomorphism (peripheral) × Perceived accuracy (core) × Perceived trustworthiness (peripheral) × Flow experience (core) × Social influence (peripheral), showing migration even without community fatigue when AI pull and social influence are strong.
• S3: Community fatigue (peripheral) × Perceived anthropomorphism (peripheral) × Perceived accuracy (peripheral) × Perceived trustworthiness (peripheral) × Flow experience (core) × Social influence (peripheral). Across configurations, flow experience was a recurring core condition; information overload and perceived accuracy frequently appeared as core; social influence tended to be peripheral in the main paths. Overall consistency of solutions ≈0.964 with raw coverage of the leading paths around 0.35.

Findings demonstrate that user migration intention from Q&A communities to generative AI is driven more by pull (quality and trust of AI responses and the ensuing flow experience) and mooring (social influence) than by push (fatigue from Q&A communities). Information overload in Q&A communities is a key antecedent of community fatigue, reinforcing literature that overload induces negative affect. Yet low content quality did not heighten fatigue, possibly because Q&A users search purposefully and filter for high-quality answers, and because quantity (overload) rather than quality primarily drives fatigue. On the AI side, accuracy and trustworthiness play larger roles than anthropomorphism in producing flow, indicating users’ utilitarian orientation in knowledge seeking. Flow and social influence exert strong direct effects on migration intention; fsQCA shows that enjoyable, immersive AI interactions combined with perceived accuracy and some social endorsement are sufficient for migration, even absent fatigue. The mixed-method evidence underscores that enhancing AI experience and leveraging social networks can substantially shift users away from traditional Q&A platforms.

This study applies the push-pull-mooring framework to explain intentions to migrate from social Q&A communities to generative AI. It shows that information overload elevates community fatigue (push), perceived accuracy, trustworthiness, and anthropomorphism enhance flow with AI (pull), and social influence (mooring) and flow are strong drivers of migration intention. The model accounts for 69% of variance in migration intention and fsQCA reveals multiple sufficient paths, with flow as a central condition. The work contributes by (1) extending Q&A user behavior research to cross-channel migration, (2) clarifying how AI-related perceptions translate into immersive experiences that promote migration, and (3) integrating cognitive and affective factors to uncover mechanisms of migration intention. Practically, Q&A platforms should curb information overload and mitigate fatigue; generative AI providers should prioritize accuracy, trust, and engaging interactions, and harness social influence to accelerate adoption. Future research should generalize to newer AI systems, incorporate additional determinants (e.g., privacy risk, dissatisfaction, habit), and examine actual migration behaviors beyond intention.

• Scope of AI systems: The study focused on prevalent generative AI (e.g., ChatGPT, ERNIE Bot); rapid AI evolution may limit generalizability to newer, more anthropomorphic models.
• Omitted variables: Other potential determinants (privacy risk, dissatisfaction, habit, switching costs) were not modeled and may affect migration intention.
• Behavioral measure: The outcome was intention rather than observed migration behavior; future work should track actual usage shifts.
• Sample characteristics: Predominantly young, highly educated users (20–29 years), which may limit generalizability to broader populations.