Abstract:

[Purpose/Significance] Existing studies of data bias have primarily examined training data, model algorithms, and output-level errors, with limited attention to how scholarly information is selected, reconstructed, confirmed, and reused after large language models (LLMs) are adopted in research. In LLM-assisted research, literature, search results, dialogue records, summaries, and generated texts are repeatedly processed and reintegrated into knowledge production. Bias may therefore arise not only before or within model generation, but also through human–LLM interaction across research tasks. To address this gap, this study adopts a dynamic view of data in knowledge production and extends the analysis of data bias from training data and model outputs to the socio-technical process of academic research. It examines the practical forms, generative conditions, and governance of data bias in the humanities and social sciences, while distinguishing observable bias forms from the technical, behavioral, and contextual conditions that shape them, and highlighting researchers as both potential amplifiers and active correctors. [Method/Process] Grounded theory was adopted to identify recurring patterns and category relations from actual research practice rather than test a predetermined causal model. Empirical materials were collected in two stages from 56 participants, including 8 undergraduates, 18 master's and doctoral students, 26 university teachers, and 4 journal editors. The first stage, conducted from December 2023 to December 2024, used purposive sampling and generated materials from 41 participants. Open coding and preliminary categorization identified initial concepts and relations requiring further elaboration. The second stage, conducted in July 2026, added 15 theoretically sampled participants and focused on underdeveloped relations involving human-LLM trust and use states, researcher capability and task configuration, AI-text detection, academic-community use norms, and disciplinary differences. Second-stage materials were compared item by item with the concepts, properties, and category relations formed in the first stage. Researcher practice records and public online discussions served as supplementary materials for constant comparison. Open, axial, and selective coding produced 86 initial concepts, 12 initial categories, and a theoretical model. Theoretical saturation was reached when later second-stage materials produced no new concepts, categories, or category relations. No fixed numerical frequency threshold was imposed. Relative stability or repeatability was identified through recurrent experiences within the same participant, convergent accounts across participants undertaking similar tasks, or recurrence across different data sources or research stages. Isolated anomalies, including one-off hallucinations, were retained as risk records unless repeated directional patterns emerged. [Results/Conclusions] Four parallel forms of data bias were identified. Knowledge-resource visibility bias concerns recurring differences in which literature, sources, or information become visible or remain omitted during retrieval and AI-assisted reading. Knowledge-reconstruction bias refers to relatively stable or repeatable shifts in information completeness, conceptual boundaries, argumentative strength, or modes of expression during summarization, explanation, translation, or rewriting. Human-LLM interaction feedback bias develops through sustained dialogue when model tendencies interact with researchers' questioning, trust, fatigue, and acceptance, thereby reinforcing particular viewpoints, problem frames, or judgment paths. Knowledge-production recirculation bias emerges when AI-mediated content, expression patterns, detection practices, and community norms are repeatedly incorporated into writing, evaluation, revision, imitation, circulation, and reuse. The four forms do not constitute a fixed sequence. They may enter at different points, interact with one another, and exhibit a possible mechanism of multiple entry points, non-linear formation, and cyclical accumulation. Technical, behavioral, and contextual conditions shape whether such tendencies are accepted, reinforced, corrected, or rejected. Researchers may amplify bias through early delegation of core tasks, accumulated trust, sustained interaction, and fatigue, while source verification, explicit correction, retention of core questions and judgments, and control of human-LLM collaboration boundaries can weaken or interrupt it. Governance should improve academic knowledge-resource provision and traceability, enhance transparency in model-based knowledge processing, strengthen researcher agency and collaboration boundaries, and develop revisable community norms for disclosure, detection, evaluation, and responsibility. The model remains exploratory: the sample is limited to the humanities and social sciences, much of the evidence is retrospective, and the two data-collection stages span a period of rapid LLM development. Future studies should combine longitudinal tracking, behavioral observation, experiments, and quantitative testing to examine boundary conditions and change over time.

Key words: large language models, academic research, data bias, knowledge production, human - LLM interaction, grounded theory