Abstract
This study examines ethical risks and workable governance for artificial intelligence in university ideological and political education (IPE) in China. Semi-structured interviews with 17 instructors from five universities in Chongqing conducted from March to June 2025 were analyzed using reflexive thematic analysis. Six themes characterize current practice: privacy and consent remain fragile in attendance, proctoring, and analytics; the teacher role shifts from authority to curator and ethical gatekeeper; recommendation and moderation shape visibility and the continuity of deliberation; assessment integrity benefits from process-based evidence and explicit disclosure; metric-driven activity targets can crowd out value reasoning; and governance and accountability depend on institutional capacity and consistent rules. The findings indicate that responsible integration requires governance at the point of use across classroom, platform, and institution, including course-level disclosure and granular consent, explainable moderation with instructor overrides and traceability, process-based assessment with AI-use disclosure, curated corpora linked to retrieval-augmented generation with citation binding, routine audits, and faculty training.
Introduction
As higher education institutions globally move to modernize through "AI + Education" initiatives, curricula that are traditionally human-centric and value-laden face unique ethical challenges. In the context of Ideological and Political Education (IPE), the stakes are particularly high: the primary objective is not merely information transfer, but the cultivation of critical judgment, conviction, and value reasoning.
This research identifies a significant gap between ambitious national policies and the practical reality of the classroom. Instructors are currently grappling with "black-box" recommendation algorithms that may narrow student perspectives, Large Language Models (LLMs) that generate polished but hallucinated arguments, and privacy frameworks that often force students into "all-or-nothing" data consent. To address these frictions, this paper proposes a "point-of-use" governance framework that addresses the specific interface where teachers, students, and AI systems interact.
Methodology and Participant Profile
The researchers employed a qualitative exploratory design using reflexive thematic analysis. Semi-structured interviews lasting 60–90 minutes were conducted between March and June 2025. Following the Braun & Clarke six-phase procedure, transcripts were coded at both semantic and latent levels using NVivo 12 to prioritize qualitative depth and interpretive sufficiency.
The study involved 17 IPE instructors from five universities in Chongqing, China. The participant pool represented a diverse range of academic ranks, from teaching assistants to associate professors, with professional experience in IPE ranging from 3 to 21 years. Their backgrounds covered various core courses, including "Introduction to Xi Jinping Thought" and "Modern Chinese History."
Key Findings: Ethical Friction Points
The thematic analysis revealed six core areas where AI integration creates ethical or pedagogical tension:
- The "Authority to Curator" Shift: Instructors are transitioning from being the sole source of knowledge to becoming "ethical gatekeepers." They must now help students deconstruct "fluent but shallow" AI outputs. As one instructor noted, while AI provides a polished surface, the teacher's role is to help the student "reconstruct the reasoning" behind the output.
- The Metric Trap: Learning Management System (LMS) dashboards often prioritize "activity targets" such as clicks, post counts, and time spent online. The study found that these metrics can "crowd out" actual value reasoning; students may learn to optimize for engagement targets without meaningfully engaging with the course material.
- Fragile Consent and Privacy: Privacy is frequently handled via a single, mandatory consent button at initial login. Instructors reported student anxiety regarding "bundled" consent, where opting out of invasive features like keystroke tracking or facial capture would result in losing access to essential course materials.
- Algorithmic Narrowing: Recommendation engines can inadvertently create "information cocoons." By feeding students content similar to their previous clicks, AI can make it difficult to surface contested viewpoints or foundational theoretical texts necessary for a well-rounded education.
- Opaque Moderation: Automated content moderation often results in "takedowns" for code-of-conduct violations without providing explanations. This lack of transparency can stifle classroom momentum and deliberation.
Proposed Governance Framework
To ensure responsible AI integration, the paper concludes that governance must be applied at three distinct levels:
- Technical Level: Implementing Retrieval-Augmented Generation (RAG) with "citation binding." This technical solution anchors AI outputs to a curated, authoritative corpus of texts, significantly reducing the risk of hallucinations and ensuring traceability.
- Pedagogical Level: Shifting toward process-based assessment. To counter the ease with which LLMs can produce final essays, instructors are encouraged to use version histories, "Research Plan Memos," and oral defenses to verify the student's ownership of ideas.
- Institutional Level: Moving beyond vendor-default settings toward "targeted clinics" for faculty. These sessions should focus on identifying algorithmic bias, managing privacy controls, and establishing explainable moderation with instructor override capabilities.
Limitations and Future Outlook
The study's findings are limited by its regional scope (Chongqing) and its reliance on self-reported data from instructors. Future research should incorporate direct classroom observations and "mechanism tracing" to observe real-time student-AI interactions. As AI models become more sophisticated, governance frameworks must remain dynamic to address evolving ethical risks.
Summary updated