"Polished Artifacts, Fragile Engagement? Tackling the Challenge of Reduced Epistemic Effort in Human-AI Knowledge Construction"

Synthesis: Drawing on CSCL research traditions, this paper conceptualizes the risk of reduced epistemic effort when learners use generative AI to produce knowledge artifacts. It identifies two strands of risk: a social-cognitive strand grounded in automation bias (attributing greater epistemic competence to AI) and an artifact-oriented strand focused on polished external artifacts inducing epistemic closure. The paper appeals to structure AI participation as an argumentative partner or challenger to preserve conflict and iterative refinement without diminishing human epistemic effort.

Key Findings

1. When learners use Generative AI autonomously and with little critical scrutiny to produce knowledge artifacts, they risk reduced epistemic effort: diminished active engagement in regulating, evaluating, and advancing shared knowledge, even when external outcomes look efficient and successful.
2. The social-cognitive strand explains this through automation bias—a systematic asymmetry in epistemic trust in which people attribute greater epistemic competence to AI than to their own judgment, weakening shared AI Regulation in Education, epistemic agency, and transactive engagement.
3. The artifact-oriented strand introduces the notion of a sense of epistemic closure: polished, rhetorically fluent AI-generated artifacts can appear exhaustive and authoritative from the start, pre-empting cognitive conflict and iterative refinement that normally drive learning.
4. These are conditional risk mechanisms rather than inevitable effects; the paper proposes design and measurement strategies to preserve distributed epistemic authority, sustain critique, and keep artifacts open as objects of inquiry.

Context: CSCL and the rise of hybrid collaboration

This paper revisits the foundational assumptions of Computer-Supported Collaborative Learning on the 20th anniversary of the International Journal of Computer-Supported Collaborative Learning. Historically, CSCL treated Collaborative Learning and knowledge construction as socially coordinated processes of meaning-making that depend on co-regulation, transactive discourse, and the epistemic agency of participants. Learning, on this view, is not merely the production of external artifacts but sustained human engagement in epistemic processes: regulating goals, engaging with others' ideas, and working with artifacts in ways that stimulate reflection and conceptual change.

Today, individuals increasingly work with large language models to get support for writing tasks. The paper conceptualizes these situations as hybrid collaboration—an individual working collaboratively with an AI partner—and interprets them through the lens of CSCL research. This is a functional rather than anthropomorphic extension: AI performs epistemic operations (generating and transforming content, detecting incompatibilities, integrating information) without consciousness or subjective experience. The central question posed is what happens to human epistemic effort and individual learning when key knowledge-related processes are increasingly performed by AI.

Empirical signs of reduced epistemic effort

Evidence on AI-supported learning is mixed. Studies of AI tutoring and chatbot-assisted guidance show that AI can improve performance, engagement, reflection, and self-regulation when embedded in instructional designs that require active learner participation. Yet a recurring, conditional pattern appears across domains: human-AI collaboration can remain highly productive at the level of external outcomes while simultaneously attenuating the epistemic processes through which learning and expertise typically develop.

In AI-supported academic writing, students using ChatGPT completed tasks faster and found them easier, but their texts showed lower originality and weaker indicators of authorship, with neurophysiological measures suggesting reduced engagement of working memory and semantic integration. When AI support was removed, these students struggled more than peers who had relied on search engines or received no support—indicating that sustained epistemic effort had not been consistently enacted. Complementary studies found that ChatGPT-supported revision primarily improved surface-level features while deeper aspects of content organization showed limited gains, and that participants tended to underestimate how much genAI contributed to jointly produced texts. Similar patterns appear in professional domains: in medicine, radiologists exposed to AI-generated recommendations were more likely to accept erroneous suggestions (automation bias), and in software development, continuous reliance on AI raises concerns about long-term Cognitive Offloading of Creativity and expertise. The reason for reduced effort is not AI use per se but the tendency to outsource epistemic work to AI.

Two strands explaining reduced learning

The paper accounts for reduced epistemic effort through two parallel explanatory strands rooted in CSCL theory.

The social-cognitive strand

This strand centers on automation bias—people's systematic tendency to attribute greater epistemic competence and decision accuracy to AI systems than to their own judgment, independent of whether that superiority is warranted. In hybrid collaboration, this attribution reshapes the distribution of epistemic responsibility: when AI-generated interpretations are assumed correct, individuals invest less effort in articulating goals, planning, monitoring progress, questioning assumptions, and evaluating quality. The result is reduced shared regulation, which in turn reduces transactivity—the active uptake of another participant's reasoning through critique, elaboration, integration, and transformation—and weakens epistemic agency (the willingness to take responsibility for advancing shared knowledge). From an SSRL (socially shared regulation of learning) perspective, this is not a simple absence of regulation but a disruption of regulatory coupling: humans remain formally involved while becoming progressively decoupled from the regulatory processes that normally sustain learning.

The artifact-oriented strand

This strand starts from the epistemic properties of polished external artifacts. Trialogical learning theory emphasizes that knowledge creation depends on collaboratively developing artifacts that remain provisional and open to criticism—artifacts function as epistemic objects precisely because their incompleteness invites inquiry and revision. The paper introduces the sense of epistemic closure as distinct from premature closure: whereas premature closure captures groups ending negotiation too soon at the interactional level, epistemic closure addresses how the perceived completeness of artifacts suppresses epistemic engagement even before critical negotiation or conflict emerge. When harmonized, apparently complete AI-generated artifacts make discrepancies less salient, learners perceive limited need to interrogate assumptions, and cognitive conflict—a central driver of conceptual change—is reduced. This weakens the co-evolution between internal knowledge structures and external knowledge products: external products continue to evolve efficiently while people's learning becomes increasingly decoupled from production.

Conclusion

GenAI introduces powerful new opportunities for collaboration and knowledge construction while challenging foundational CSCL assumptions when AI performs substantial parts of the epistemic work involved in producing external artifacts. Reduced human learning in hybrid collaboration can be understood through the social-cognitive strand (automation bias weakening shared regulation, transactivity, and epistemic agency) and the artifact-oriented strand (sense of epistemic closure suppressing criticism and elaboration). These are conditional risk mechanisms, not inevitable effects. The next phase of CSCL research should integrate automation bias more systematically, validate the sense of epistemic closure as a construct, and design AI systems that preserve distributed epistemic authority, sustain critique, and maintain artifacts as objects of inquiry rather than finalized products.

What this means for practice

• Designers. Build knowledge-building assistants that prompt users to articulate, justify, and evaluate their own goals — counter-questions, requests for justification, or deliberate withholding of a complete solution — rather than supplying finished content, because automation bias lets learners attribute more epistemic competence to the model than to their own judgment.
• Designers. Frame AI contributions as provisional and give artifacts explicit "uptake obligations" — deliberate gaps that require transformation or justification — so the co-produced artifact stays a relational object that demands epistemic work instead of a resource that can be accepted as is.
• Designers. Counter the sense of epistemic closure by design: offer multiple alternative drafts, deliberately incomplete representations, annotations of uncertainty, and outputs that embed counter-perspectives within themselves, so that discrepancies stay salient and cognitive conflict is not pre-empted by fluency.
• Researchers. Measure learning separately from product quality: use process analyses of which epistemic operations the human or the AI performed, trace data on contribution uptake, Cognitive Offloading measures, delayed tests, unaided explanations, and transfer tasks, since efficient external outcomes can coexist with attenuated learning.
• Researchers. Take up mechanisms CSCL has left untouched — automation bias is well studied in medicine and human-machine interaction but has not yet been examined in CSCL research on human-AI collaboration — and validate the new construct rather than assuming it.

Limitations

• This is a conceptual paper, not an empirical study: its author presents the two-strand framework as "a theoretically grounded extension of CSCL concepts that phrases propositions for future research rather than an empirically established model".
• Its strongest empirical warrant is secondhand and contested: the longitudinal ChatGPT writing study by Kosmyna et al. (2025) has been criticized for its small sample size, the interpretation of its neurophysiological measures, and its preprint status.
• The central new construct is unvalidated: the sense of epistemic closure is offered as a possibly relevant concept that "needs to be theoretically and empirically validated through adequate studies", and automation bias has not yet been taken up in CSCL work on human-AI collaboration.
• The evidence it assembles is mixed and largely correlational — AI tutoring improves performance and reflection where designs demand active participation, while AI-supported writers show lower originality and weaker retention without AI — so the risk mechanism is described as conditional rather than demonstrated.

Citation

Kimmerle, J. (2026). Polished Artifacts, Fragile Engagement? Tackling the Challenge of Reduced Epistemic Effort in Human-AI Knowledge Construction. EdArXiv preprint.