Research Article
Scaffolding critical thinking with generative AI: Design principles for integrating large language models in higher
Synthesis: Vendrell & Johnston (2026) propose a design-oriented pedagogical framework for integrating LLMs into higher education in ways that strengthen rather than displace critical thinking. Drawing from cognitive psychology, educational theory, and AI ethics, the framework addresses the risk that unstructured LLM use leads to cognitive offloading, metacognitive disengagement, and reduced epistemic agency.
The Problem: Unstructured LLM Use
A global survey reports 86% of university students now use AI in their studies, with over half engaging weekly — primarily for summarization, grammar checking, and drafting (Digital Education Council, 2024). However, LLMs are not designed with educational goals in mind; they simulate understanding via probabilistic language modeling without intentionality or comprehension.
Research shows that unguided AI use correlates with:
- Weaker arguments and lower cognitive engagement (Stadler et al., 2024)
- Diminished attentional modulation (Kosmyna et al., 2025)
- Reduced independent evaluation, especially among younger students (Gerlich, 2025)
- Decreased scrutiny of outputs, increasing susceptibility to misinformation (Yatani et al., 2024)
- Shifted attitudes toward AI biases even when recognized (Fisher et al., 2025)
- Overall erosion of critical thinking, decision-making, and analytical reasoning (Zhai et al., 2024)
However, when students first engage independently before consulting AI, their outputs are significantly stronger (Kosmyna et al., 2025). This mirrors earlier technology integrations — calculators, search engines — where effects depend on pedagogical design, not the tool itself.
Six Essential Processes
The framework identifies six interconnected cognitive and metacognitive processes that must be deliberately supported in AI-enhanced learning environments:
| Process | Definition |
|---|---|
| 1. Conceptual interpretation | Actively constructing meaning by selecting, organizing, and integrating information; distinguishing core ideas from peripheral details. Aligned with Bloom's "Understand" and "Analyze" levels. |
| 2. Inferential reasoning | Generating warranted conclusions from evidence; identifying assumptions, discerning logical relationships, and predicting implications. The cognitive bridge between information and action. |
| 3. Evaluative judgment | Assessing credibility, coherence, and evidentiary support of claims. What Bielik & Krell (2025) call epistemic vigilance — the capacity to critically assess both source credibility and claim validity. |
| 4. Metacognitive AI Regulation in Education | Monitoring, evaluating, and strategically controlling one's cognitive processes. Transforms critical thinking from episodic acts into sustained, self-directed practice. |
| 5. Intellectual curiosity | The motivational disposition to explore ideas, ask questions, and pursue knowledge beyond instrumental goals. Fuels cognitive persistence and epistemic openness. |
| 6. Epistemic integrity | The ethical orientation to seek truth, evaluate claims fairly, and engage with complexity conscientiously. Combines intellectual honesty with critical reflexivity. |
Eight Design Principles
| Principle | Core Mechanism | Supported Processes |
|---|---|---|
| P1. Preserve cognitive friction | Require independent thinking before AI; use AI to generate counterarguments | Conceptual interpretation, Inferential reasoning |
| P2. Scaffold LLMs as thinking partners | Position LLMs as provisional collaborators, not authoritative sources | Inferential reasoning, Curiosity, Epistemic integrity |
| P3. Embed evaluation as standard practice | Structured checkpoints for cross-referencing and criteria-based assessment | Inferential reasoning, Evaluative judgment |
| P4. Activate metacognitive self-regulation | Planning templates, reflective journals, AI prompt logs | Metacognitive regulation |
| P5. Encourage intellectual humility and curiosity | Examine AI limitations, explore alternative perspectives, identify omissions | Curiosity, Epistemic integrity |
| P6. Foster epistemic integrity | Require justification of claims, multiple perspectives, reasoning under uncertainty | Epistemic integrity |
| P7. Align assessment with intended cognition | Reward reasoning quality over surface fluency; assess how students interpret, question, and integrate AI content | Evaluative judgment, Metacognitive regulation |
| P8. Balance AI-mediated and AI-free phases | Sequence pre-AI, during-AI, and post-AI activities; deliberate AI-free zones | Conceptual interpretation, Metacognitive regulation |
Connection to Existing Frameworks
This model aligns with the A principled way to think about AI in education: guidance for educators and policy makers based on goals, models principle of augmenting rather than displacing human cognition. The emphasis on sequencing echoes findings in The Evidence Base on AI in K-12: A 2026 Review that pedagogically-designed AI outperforms raw LLM chatbots on learning outcomes. The metacognitive focus connects to Self-Regulated Learning cycles of planning, monitoring, and evaluation, while drawing on feedback literacy principles (Carless & Boud, 2018).
P1's use of counterarguments as cognitive stimuli parallels the Socratic Method approach of using questions to foster expert-like reasoning. P7's assessment alignment extends Formative Assessment principles into AI-mediated contexts, and the emphasis on preserving learner agency connects to Building AI Companions that Prioritise Learning over Performance which prioritizes learning over performance. The focus on epistemic integrity and the risk of over-reliance also speaks to Transfer of Learning — the central question of whether AI-assisted gains persist when tools are removed.
P2's scaffolded positioning of LLMs aligns with AI Literacy frameworks that progress from understanding to critical engagement to creative application, and the framework's overall design-based research methodology parallels the evidence standards discussed in The Evidence Base on AI in K-12: A 2026 Review.
Practical Application
Two illustrative scenarios demonstrate implementation:
Scenario A: Prompt Crafting and Critique. Students draft prompts individually without AI access, predict possible responses and limitations, analyze outputs using structured criteria (Which perspectives are prioritized? What is absent? What claims lack justification?), revise prompts with reflection, and identify ethical implications of prompt framing.
Scenario B: AI-Mediated Debate and Synthesis. Students engage in structured debate with AI as a provisional interlocutor, followed by independent written synthesis requiring assumption identification, counterargument consideration, and integration of AI-generated content with original reasoning.
Both scenarios emphasize the critical sequence: independent thinking first, AI interaction second, reflective integration third.
Open Questions
- How do these principles scale across class sizes, disciplines, and institutional contexts?
- What is the minimum threshold of AI literacy needed for students to benefit from scaffolded integration?
- How should faculty development programs train instructors to implement these principles at scale?
- What empirical evidence is needed to validate the framework's efficacy claims?
What this means for practice
- Instructors. Preserve cognitive friction: require independent thinking and a first draft before any AI consultation, then use the LLM to generate counterarguments rather than answers (P1).
- Instructors. Embed evaluation as standard practice rather than an optional add-on, with structured checkpoints where students ask which perspectives an AI output prioritizes, what is absent, and which claims lack justification.
- Instructors. Align assessment with the cognition you intend: reward reasoning quality over surface fluency and assess how students interpret, question, and integrate AI-generated content rather than how polished the output reads (P7).
- Faculty developers. Prepare educators for the shift from facilitator to "cognitive orchestrator" (UNESCO, 2023), covering how LLMs work, how to sequence AI-free and AI-mediated phases, and ethical sensitivity to data provenance, algorithmic bias, and platform dependency.
- Faculty developers. Plan for systemic constraints rather than assuming good design spreads on its own: reported barriers include ethical concerns, fragmented curricula, inadequate infrastructure, and limited faculty training, and institutional policy frameworks remain uneven on data privacy, equity, and continuous evaluation.
Limitations
- The study is conceptual and theoretical: it reports no human participants, no dataset, and no empirical validation of the framework's effectiveness.
- The design principles rest on established theory and recent research but are untested in classrooms; the authors call for classroom-based interventions, design-based research, or comparative studies before efficacy claims are made.
- The authors state that effectiveness and applicability may vary across disciplines, institutional contexts, and learner populations, since higher education systems differ in technological infrastructure, faculty expertise, and policy environments.
- The analytics indicators the framework proposes — prompt formulation and revision patterns, draft modification sequences, transitions between AI-free and AI-mediated phases — are not directly observable measures of learning and are described as requiring pedagogical interpretation rather than automated classification.
Citation
Vendrell, M., & Johnston, S.-K. (2026). Scaffolding critical thinking with generative AI: Design principles for integrating large language models in higher education. Computers and Education: Artificial Intelligence, 10, 100572.