Introducing HAIML: A Human-Centered AI Metacognitive Learning Model — A Framework for Human Agency and Reflective Learning

Synthesis: HAIML is a human-centered framework for learning in AI-supported environments that preserves human agency, metacognitive awareness, ethical reasoning, and personal responsibility. Grounded in self-efficacy, Self-Regulated Learning, Experiential Learning, Metacognition, and automation-bias research, the model spans three interconnected layers — Experiential AI Use, Metacognitive Reflection, and Ethical Decision-Making — guiding learners from direct engagement with AI to reflection on how AI shapes their thinking and then to responsible judgment. A companion set of Four AI Use Levels aligns AI permission with an assignment's objectives rather than ranking assignments, with metacognitive reflection required at every level, including Level 1. The distinction worth keeping: AI Literacy frameworks ask what learners know about and can do with AI, while HAIML asks how AI changes their thinking, confidence, and decisions. The paper reports no sample, no outcome measures, and no test of the model, so its claims are design guidance rather than evidence.

Key Findings

1. A gap in existing frameworks. AI-literacy, technical-proficiency, responsible-use, and academic-integrity frameworks give little attention to how AI influences the learner's thinking and decision-making processes.
2. Three-layer model. HAIML organizes learning around Experiential AI Use, Metacognitive Reflection, and Ethical Decision-Making, moving learners from direct engagement with AI to reflection on how AI shapes their thinking, and finally to informed, responsible decision-making.
3. Agency-preserving design. The framework builds on human agency, self-regulated learning, metacognition, and automation-bias research to keep learners in control of AI-mediated cognition. Following Bandura, AI can generate recommendations but cannot hold responsibility, so judgment stays with the learner.
4. Four AI Use Levels. Levels run from no AI use to AI-integrated creation and evaluation, aligning AI permission with instructional goals rather than ranking assignments from better to worse; more AI is not treated as always preferable.
5. Reflection as the bridge. Metacognitive Reflection carries the framework's distinctive claim: students examine how AI shaped their thinking, confidence, revisions, and understanding at every level, including Level 1.
6. Two modes of thinking. Kahneman's System 1 and System 2 explain why fluent AI output invites fast acceptance, and why HAIML asks students to slow down and verify.

The Problem HAIML Was Designed to Solve

Generative AI now participates in cognitive processes traditionally associated with human reasoning — generating ideas, drafting arguments, summarizing research, and proposing solutions — so the question is no longer whether students will use these tools but how AI shapes their thinking. Reardon argues that learning is fundamentally a cognitive process, and that the central challenge is ensuring students remain active participants in thinking rather than passive recipients of AI outputs.

A primary concern is Cognitive Offloading, where students rely on external tools to reduce mental effort. While beneficial for higher-order focus, generative AI now supports increasingly complex thinking, risking gradual disengagement from the very processes education is intended to develop. A second concern is automation bias — the tendency to place excessive trust in automated systems — made worse because AI outputs appear fluent and authoritative even when incomplete or incorrect. Closely related is the illusion of understanding, where exposure to a polished AI explanation is mistaken for genuine comprehension. The framework also addresses authorship and responsibility, treating them through reflection and transparency rather than punishment.

Theoretical Foundations

HAIML is grounded in established psychological theory and learning science. Bandura's work on Self-Efficacy and human agency anchors the model: students must continue to view themselves as capable participants even as AI grows more capable, since AI can extend but never possess the intentionality, forethought, and self-reflection that make agency uniquely human. The paper quotes Bandura (2001): "to be an agent is to intentionally make things happen by one's actions" (p. 2).

Research on Self-Regulated Learning (Zimmerman) explains why learners must actively plan, monitor, and evaluate their interactions with AI. Experiential Learning theory (Kolb) supports the first layer, holding that learning emerges from engaging with experience and making sense of it rather than from passive receipt of information. Metacognition (Schraw & Dennison) forms the reflective core, and the work of Bjork, Dunlosky, and Kornell on illusions of learning warns that AI-generated explanations can mask real comprehension gaps. Kahneman's dual-process theory, Klein's naturalistic decision-making, and automation-bias research (Parasuraman & Riley; Cummings) round out the foundations — together reinforcing that technology alone does not produce learning and that responsibility remains human.

The Three Layers of HAIML

Layer One: Experiential AI Use. Students engage directly with AI through authentic, structured learning experiences — brainstorming, exploring perspectives, receiving Feedback, analyzing information, or running decision simulations. AI here functions as Scaffolding, supporting rather than replacing learning, so students leave with greater confidence in their own abilities rather than dependence on technology.

Layer Two: Metacognitive Reflection. Reflection is the bridge between AI use and meaningful learning and the feature that most distinguishes HAIML. Students monitor not only their own thinking but AI's influence on it — when AI expanded their thinking, challenged assumptions, or reduced deeper engagement — and they record what they verified and how the work might have differed without AI. Regular reflection serves as a safeguard against cognitive dependence and automation bias, turning experience into understanding.

Layer Three: Ethical Decision-Making. Students evaluate their AI experiences through lenses of responsibility, transparency, authorship, accountability, fairness, and judgment. AI can generate information and content, but it cannot assume responsibility for decisions; learners must critically evaluate and verify outputs, consider biases, and decide how information should be used.

The Four AI Use Levels

To make the framework actionable, HAIML is paired with Four AI Use Levels that align AI use with instructional goals: Level 1 (No AI Use) for assessing independent knowledge and foundational skills, where students still reflect on why independence mattered; Level 2 (AI for Brainstorming and Support) for idea generation, clarification, examples, and planning while final work remains student-generated; Level 3 (AI Collaboration) positioning AI as a collaborative partner while students evaluate and decide; and Level 4 (AI-Integrated Creation and Evaluation) for extensive use with emphasis on evaluating, comparing, and revising AI outputs. The paper is explicit that the levels do not rank assignments from better to worse and that more AI is not always preferable. Across all four levels, metacognitive reflection remains central, distinguishing HAIML from policies focused solely on disclosure or compliance.

Applications and Implications

Although developed in higher education, HAIML adapts to workforce development, professional training, leadership, communication, and healthcare education — any setting where human judgment remains essential. For faculty, implementation means aligning AI use to learning objectives, normalizing reflection as part of the learning process, and prioritizing transparency and meaningful cognitive engagement over detection and enforcement.

For Metacognition and Self-Regulated Learning, HAIML offers a structured way to make AI interactions objects of reflection, addressing Cognitive Offloading and over-reliance risks by keeping learners aware of how AI influences their thinking. It complements AI Literacy frameworks by adding a metacognitive and ethical layer focused on internal decision-making rather than tool proficiency alone.

The model's human-centered orientation connects to human-AI collaboration design and to Ethics in education. As a framework paper, it would benefit from empirical validation of whether the three-layer progression measurably improves agency and reflection in AI-supported learning.

What this means for practice

• Instructors. Assign an AI Use Level to each task before students start it — Level 1 (no AI) where you are assessing independent knowledge, Level 2 for brainstorming with final work still student-generated, Level 3 for AI as a collaborative partner, and Level 4 for extensive AI-integrated creation with the emphasis on evaluating and revising outputs.
• Instructors. Treat metacognitive reflection as a required, assessed layer of AI-supported work rather than an optional add-on, and keep it in place even at Level 1. Ask learners to report where AI expanded the reasoning, challenged an assumption, or reduced engagement, and normalize that reflection as part of the process rather than as a compliance disclosure.
• Instructional designers. Keep analysis, evaluation, decision-making, and reflection with the student at Levels 2 to 4, using AI for brainstorming and feedback without letting it produce the final artifact.
• Learners. Record, for each AI interaction, which decisions were yours and how you verified the output, so that authorship, bias, and accountability stay visible in your own record rather than in the tool's.
• Faculty developers. Reframe AI workshops around this metacognitive and ethical layer instead of tool proficiency alone, since HAIML is pitched as the layer existing AI Literacy frameworks leave out.

Limitations

• The framework is conceptual: the paper reports no sample, no intervention, and no data, so nothing in it shows that moving through the three layers measurably improves Learner Agency, reflection, or ethical judgment — a validation gap the author states would need empirical work.
• The layers and levels are not operationalized: the paper offers reflection prompts but no validated instrument or fidelity measure, so implementations will vary in what counts as reflection.
• Its foundations are entirely secondary — Bandura on Self-Efficacy and human agency, Zimmerman on Self-Regulated Learning, Kolb on Experiential Learning, Schraw and Dennison on Metacognition, work on illusions of learning, and automation-bias research — so its coherence rests on how well those theories transfer to AI-mediated learning, which is asserted rather than tested.
• The Four AI Use Levels are proposed without validation: no reliability data, no evidence that different instructors place the same task at the same level, and no evidence that students respond differently across the four.
• It is a single-author preprint (Version 1.0, April 2026; EdArXiv June 2026, updated August 2026) that builds its case from prior theory rather than a systematic review, and the author notes that HAIML 2.0 is in preparation.

Citation

Reardon, C. M. (2026). Introducing HAIML: A Human-Centered AI Metacognitive Learning Model — A Framework for Human Agency and Reflective Learning in the Age of Artificial Intelligence. EdArXiv preprint (v2). https://doi.org/10.35542/osf.io/qhjgs_v2