AI Can Boost Your Work While Quietly Eroding Your Learning, New Model Warns

Generative artificial intelligence has quietly created a paradox in classrooms and workplaces: the same tool that helps people produce polished essays, clean code, and confident analyses may be undermining the very learning that makes such expertise possible. A new theoretical paper published in Communications Psychology introduces a framework called the SYNC model—Synergetic Notion of Human-AI Skills and Knowledge Construction—that attempts to explain when working with AI builds durable human competence and when it merely manufactures a convincing performance illusion. The authors, led by Yizhou Fan and Dragan Gašević, argue that fluent, coherent AI assistance can improve immediate task performance while weakening the metacognitive engagement and deeper processing required for lasting learning.

The central insight of the SYNC model is that learning and performance, long treated as two sides of the same coin, can decouple when AI contributes substantial cognitive and metacognitive processing to a task. In the classical view of education, improved task performance is taken as evidence of learning: as competence grows, performance improves, and better results signal better skills. Generative AI breaks this coupling. A learner may appear highly competent while using AI support yet show little lasting change in knowledge or skills once that support is removed. The authors illustrate this with a driving analogy: a person with minimal driving skill might still reach a destination safely if an autonomous system handles perception, planning, and control. Task performance can be achieved with external support even when the learner’s internal skill has not changed.

To map this territory, the model places learning on a vertical axis and performance on a horizontal axis, creating four quadrants that describe distinct states of human-AI interaction. The first quadrant, called Floundering, captures learners who show both low learning and low performance, often because they lack prior knowledge, adequate scaffolding, or effective task support. The second, Effortful Learning, describes learners who carry most of the cognitive processing themselves while scaffolding—temporary, contingent support from a teacher, peer, or AI agent—helps them regulate that effort. The third quadrant, Autopilot Performance, is where the danger lies: AI-supported task performance exceeds what learners could achieve unaided, while learning remains limited. The fourth, Synergetic Mastery, represents the ideal state in which AI support raises performance while learners retain responsibility for planning, monitoring, and evaluation.

The model builds on a key conceptual distinction the authors introduce: metacognitive offloading versus metacognitive onloading. Cognitive offloading—delegating storage, search, or computation to an external resource—can be beneficial, freeing mental capacity for deeper work such as comparing claims or testing edge cases. But metacognitive offloading occurs when learners go further and delegate or bypass their own planning, monitoring, evaluation, and judgment calibration to AI. In such cases, fluent AI-supported performance biases their sense of understanding and reinforces future reliance. The authors describe five problematic scenarios: learners may skip metacognitive effort on tasks perceived as difficult or low-value; they may lack the domain knowledge to activate metacognition on novel tasks; they may outsource strategy evaluation to AI; misleading feedback may reinforce the offloading loop; and fluent, cost-effective interactions may inflate confidence and create the illusion of competence that does not exist independently.

Empirical evidence cited in the paper lends weight to these concerns. In a controlled experiment on academic writing, students using generative AI produced higher-quality essays but engaged far less in planning, monitoring, or evaluation. In high school mathematics, students with access to an AI tutor performed better during AI-supported practice yet scored lower on a subsequent closed-book exam when the tool was removed. Longer-horizon classroom evidence from an introductory computer science course reported strong gains on midterm and quiz assessments completed with AI but no significant advantage on a final exam completed without it. Lower-performing students in that course showed a pattern of high dependence and low trust that signaled reliance without effective evaluation. The authors also cite a meta-analysis by Vaccaro and colleagues finding that when AI outperforms humans in decision-making tasks, human-AI interactions often fall below the AI-only baseline.

The paper identifies four specific pitfalls that trap learners in the autopilot quadrant. The Pitfall of Fluency arises because AI’s instant, well-phrased responses create cognitive ease, tempting learners to treat easy processing as proof of understanding. The Pitfall of Pandering stems from AI systems optimized to be agreeable, which can drift into affirming learners’ beliefs rather than challenging them, producing an echo chamber that discourages misconception repair. The Pitfall of Coherence occurs when AI outputs appear internally consistent and well-structured, making unsupported claims feel plausible because they fit smoothly into a narrative. The Pitfall of Transparency emerges when visible rationales, step-by-step explanations, and citations make the system appear inspectable, leading learners to treat explanation visibility as a guarantee of correctness. Research cited in the paper shows that explanations can increase reliance on AI responses even when those responses are wrong, and can reduce users’ tendency to ask follow-up questions—precisely the opposite of what transparency is intended to achieve.

For each pitfall, the authors propose a corresponding strategy grounded in what they call metacognitive onloading. Against fluency, they recommend desirable frictions: intentionally introduced, learner-controlled micro-interruptions that slow unreflective fluency at key moments, such as when a learner is about to accept an AI output. An AI acting as scaffolding might first check whether it has correctly understood the learner’s intention, prompt the learner to think about quality criteria before and after receiving an answer, and remind them what judgments they made too easily in the previous interaction. Against pandering, they propose engineered dissent, in which AI generates counterarguments based on data beyond what the learner provides, acts as multiple agents thinking from different perspectives, and explicitly labels its own assumption-dependence and uncertainty. Against coherence, they recommend balanced aid that defaults to the minimum necessary help—a single next-step hint or diagnostic question—with deeper assistance unlocked only after learners actively request it. Against transparency, they advocate contextualized verification, in which AI acts as an equal collaborator linking sources of information and addressing uncertainties within the learner’s specific context rather than as an authoritative tutor.

Scaffolding itself must be calibrated rather than maximal, the authors caution. Effective scaffolding operates at three levels: micro-level supports within a single prompt, such as guiding questions and verification steps; meso-level supports across a task, such as milestones and clear roles for AI versus human judgment; and macro-level supports over time, where assistance is phased out as competence develops. Excessive scaffolding can impede metacognitive development, disrupting learner autonomy and weakening self-regulation. The goal is calibrated, fading scaffolding that preserves learner self-regulation while ensuring metacognition is engaged at the right moments. Scaffolding triggers should be based on task novelty, high consequences of errors, or low learner calibration, and should be removed once learners consistently include planning, prediction, and checking in their learning process.

The authors are explicit that the SYNC model is a conceptual, theory-building framework rather than an empirically validated one. Its proposed mechanisms and outcomes have not yet been systematically tested, and few established measures exist for key constructs such as metacognitive offloading and onloading. The model is most relevant in contexts where durable human competence, assessment validity, safety, or transferable expertise matter, not in purely instrumental uses such as automating routine tasks or drafting one-off projects where cognitive offloading may be an efficient use of effort. The authors also note that the evidence base remains emerging, limiting strong claims about generalisability across educational, professional, and cultural contexts. They propose future research questions, including which learning tasks most strongly decouple AI-supported performance from durable learning, and which observable indicators can validly distinguish metacognitive offloading from metacognitive onloading during human-AI interaction.

The practical stakes extend beyond the classroom. In an era defined by pandemics, energy crises, and climate change, the authors argue that the synergy of AI capabilities with durable human skills is becoming indispensable. AI can accelerate data-informed insights and automate routine work, but human professionals remain essential for formalising ill-defined problems, generating novel hypotheses through theory-based causal reasoning, guiding ethical decisions, and preserving the creativity and adaptability that underpin innovation. The challenge, as the SYNC model frames it, is to reap AI’s benefits without eroding the learning process that makes such human-AI interaction effective. The authors call on researchers, educators, and developers to co-design AI agents and learning environments with contingent, fading scaffolds and process-sensitive evaluations so that AI boosts productivity without eroding learning—ensuring that the next generation of learners builds genuine expertise alongside powerful tools rather than a fragile dependence on them.

Subject of Research: A theoretical model of metacognitive offloading and onloading in human-AI interaction and its implications for durable learning

Article Title: A metacognitive approach to learning and performance in human-AI interaction

Article References: Fan, Y., Yan, L., Molenaar, I., Greiff, S., & Gašević, D. (2026). A metacognitive approach to learning and performance in human-AI interaction. Communications Psychology, 4(1), Article 150. https://doi.org/10.1038/s44271-026-00548-y

Image Credits: AI Generated

DOI: 10.1038/s44271-026-00548-y

Keywords: generative AI, metacognition, learning, human-AI interaction, scaffolding, cognitive offloading, SYNC model, education, overreliance, performance, durable learning, Communications Psychology