Agentivism: a learning theory for the age of artificial intelligence

Synthesis: Yan and Gašević (2026) argue that generative and agentic AI change the conditions of learning by making it easy to delegate explanation, writing, and problem solving to systems that generate, recommend, and sometimes act on a learner's behalf. Successful performance then stops being evidence of learning: learners can complete tasks well while developing less understanding, weaker judgment, and limited transfer. The paper reviews what behaviorism, cognitivism, constructivism, and connectivism still explain about learning with AI and where each stops short, then proposes Agentivism, a mid-range theory for human-AI interaction. Agentivism defines learning as durable growth in human capability through selective delegation, epistemic monitoring and verification of AI contributions, reconstructive internalization of AI-assisted outputs, and transfer under reduced support. Monitoring and reconstruction may recur rather than form a fixed sequence. The authors derive six testable propositions from these mechanisms, treating verification and reconstruction as part of learning rather than as add-ons for responsible use.

Key Findings

• Agentivism is a mid-range learning theory for human-AI interaction that defines learning as durable growth in human capability, not as successful task completion with AI assistance.
• Its four mechanisms are delegated agency, epistemic monitoring and verification, reconstructive internalization, and transfer under reduced support, the last of which is the criterion distinguishing learning from assistance.
• Classical theories stop short: behaviorism explains habitual AI reliance, cognitivism offloading, constructivism dialogue, and connectivism distribution, but none explains when assisted performance becomes retained capability.
• Six testable propositions follow, including that learning is stronger when AI support preserves learner responsibility for problem framing, criteria setting, and justification than when it delivers answers.
• Verification should support transfer, and its absence should cost: requiring source checking or justification should improve delayed performance, while repeated low-friction delegation without reconstruction is predicted to weaken learners' calibration of their competence.

Why classical learning theories stop short

Earlier theories responded to earlier conditions of learning and remain useful, but each leaves a gap that intelligent delegation exposes. Behaviorism explains why AI support is reinforcing, since it is fast, fluent, and effort reducing, yet reinforcement cannot show whether competence stays with the learner once support is withdrawn. Cognitivism explains Cognitive Offloading and the divergence between feeling of learning and actual learning, but not when offloading stays productive. Constructivist accounts explain dialogue, yet conversation with a fluent system does not guarantee justified learning: it can occupy the position of a knowledgeable other without being trustworthy. Connectivism explains distribution, but network nodes can now generate content, infer intent, and sequence a learner's activity.

The four mechanisms of Agentivism

Delegated Learner Agency describes how influence over planning, drafting, explanation, evaluation, and revision is distributed between learner and system. Delegation preserves learning when the learner still frames the problem, sets criteria, and decides what counts as acceptable reasoning, and it undermines learning when those functions migrate to the system and the learner becomes mainly a selector of fluent output. Epistemic monitoring and verification covers checking output for truthfulness, relevance, adequacy, provenance, and fit to task demands, keeping the learner engaged instead of mistaking fluency for understanding. Reconstructive internalization is when an accepted output becomes the learner's own capability: they can say why it is appropriate, when it would fail, and how to reproduce the reasoning. Transfer under reduced support is the criterion, and it does not require AI independence.

Propositions and design conditions

As a mid-range theory, Agentivism generates falsifiable claims rather than verdicts on AI. Process measures taken during interaction, such as prompt trajectories, revision sequences, evidence-checking moves, and explanation quality, should predict later learning better than final product quality. Design and institutional conditions matter as conditions rather than substitutes: interface design, rules, accessibility provisions, and norms for acceptable assistance decide whether verification and reconstruction are invited or whether passive uptake is the path of least resistance, most of all where learners differ in resources and prior knowledge. Repeated reliance on the same systems can also narrow the diversity of sources, framings, and questions learners meet, making epistemic diversity a learning concern as much as a fairness one.

What this means for practice

• Preserve the learner's share of the work. Keep problem framing, criteria, and justification with students; an AI that delivers answers removes the operations that build independence.
• Require verification before uptake. Build source checks or a short justification into AI-supported tasks; slowing acceptance down reduces over-reliance even when it feels less convenient.
• Ask for reconstruction, not just submission. Have learners re-explain the answer, adapt it to a new case, or say when it would fail, and assess that work.
• Assess later, with less AI. Use delayed explanation and novel problems under reduced support as evidence of learning, and treat rising productivity during AI use as no evidence of learning.
• Make delegation expectations explicit. State what assistance is acceptable for a task and pair it with accessible support; the same arrangement is not equally agentic for every learner.

Limitations

• No formal model. As a mid-range conceptual theory, it names constructs and proposes tests, but does not specify a universal sequence or reduce learning with AI to one principle.
• Constructs resist simple measurement. Delegated agency, verification, reconstruction, and transfer are unlikely to be captured by one method; research will need trace data, process measures, and delayed assessments together.
• Evidence is short-term. Much current evidence comes from brief tasks while the strongest claims concern durable capability, and productive delegation likely differs across tasks and learners.

Connected Concepts

• Agentic AI
• Learner Agency
• Theories and Frameworks
• Theory Development in AI in Education
• Human AI Collaboration
• Cognitive Offloading
• Learning Theories
• Metacognition
• Transfer of Learning
• Trust Calibration
• Generative AI

Connected Articles

• Building AI Companions that Prioritize Learning over Performance — Building AI Companions that Prioritize Learning over Performance
• Distinguishing performance gains from learning when using generative AI — Distinguishing performance gains from learning when using generative AI
• Toward Convergence in Student-LLM Interactions: A Rapid Scoping Review and Taxonomy for Learning-Oriented Use — Toward Convergence in Student-LLM Interactions

Citation

Yan, L., & Gašević, D. (2026). Agentivism: a learning theory for the age of artificial intelligence. Computers and Education: Artificial Intelligence.