Can AI Really Teach Us to Learn? Exploring the New Theory of Agentivism

Evolution of Learning Conditions

Over the past century educational theory has progressed from behaviorism (stimulus‑response), through mid‑century cognitivism (the brain as a computer), to constructivism (knowledge built through interaction with real environments), and finally to connectivism (knowing where to find information is more valuable than memorising it). The rapid diffusion of generative AI and autonomous agents now creates a new learning condition that breaks the previous paradigm: task performance no longer guarantees genuine understanding.

Performance Does Not Equal Learning

Traditional tools—books, calculators, search engines—assist humans but still require the learner to synthesise, reason and write. Modern generative AI can perform these core cognitive tasks instantly, producing polished reports or code. When AI handles the heavy lifting, learners risk losing the ability to reason independently, diagnose problems, or transfer knowledge to novel contexts, much like a driver who cannot steer when autopilot fails.

Agentivism: Four Pillars for AI‑augmented Learning

Agentivism, proposed by a Tsinghua University research team, defines learning in the AI era as a dynamic interaction with intelligent systems that must be guided by deliberate strategies. The framework consists of four inter‑dependent pillars.

Selective Delegation Identify tasks that can be fully automated (e.g., data collection, formatting, routine calculations) and separate them from tasks that require human insight (e.g., hypothesis generation, ethical judgement, deep conceptual reasoning). Explicitly document the boundary so that core reasoning remains under the learner’s control.

Cognitive Monitoring and Verification Treat every AI output as a hypothesis that must be audited. Verify data sources, cross‑check facts, and reconstruct the reasoning chain presented by the model. This continuous verification keeps the mind actively engaged and guards against hallucinations.

Reconstructive Internalisation After receiving AI‑generated content, de‑construct it: re‑phrase in one’s own words, map the ideas onto personal knowledge structures (e.g., mind‑maps or concept graphs), and, where possible, reproduce the solution without digital aid. This “re‑assembly” transforms external information into durable internal knowledge.

Deliberate Practice Without AI Regularly schedule sessions where the learner solves problems or writes explanations without any AI assistance. Examples include handwritten derivations of algorithms, oral presentations of a research summary, or manual coding of a core routine. Such practice preserves and strengthens independent problem‑solving abilities.

Practical Guidelines for Applying the Pillars

Task‑Boundary Mapping : Create a checklist that labels each step of a workflow as either delegate‑able or human‑critical . For a research report, delegate literature search and formatting, but retain hypothesis formulation and critical discussion.

Verification Workflow : Use a three‑step audit—(1) source validation, (2) logical consistency check, (3) independent re‑derivation. Record findings in a log to build a habit of skeptical scrutiny.

Internalisation Exercises : After an AI‑generated answer, close the screen and write a summary from memory. Then compare with the original to spot gaps. Optionally, draw a diagram that captures the underlying process.

AI‑Free Challenge Sessions : Allocate fixed time blocks (e.g., 30 minutes per day) where no AI tools are permitted. Choose tasks that directly exercise the skills you wish to retain, such as solving a physics problem on paper or coding a sorting algorithm from scratch.

By consciously delegating low‑value work, rigorously monitoring AI outputs, reconstructing knowledge internally, and maintaining regular AI‑free practice, learners can convert fleeting AI assistance into lasting human capability.

Reference: https://arxiv.org/pdf/2604.07813