Generativism: Toward a Learning Theory for the Age of Generative Artificial Intelligence

Synthesis: Li and Zheng argue that the four dominant learning theories (behaviorism, cognitivism, constructivism, and connectivism) each rest on an assumption that generative AI breaks, and propose Generativism: learning as the deliberate co-construction of knowledge through iterative human-AI interaction. The framework names four constructs, epistemic partnership, distributed agency, generative literacy, and adaptive metacognition, and derives assessment indicators for each. It is a position paper that synthesizes existing evidence rather than testing a model of its own.

Key Findings

1. Observable output no longer certifies learning. Behaviorism equates learning with a change in observable behavior, but a learner using an LLM can produce essays, analyses, and code indistinguishable from competent work. What is left is performance, and the dissociation is measurable: in a randomized field experiment with roughly a thousand high school students, access to a standard GPT-4 interface raised practice-problem performance by 48 percent while the unassisted exam score fell 17 percent below students who had practiced without AI. A tutored version removed the harm but produced no lasting advantage (Distinguishing performance gains from learning when using generative AI).
2. Cognitive labor is redistributed, not merely accelerated. Learners now externalize synthesis, argumentation, analysis, and production, which under cognitivist theory would count as evidence of learning if a person did them. In a scientific inquiry task, students working with an LLM reported significantly lower mental effort and produced reasoning of significantly lower quality than students using conventional web search.
3. Effortful meaning-making is what gets skipped. Constructivism and generative learning theory both require learners to select, organize, and integrate material themselves, and the ICAP framework predicts stronger learning as engagement moves from passive through active and constructive to interactive. Students instead often adopt AI-generated solutions before doing that work, and fluent AI output also reduces the peer dialogue that social constructivists treat as the driver of co-construction.
4. Connectivism's "know-where" is subsumed by co-generation. Connectivism was formulated when non-human nodes stored and transmitted information; generative AI produces novel, query-responsive output, so the scarce skill shifts from navigating a network to evaluating, guiding, and integrating generated content. The older skill was not well developed either: national samples of high school students could not reliably separate credible from untrustworthy online content, and even educated adults judged unfamiliar sites poorly without professional fact-checking strategies.
5. Human-AI teams are not automatically better. A meta-analysis of 106 experiments found human-AI combinations performing significantly worse on average than the best of humans or AI alone, with losses in decision-making tasks and gains in content creation. Whether AI helps depends on the quality of the interaction rather than on access to the tool.
6. The framework converts into assessable indicators. For each construct the authors propose what evidence would show it: whether learners question, verify, and redirect AI output; whether they can explain what they delegated and why; whether they revise prompts and integrate output with prior knowledge; and whether they detect confusion or overreliance and adjust.

Where the four learning theories break down

Each critique targets one assumption. Behaviorism fails on the behavioral equivalence problem: identical output, different learning. Cognitivism fails because cognitive operations previously exclusive to human minds are now performed outside them, a redistribution the theory does not model, which shows up as reduced self-regulated learning activity and as a negative association between frequent AI use and critical thinking that cognitive offloading mediates (AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking). Constructivism fails because ready-made, already-organized output invites consumption instead of construction. Connectivism fails because generation replaces retrieval as the hard part.

A shared mechanism runs through the cognitivist and constructivist critiques: effort moved out of working memory takes schema-building with it. AI output arrives pre-organized, so learners retain it without reorganizing it, and the ease of processing that accompanies fluent text is exactly the cue people misread as understanding (Metacognition).

The four constructs

Epistemic partnership treats AI as an active participant in a cognitive system rather than an external artifact. Distributed cognition's charts and logs stored and displayed information; a model produces explanations, adapts to the interaction, and shapes what the learner notices. Partnership quality, not the presence of a tool, determines outcomes, and learners approach AI with different epistemic stances that decide whether it functions as a shortcut to answers or as a partner in inquiry (Distributed Cognition).

Distributed agency describes how cognitive work is allocated between learner and system. The learner sets goals, frames problems, evaluates responses, and decides when output is inadequate, while the system makes some possibilities more visible than others. Agency is negotiated rather than transferred, and it is vulnerable upstream: when a system steers which options feel available, learners can adopt intentions they did not fully author.

Generative literacy is critical literacy for prompting, evaluating, refining, and integrating generated knowledge, condensed here into four capacities: prompt competence, evaluative judgment, epistemic vigilance, and integrative synthesis (Evaluative Judgment). The authors treat it as a condition for learning with AI rather than an optional technical add-on, because it is what turns output into understanding.

Adaptive metacognition extends self-regulated learning to a setting where part of the cognition is external. The established models assume the learner regulates only their own processes; here learners must also monitor a system that is performing cognitive operations, across decisions made before, during, and after use (Self-Regulated Learning, Cognitive Offloading).

How the framework treats agency and responsibility

Authoring a knowledge artifact with AI may distribute authorship, but the framework treats responsibility as non-delegable: the human partner remains accountable for the accuracy, appropriateness, and ethical use of AI-supported work. That position has direct consequences for assessment and for academic integrity policy, since it makes the learner's ability to explain and defend the division of labor part of what is being evaluated rather than a disclosure requirement bolted on afterward (Academic Integrity, Human AI Collaboration).

What this means for practice

• Instructors. Separate contexts where AI collaboration is the target competence from contexts where independent performance is the outcome, and assess each accordingly: in the first, examine how learners question, verify, and redirect AI output and what they understand afterward; in the second, restrict AI use and use unaided explanation or transfer tasks.
• Instructors. Structure tasks so the model has to be interrogated rather than consulted: require an attempt before AI use, ask learners to diagnose and revise AI-generated work, or have them reconcile competing AI perspectives. Preserve the difficulty that carries the learning.
• Instructors. Make the division of cognitive labor visible and treat it as assessable evidence. The authors' indicators can be gathered from process logs, annotated prompts, brief reflections, oral defenses, and comparisons of AI drafts against learner revisions.
• Instructors. Do not assume learners allocate agency well by default. Algorithm appreciation and aversion both appear in the literature, so prompt explicit decisions about when to rely on AI, when to question it, and when to keep the judgment.
• Researchers. Treat the four constructs as testable rather than settled: examine whether they are empirically separable, whether each contributes unique variance to learning outcomes in human-AI settings, and whether phenomena fall outside their joint scope. The framework's own authors name this as the next step.

Limitations

• Generativism is a conceptual framework that has not been empirically tested. The constructs are derived from a synthesis of existing evidence and require systematic validation, including the question of whether the four are separable at all.
• The authors state that the framework does not fully address access and equity, specifically how differential access to generative AI tools across socioeconomic contexts affects learning dynamics.
• The rapid pace of AI development means the specific technologies discussed may evolve in ways the framework cannot anticipate.
• It is a position paper with no new data, so it can propose constructs and indicators but cannot establish effects on learning.

Citation

Li, S., & Zheng, J. (2026). Generativism: Toward a Learning Theory for the Age of Generative Artificial Intelligence. arXiv preprint.