Generative AI as a Mediational Agent: Rethinking Learning in Sociocultural Theory

Synthesis: Warschauer, Tate, and Ritchie (2026) argue that Generative AI challenges a foundational distinction in sociocultural theories of learning: the separation between mediational means and social interaction. Traditional tools such as language, writing, and educational technologies mediate human activity while development arises through social participation; generative AI both mediates activity and generates context-sensitive contributions that shape ongoing interaction. Because treating AI as a mere tool underestimates its interactional influence, while treating it as a collaborator wrongly attributes intentionality, accountability, and community membership, the essay proposes the mediational agent — a responsive but non-accountable system that contributes explanations, critiques, questions, and suggestions while occupying no position in a community of practice. The authors derive five human-first habits — primacy of human cognition, purposeful engagement, supervisory agency, epistemic vigilance, and reflective self-AI Regulation in Education — meant to preserve learners' judgment, Learner Agency, and responsibility. The framing offers a principled middle path for AI education, shifting attention from technological capability to forms of participation.

Key Findings

1. Primacy of human cognition. Learners should develop ideas, interpretations, or problem framings before turning to AI, building the intellectual foundation needed to evaluate later contributions rather than adopt them.
2. Purposeful engagement. AI use is directed by human intentions: learners enlist the system strategically to extend, challenge, or refine their thinking rather than to replace it.
3. Supervisory agency. Because the system generates contributions without responsibility for consequences, learners remain accountable for every decision about whether and how those contributions enter their work.
4. Epistemic vigilance. Fluent language and confident explanations must never be confused with truth or disciplinary validity; learners interrogate claims, corroborate evidence, and judge reasoning by disciplinary standards.
5. Reflective self-AI Regulation in Education. Since AI can obscure the boundary between human and machine contributions, learners consciously monitor what they genuinely learned and how the system shaped their thinking.

Why "Tool" and "Collaborator" Both Fail

The essay starts from a conceptual gap. Sociocultural theory has long distinguished mediational means — language, writing, technologies — from social interaction with teachers, peers, and communities. Two dominant framings of generative AI are each insufficient. Treating AI as a tool underestimates its interactional influence: an Large Language Models (LLMs) generates contingent, context-sensitive turns — explaining, critiquing, questioning, suggesting, revising — and so actively shapes the trajectory of activity rather than sitting inertly within it. Treating AI as a collaborator instead risks attributing intentionality, accountability, and community membership the system lacks; a collaborator is answerable to shared norms, occupies a position in a community of practice, and shares a history. This mirrors the knowledge base's wider debate over AI as neutral instrument versus quasi-partner in human–AI collaboration, and the paper offers a theoretically rigorous middle path.

The Mediational Agent

The authors propose mediational agent as a distinct category: a system that simultaneously mediates activity and generates context-sensitive contributions shaping unfolding interaction, without intentionality, social membership, or accountability. Like a tool, generative AI reorganizes how learners write, reason, solve problems, and communicate; unlike one, it contributes novel utterances that become part of the activity itself. Unlike a human participant, it holds no position in a community of practice, pursues no goals, and bears no responsibility for consequences. The concept therefore names a hybrid form of participation rather than a midpoint on a continuum between tools and people — the novelty lies not in intelligence but in participation. Because activity systems containing mediational agents are organized differently, learners must continually decide whether to accept, reject, revise, interrogate, or redirect AI-generated contributions, making those judgments part of the learning activity itself.

The Developmental Paradox

Generative AI is exceptionally effective at improving immediate performance — fluent prose, working code, plausible explanations — yet learning depends on process: generating ideas, wrestling with uncertainty, evaluating alternatives, revising understanding. Systems designed to reduce cognitive effort therefore create a developmental paradox, since the same capabilities that improve immediate performance can reduce opportunities for intellectual growth when they replace rather than support learners' own thinking. This is the paper's link to the Reducing AI Misuse concern about performance–learning gaps and to productive struggle. The five habits respond by regulating the relationship between one's own thinking and the system's contributions rather than by optimizing tool use. They are dispositions, not a fixed Pedagogies and Teaching Strategies: supervisory agency in historical inquiry differs from software engineering, and epistemic vigilance takes different forms in scientific modeling than in literary interpretation.

Research agenda

Viewing AI as a mediational agent redirects researchers from whether AI improves outcomes toward how patterns of participation change. Scaffolding traditionally meant support from teachers, peers, or designed environments; mediational agents generate support dynamically without the social relationships that gave scaffolding its developmental significance. Distributed cognition asks how cognitive work is shared across people and artifacts; generative AI raises the sharper question of how cognitive responsibility should be distributed when artifacts contribute to reasoning itself. The framing supplies vocabulary for debates that otherwise oscillate between techno-optimism and alarmism, complementing work on Framing AI Use for Students, AI Literacy, Critical Thinking, Ethics, and Philosophy of AI in Education, while cautioning against non-accountable persuasive output such as AI Sycophancy that erodes the human judgment these habits protect.

What this means for practice

• Instructors. Require learners to develop their own ideas, interpretations, or problem framings before consulting generative AI, so that they hold the intellectual foundation needed to evaluate the system's contributions rather than adopt them.
• Instructors. Assess the judgment, not only the output: make accepting, rejecting, revising, interrogating, or redirecting an AI-generated contribution an explicit and graded part of the task, since the mediational-agent framing treats those decisions as part of the learning activity itself.
• Instructors. Defend process over polish: because generative AI improves immediate performance while learning depends on generating ideas, wrestling with uncertainty, evaluating alternatives, and revising, design tasks in which the system supports that struggle instead of replacing it.
• Instructors. Ask learners to account for what they genuinely learned and how the system shaped their thinking, since reflective self-AI Regulation in Education is proposed precisely because AI obscures the boundary between human and machine contributions.
• Researchers. Study patterns of participation and the negotiation of authority, responsibility, and judgment during interaction rather than treating AI use as a binary condition or measuring learning only through performance gains.

Limitations

• This is a conceptual essay (an EdArXiv preprint), not an empirical study: it presents no sample, intervention, or measured outcome, so it can neither show that the mediational-agent framing improves learning nor establish that learners actually enact the proposed habits.
• The five human-first habits — primacy of human cognition, purposeful engagement, supervisory agency, epistemic vigilance, and reflective self-AI Regulation in Education — are derived analytically from sociocultural theory rather than tested, and the authors state that supervisory agency and epistemic vigilance take different forms across disciplines, so the set is not a validated curriculum.
• The authors place the concept's own success criterion in the future tense — its value "will depend on whether the concept helps the field to notice phenomena that existing categories leave unexplained" — and the essay offers no operationalization or measurement for that criterion.
• The argument treats generative AI as a single undifferentiated category, without distinguishing systems by capability, interface, or deployment, so its conceptual claims cannot be bounded to particular tools or settings.

Citation

Warschauer, M., Tate, T., & Ritchie, D. (2026). Generative AI as a Mediational Agent: Rethinking Learning in Sociocultural Theory. EdArXiv preprint.