Critical AI Tutors: Empower or Enslave?

Created: 2026-07-29 | Tags: intelligent-tutoringcritical-thinkingover-reliance
Critical AI Tutors: Empower or Enslave? — A position paper presented at the AIED 2025 workshop that issues a stark warning: unchecked use of AI tutors risks creating a generation of cognitively atrophied learners who have traded genuine understanding for the illusion of competence. Drawing on cognitive science and pedagogical theory, the authors argue that AI tutors without intentional guardrails lead to cognitive atrophy, loss of agency, emotional risks, and serious ethical concerns around privacy and academic integrity. The paper advocates for critically informed, transparent AI use that empowers rather than diminishes the learner, and calls for student perspectives to be centered in the design and deployment of AI tutoring systems.

Authors: Favero et al. · Workshop: AIED 2025 · arXiv: 2507.06878

Key Findings

This position paper occupies a critical counterpoint in the intelligent-tutoring-systems literature. While much of the field focuses on optimizing AI tutor performance and learning gains, the authors argue that effectiveness metrics alone are dangerously insufficient — what matters equally are the cognitive and developmental costs that AI tutors may impose on learners.

Cognitive atrophy and the shortcut problem. The central argument draws on cognitive-load-theory and the well-established finding that effortful cognitive processing is essential for durable learning. When intelligent-tutoring systems provide immediate, high-quality answers and solutions, they effectively short-circuit the very cognitive processes — struggling with problems, retrieving from memory, constructing explanations — that produce deep understanding. This phenomenon, which the authors term cognitive atrophy, mirrors concerns raised in the efficiency-gain-illusion-ai-overreliance and cognitive-offloading literatures: learners may feel more productive while learning less.

Loss of agency and dependency. Beyond cognitive effects, the paper identifies a broader threat to learner agency. Prolonged reliance on AI tutors can produce over-reliance — a state where students lose confidence in their own reasoning abilities and become dependent on AI assistance even for tasks they could complete independently. This dependency dynamic connects to the correct-answer-trap-ai-tutor problem and the finding that ai-assistance-reduces-persistence — students give up more quickly when AI help is available.

Emotional and well-being risks. The authors highlight underexplored emotional dimensions: AI tutor interactions can erode self-efficacy when students compare themselves unfavorably to flawless AI outputs, contribute to ai-fatigue-academic-contexts, and diminish the relational aspects of learning that teacher-ai-coagency frameworks seek to preserve.

Ethical concerns. The paper catalogs significant ethical risks including academic-integrity erosion, questionable privacy practices in educational AI systems, and the broader societal implications of ai-making-us-stupid — a provocative framing that challenges the edtech optimism narrative.

Implications

The paper's most important contribution is its call for critical AI literacy as a prerequisite for AI tutor deployment. Rather than banning AI tutors, the authors argue for what they call "critically informed use" — a framework where students, educators, and institutions actively interrogate AI tools rather than passively accepting them. This aligns with the critical-thinking tradition in education and extends it to the domain of ai-literacy-power-knowledge.

For system designers, the paper implies that ai-tutor-safety-harms frameworks must expand beyond immediate harm prevention to include long-term developmental impacts. An AI tutor that never harms a student in the moment but gradually erodes their cognitive independence should be considered unsafe. This reframes pedagogical-safety as encompassing not just what the tutor does but what kind of learner it produces.

The authors' emphasis on student voice — centering learner perspectives in design decisions — connects to the student-experience literature and the growing recognition that generative-ai-guardrails-harm-learning must be co-designed with the very populations they aim to protect.

For educators and policymakers, this paper provides intellectual ammunition for resisting uncritical AI adoption. It suggests that regulation of educational AI should consider not just data privacy and bias but also cognitive and developmental outcomes — a position that resonates with calls in the ai-governance-education community for holistic evaluation frameworks.

Related Pages