From Algorithmic Scaffolding to Cognitive Withdrawal: A Three-Phase Classroom Experiment on AI Regulation in EFL Debates

This study examines how varying levels of AI access reshape oral performance, linguistic choice, and critical engagement in EFL academic debates. Thirty undergraduate students participated in a three-week structured debate intervention under escalating regulation: (1) unrestricted phone use, (2) AI reliance discouraged, and (3) complete device prohibition. Debate rounds were systematically coded for language use (English dominance), response latency (immediate, hesitant, delayed), and response quality (critical independence vs. AI-script dependence). Results reveal a paradoxical trajectory. In Week 1, unrestricted AI access produced fluent but script-dependent English delivery, characterized by delayed responses linked to live translation and reading behaviors. In Week 2, verbal warnings did not significantly reduce AI-script dependence. In Week 3, device prohibition eliminated scripted reading but led to increased hesitation and code-switching, with many immediate responses occurring in the native language rather than English. Statistical comparison across weeks (χ² tests) indicates significant shifts in response immediacy and language dominance (p < .05). The findings introduce the construct of algorithmic scaffolding withdrawal, suggesting that abrupt removal of AI support exposes underlying linguistic insecurity rather than restoring spontaneous English fluency. The study challenges simplistic policy assumptions that banning AI enhances authenticity and argues for graduated scaffolding models that rebuild internalized language competence while preserving critical reasoning capacity.