Visual Scaffolding in AI-Assisted Second Language Learning : Designing and Evaluating Image-Based Conversation Prompts for Intermediate English Learner

Sammanfattning

Cognitive Load Theory suggests that intermediate second language (L2) learners may face a cognitive bottleneck in AI-assisted conversation practice, simultaneously managing content generation, social anxiety, and linguistic formulation under finite attentional resources. Motivated by this account—and by the long-standing classroom use of pictures as speaking prompts—this thesis proposes and tests visual scaffolding, here operationalised as image-based conversation prompts displayed during interaction with an AI voice tutor, and examines whether such prompts in fact reduce that load. Study 1 conducts a within-subjects experiment (n=22) with adult English learners, comparing a visual scaffolding condition against a free-talk condition and measuring perceived cognitive load via NASA-TLX and linguistic performance via the Complexity, Accuracy, and Fluency (CAF) framework. Visual scaffolding produced a 21% increase in speech volume and a significant narrowing of lexical diversity but did not reduce perceived cognitive load at the sample level; a performance-based subgroup analysis revealed a significant Frustration increase (+11.8 points, p = .045) for lower-performance learners that was masked in the aggregate by neutral effects in higher-performance learners — an expertise-reversal pattern. The findings reframe visual scaffolding as a content scaffold (shaping what learners say and how much) rather than a process scaffold (reducing the cognitive cost of saying it). Study 2 employs a Research through Design (RtD) approach, conducting co-design workshops with ESL teachers to derive design principles for stage-adaptive visual scaffolding—from concrete realistic images to abstract diagrams—supporting learners' progression toward autonomous conversation. Together, the two studies contribute empirical evidence on the expertise-dependent effects of visual scaffolding in AI-mediated L2 practice and three sets of design principles — for visual prompts, verbal instructions, and progression — that specify how an adaptive AI tutor should support intermediate learners at risk of the 'intermediate plateau.'