Abstract
This study investigates the role of code-switching (CSW)—the alternation between a native language and a target language—within AI-mediated English as a Foreign Language (EFL) speaking practice. By analyzing interactions between Korean learners and a GPT-4o-based tutor alongside expert teacher evaluations, the research identifies a significant "pedagogical gap." While Large Language Models (LLMs) excel at providing accurate translations, they often lack the nuanced scaffolding and emotional intelligence that human teachers employ to transform CSW into meaningful learning moments. The findings suggest that CSW serves as a vital tool for reducing anxiety and expressing cultural identity, though current AI tutors remain overly reliant on direct translation rather than conversational recasts.
Introduction
In traditional EFL environments, "English-only" policies often induce high anxiety, causing learners to freeze when they encounter linguistic barriers. Code-switching (CSW) is a natural strategy used by learners to bridge these gaps. However, most AI speaking tutors are designed to be English-centric, either ignoring first-language (L1) input or responding with robotic translations. This paper seeks to understand the functional purpose of CSW in human-AI interaction and how AI tutors can be designed to mirror the pedagogical finesse of expert human teachers.
Methodology
The researchers conducted a two-part study to bridge the gap between learner behavior and expert pedagogy:
- Learner Study: 20 Korean EFL learners engaged with a custom AI speaking platform powered by GPT-4o over six weeks. The system utilized four feedback strategies: explicit feedback, implicit feedback, clarification, or ignoring the CSW. A total of 325 dialogues (1,032 CSW instances) were analyzed to categorize the intent behind each switch.
- Teacher Study: 9 experienced English teachers reviewed the learner-AI logs. They provided spontaneous responses to learner CSW, "reconstructed" ideal pedagogical responses, and participated in interviews to explain their reasoning.
The team employed thematic analysis to build a taxonomy of CSW functions and used Sankey diagrams to visualize "learner uptake"—the degree to which students successfully integrated suggested English terms into their subsequent speech.
Functional Taxonomy of Code-Switching
The study revealed that CSW is far more than a simple "lexical gap" solution. While approximately 75% of switches were used to replace unknown words, learners also utilized their native language to express complex cultural identities and emotional nuances that they felt were lost in translation.
As shown in the data, learners frequently switched to Korean when discussing specific cultural items (e.g., "kimchi jjigae") or when attempting to convey specific stances and emotions. Interestingly, over the six-week period, learners shifted from using CSW for basic lexical compensation to using it for expressive functions, indicating increased comfort and a desire for deeper communication with the AI.
Key Results and Pedagogical Divergence
The analysis highlighted a stark difference between how LLMs and human teachers handle CSW:
- The Translation Trap: LLMs relied on direct translation in roughly 40% of their responses. In contrast, human teachers preferred recasts—implicitly rephrasing the learner's L1 into English within a natural sentence—to maintain flow and reduce anxiety.
- Learner Uptake: Despite the LLM's limitations, the system achieved a 68.2% successful uptake rate, where learners adopted the suggested English expressions in their next turn.
- Anxiety and Willingness: Learners reported that the ability to code-switch significantly reduced speaking anxiety (5.72/7) and increased their willingness to communicate (5.50/7).
The visualization of learner uptake demonstrates that while explicit feedback often leads to successful adoption of new vocabulary, implicit strategies like recasting are vital for maintaining the conversational thread, a balance human teachers strike more effectively than current LLM implementations.
Discussion and Future Work
The research underscores that CSW is an "informed resistance" tool; learners sometimes rejected AI translations to preserve the cultural resonance of their native terms. Future AI tutor development should focus on:
- Intent Classification: Automatically detecting whether a code-switch is a request for help or a deliberate cultural choice.
- Mixed-Initiative Scaffolding: Determining the optimal moment to provide a correction versus allowing the conversation to proceed.
- Persona Design: Exploring how different AI personas (e.g., a peer vs. a formal instructor) influence a learner's willingness to negotiate meaning through CSW.
While limited to the Korean-English context, this study provides a foundational framework for designing more empathetic and pedagogically sound AI language tutors.
Summary updated