1 Introduction

This chapter describes the challenges teachers face in addressing diverse student backgrounds, abilities, and interests in language classrooms, with particular attention to beginner-level Chinese L2 learners who struggle with reading and writing due to the language's unique characteristics. It introduces differentiated instruction as a pedagogical approach that tailors learning processes to individual needs, while acknowledging practical constraints like limited time and resources. The emergence of ChatGPT and generative AI is presented as a potential solution, with Custom GPTs offering enhanced capabilities for domain-specific content generation. The study outlines its research framework: examining differentiated instruction concepts and implementation strategies, applying the component skills approach to assess student readiness, conducting comprehensive evaluations of learning styles and interests, and exploring ChatGPT applications for differentiated instruction. The methodology involves comparing standard ChatGPT with customized versions through GPT builders, analyzing their characteristics, strengths, and limitations. Generated instructional materials will be evaluated by ten professional Chinese L2 teachers, with t-test analysis used to statistically compare content quality between the two approaches.

2 Related Work

This chapter establishes the theoretical and practical foundations for examining ChatGPT's potential in designing differentiated instructional activities for Chinese L2 reading and writing classrooms. Differentiated instruction is presented as both a teaching philosophy and classroom practice that requires teachers to understand students' strengths, needs, differences, and similarities. Grounded in Tomlinson's framework, effective differentiated instruction centers on three key student characteristics: readiness (alignment between tasks and current skill levels), interest (student curiosity and passion tied to choice), and learning profile (preferred learning methods encompassing styles, intelligence preferences, gender tendencies, and cultural backgrounds). These characteristics enable teachers to differentiate three core curriculum components: content, process, and product.

The component skills approach to reading provides the cognitive and linguistic foundation for understanding L2 Chinese reading development. This approach identifies multiple complementary abilities including phonological awareness, orthographic knowledge, morphological awareness, and syntactic awareness. Phonological awareness, particularly through Pinyin recoding, contributes to orthographic and semantic learning despite having moderate rather than strong effects on Chinese reading compared to alphabetic languages. Orthographic knowledge emphasizes the critical importance of memorizing character visual forms—strokes, radicals, and overall structure—with semantic and phonetic radical mastery significantly enhancing recognition. Morphological awareness facilitates vocabulary acquisition and directly impacts reading comprehension through mechanisms like lexical inference. Syntactic awareness, especially word order knowledge, uniquely contributes to passage-level comprehension even after controlling for multiple variables, while word segmentation skills address Chinese writing system characteristics where boundaries remain unmarked.

Regarding generative AI applications, ChatGPT has attracted substantial educational attention since November 2022, though research remains predominantly theoretical with limited empirical investigation in international Chinese education contexts. The technology demonstrates potential as a personal tutor through four functional aspects: collecting and inputting student information regarding knowledge levels, interests, and learning styles; generating differentiated resources aligned with readiness, interests, and learning profiles; enabling teacher review and refinement of AI-generated materials to ensure contextual appropriateness and curriculum alignment; and supporting ongoing updates through progress monitoring and adaptive strategy generation. Custom GPT features—including preloaded context, task-specific configurations, integrated knowledge databases, and definable instructions with stylistic preferences—offer enhanced effectiveness by ensuring clearer outputs and reducing complex prompting requirements, potentially enabling more effective classroom implementation.

3 Methodology

This chapter outlines the methodological framework for comparing two ChatGPT-based approaches to differentiated instruction in Chinese L2 classrooms. The study systematically implements and evaluates both standard ChatGPT-4o and a Customized GPT designed specifically for differentiated instruction, examining their relative effectiveness across multiple dimensions.

The research design involves generating five sets of instructional materials from each model, with each set containing parallel materials targeting identical textbook content and learning objectives. These materials address five differentiation strategies: content based on readiness, process based on readiness, product based on readiness, interests and learning profile combined, and content based on interests with process based on learning profile. Ten professional Chinese L2 teachers independently evaluate all materials using three 100-point criteria—differentiated instructional effectiveness, learning objective alignment, and instructional feasibility—with weighted averages calculated and t-tests applied to identify statistically significant differences between models.

Student perspectives are incorporated through a 10-item Likert-scale satisfaction survey and three open-ended questions administered at semester's end, providing comprehensive feedback on material effectiveness. The participant pool comprises five first-year UCD students with diverse proficiency levels and cultural backgrounds, all in their second semester of Chinese as a minor subject. Learning content draws from 15 lessons of the HSK 2 Standard Course textbook, with texts randomly selected for comparative analysis.

Student readiness is assessed through a four-component test measuring phonological awareness (20%), orthographic knowledge (30%), morphological awareness (25%), and syntactic awareness (25%), yielding three proficiency groupings: beginning (1 student), intermediate (2 students), and high (2 students). This classification enables readiness-based differentiation. Supplementary data collection includes demographic information and two questionnaires examining student interests and learning profiles, with all compiled information formatted for model input to support personalized instructional design.

4 Customizing GPTs and Comparison Between ChatGPT-4o and Custom GPT

This chapter explains the process of customizing GPTs for differentiated instruction and compares ChatGPT-4o with Custom GPT in terms of performance and output quality. The GPT Builder enables users to create specialized GPTs without coding knowledge by uploading files for semantic retrieval, configuring instructions, and integrating third-party APIs. For this research, four file types were utilized: student information, differentiated instruction theory and strategies, teaching content, and supplementary Chinese language materials. The customization process involves three steps: creating a tailored GPT with specific name, description and instructions; uploading relevant files to enhance the knowledge base; and requesting differentiated activities based on student readiness, interests, or learning profiles.

The comparison between ChatGPT-4o and Custom GPT focused on time efficiency and prompt requirements when designing reading activities for different proficiency levels targeting the grammar point 最. Both models required approximately 4-5 minutes from prompt input to output generation. However, Custom GPT demonstrated superior efficiency through more concise prompts and reduced repetitive inputs, with this advantage becoming more pronounced through continued use.

Professional Chinese L2 teachers evaluated the generated materials using three criteria: effectiveness of differentiated instruction, alignment with learning objectives, and instructional feasibility, with scores ranging from 0-100 across six performance levels from "Not Applicable" to "Excellent."

5 Data Analysis and Result

This chapter compared the performance of ChatGPT-4o and Custom GPT in generating differentiated instructional materials. Custom GPT demonstrated clear advantages through its dedicated knowledge base, enabling continuous accumulation of instructional resources, reduced repetitive workload, and efficient integration of retrieval capabilities with stored knowledge to produce higher-quality content. Statistical analysis confirmed these advantages: an independent-samples t-test (t=-3.39, p=0.0044) revealed that Custom GPT-generated materials received significantly higher scores across three evaluation criteria—effectiveness of differentiated instruction, alignment with learning objectives, and instructional feasibility. The Custom GPT materials showed higher median scores and narrower distribution ranges, indicating greater consistency, while ChatGPT-4o exhibited wider variability in quality. Student feedback toward differentiated instruction was generally positive, with an average score of 44.6 out of 50. Students particularly valued the approach for enhancing content understanding, increasing enjoyment through interest-based tasks, and providing clear instructions. However, some students noted that tasks did not always match their proficiency levels adequately. Students identified reading comprehension and writing tasks as most beneficial, reporting improvements in comprehension, vocabulary retention, and character recognition. They expressed desire for more frequent and personalized differentiated instruction.

6 Conclusion and Future Work

This chapter discusses the effectiveness of ChatGPT in designing differentiated instructional activities for Chinese L2 reading and writing classrooms. Custom GPT consistently outperformed ChatGPT-4o in generating high-quality differentiated instructional materials that better aligned with educational objectives and were more easily implemented by teachers, as confirmed by t-test analysis. Students generally responded positively to differentiated instruction, appreciating its helpfulness for understanding content, engaging in interest-based tasks, and following clear instructions, though some raised concerns about proficiency-level alignment and requested more personalized and frequent instruction. The study acknowledges several limitations: it did not assess student performance before and after implementation, included only a small sample, and focused primarily on comparing model attributes and teacher evaluations rather than measuring learning outcomes. Future research should incorporate more rigorous pre- and post-intervention assessments, explore ChatGPT's potential for developing listening and speaking skills through speech input and dialogue features, and investigate its impact on intermediate and advanced proficiency levels. Hallucination remains a significant concern requiring further investigation, with three key issues identified in this regard.

1 Higher Hallucination Tendency in ChatGPT-4o

Both ChatGPT-4o and the other model exhibited hallucinations, though ChatGPT-4o demonstrated a notably stronger tendency toward this issue. This higher hallucination rate necessitated additional time investment for fine-tuning the model's outputs and demanded more meticulous prompt engineering to achieve reliable results.

2 Exacerbated Hallucination with Extended Prompts

This chapter examined how hallucination intensified with extended prompts in ChatGPT. The problem became more severe when prompts were overly long or required extracting substantial information. Both models frequently produced fabricated details about students' learning profiles when designing differentiated learning processes based on learning styles, particularly at the initial response generation stage.

3 Limited Creativity in Custom GPT Responses

This chapter discussed limitations in Custom GPT's creative capabilities and proposed strategies to minimize hallucination. Enhancing data quality is crucial, with the model performing better on English documents than Chinese PDFs due to processing difficulties. Optimizing prompt design through clear instructions and logically sequenced prompts—using Prompt Recipes as predefined templates—improves accuracy by preventing fabricated content. Implementing Retrieval-Augmented Generation (RAG) combines generative capabilities with external knowledge retrieval, enabling more accurate, contextually grounded responses. Developing RAG models customized for Chinese language instruction could further enhance precision and alignment with instructional goals. Addressing these limitations through advanced strategies can strengthen ChatGPT's effectiveness for supporting differentiated instruction in Chinese language education.