Comparing Linguistic Features between Human-written High-Scoring IELTS Essays and AI-Generated ones

Abstract

Using two self-constructed corpora (25 essays each for Task 1 and Task 2), this study systematically compares the linguistic characteristics of high-scoring human IELTS essays (Simon’s model essays, Band 7+) with Doubao-generated texts across four core dimensions: lexical resources, syntactic complexity, cohesion and coherence, and task response. Quantitative analyses employ lexical indices from McCarthy & Jarvis (2010), syntactic complexity metrics from Lu (2010), Halliday & Hasan’s (1976) cohesion framework, and Stapleton & Wu’s (2015) task response evaluation model. Results indicate that AI-generated essays slightly outperform human essays in lexical diversity and complexity (e.g., Shannon Entropy, MM Entropy) and maintain high stability in task response metrics (SSI, RQA, SRM). Syntactic complexity, however, exhibits task-specific patterns: in Task 1, AI demonstrates “breadth complexity,” expanding surface structures with 1.7–2 times more basic units (e.g., words, verb phrases) than humans and relying heavily on coordination and complex noun phrases. In Task 2, humans achieve “deep logical complexity,” constructing nested dependent structures (higher DC/C and DC/T) and producing denser verb phrases (VP/T), supporting more efficient and precise argumentation. In cohesion and coherence, AI exhibits notable weaknesses, including limited referential diversity, repetitive conjunctions, and insufficient implicit cohesion, whereas human texts display superior logical layering, natural collocation, and deep semantic association. These findings highlight the complementarity between AI and human writing, offering empirical support for leveraging AI in IELTS teaching and guiding future model optimization to balance structural precision with nuanced contextual contextual adaptability.