Statistical Analysis of Human and AI Machine Translation Accuracy, Style, and Context in Academic English

Clinical Snapshot

Key Findings

• P Value: p < 0.001 for all primary comparisons

• Effect Size: Cohen's d ranging from 1.53 to 2.81 (very large by conventional standards); grammatical and semantic error rates favoured HT with Cohen's d = 2.81

• Primary Outcome: Overall translation quality scores: Human Translation mean 4.76 vs. AI Machine Translation mean 3.73 (on an unspecified Likert-type scale), with statistically significant differences across accuracy, stylistic quality, contextual appropriateness, and overall suitability

• Nnt Or Sensitivity: Not applicable to this study design; among AI tools, DeepL performed best (M = 4.16, SD = 0.37) compared to ChatGPT and Google Translate, though pairwise statistics are not detailed in the abstract

• Confidence Interval: Not reported

Clinical Application

The study's conclusions — that human translators outperform AI tools for academic translation — are not directly actionable in clinical settings without replication in healthcare-specific translation contexts. The use of AI translation tools in clinical documentation, patient information materials, or medical research dissemination requires dedicated evaluation studies with appropriate clinical outcome measures. In the Australian context, the Therapeutic Goods Administration (TGA) and the National Health and Medical Research Council (NHMRC) have not issued specific guidance on AI translation tools for clinical or regulatory documents. The RACGP does not currently endorse specific AI translation platforms for patient communication. Australia's culturally and linguistically diverse (CALD) population creates genuine demand for high-quality medical translation; however, this study's findings cannot be directly applied to inform those clinical translation decisions. The National Accreditation Authority for Translators and Interpreters (NAATI) sets professional standards for medical translation in Australia, and certified human translators remain the standard of care for critical clinical communications. This study does not engage with any of these frameworks. This study does not address a clinical population. Its findings pertain to academic Chinese-to-English translation quality as assessed by domain experts. Any extrapolation to clinical, medical, or patient-facing translation contexts would be entirely speculative.

Abstract

AI-driven Machine Translation (MT) technology has raised doubts about whether it can replace Human Translation (HT), particularly in the academic English field. This research contrasts AI-based MT and HT of academic English on accuracy, stylistic quality, errors, contextual appropriateness, and overall suitability. A quantitative comparative design was employed for the translation of eight Chinese texts with various genres. Using purposive sampling, 4 certified human translators were selected to produce the HT, and also used AI-based MT systems such as DeepL, ChatGPT, and Google Translate to translate the text. 69 experts evaluated the translated versions, and this data was examined using SPSS software. HT produced better performance than MT through mean accuracy, stylistic quality, contextual appropriateness, and overall suitability (Human: 4.76, AI: 3.73), revealed statistically significant differences (p < 0.001), and effect sizes (Cohen's d = 1.53 to 2.81). Compared to HT, the MT showed higher grammatical and semantic errors (Cohen's d = 2.81). The fluency, tone, and register of stylistic quality were evaluated at higher levels for HT (p < 0.001). Among AI-MT tools, DeepL shows higher performance in translation (M = 4.16, SD = 0.37). The genre-wide comparison of MT and professionally certified HT, and the comparison among multiple AI tools, towards various translation dimensions, provides insights into the differences in translating quality. The research indicates a continued reliance on human professionals for important scholarly translations while encouraging continued development in AI translation software for better quality.

References

1. 1.Shi, B. (2025). Statistical analysis of human and AI machine translation accuracy, style, and context in academic English. International Journal of English Linguistics, 15(1), 17. https://doi.org/10.5539/ijel.v15n1p17