Semantic-Fidelity Risk in AI-Mediated Revision of Multilingual Academic Abstracts

Abstract

Artificial intelligence (AI)-mediated revision can improve the fluency and presentability of English-medium academic writing while also altering meaning, method, claim strength, epistemic caution, disciplinary stance, or contribution framing. This mixed-methods study defines semantic-fidelity risk as an unjustified change to a source abstract’s communicative commitments and examines 40 human-authored abstracts paired with 40 outputs generated through one locked ChatGPT revision procedure. Automated comparison indicators were integrated with semantic-fidelity coding, independent double coding, adjudication, and evaluation by four experts. Methodological alteration was the most frequent primary category (7/40), while semantic drift (1/40) and stance shift (2/40) were treated as illustrative observations only. The candidate-risk screen identified 11 of 29 cases in the composite human high-risk reference (sensitivity = 0.379) and produced no false positives in this corpus (specificity = 1.000), demonstrating that low metric change did not establish fidelity. Oversight-decision agreement was high in percentage terms (14/16, 87.5%) but moderate after chance correction (Cohen’s kappa = 0.55), indicating genuine interpretive thresholds. The study offers a bounded framework for source-based human oversight rather than an estimate of general AI-system performance.