Differences in Translators' Cognitive Load under Generative-AI Post-Editing: An Eye-Tracking Study of English-Chinese Scientific Texts

{'en': 'Generative artificial intelligence has entered the translation workflow faster than evidence about its effect on the translator has accumulated. This study asks whether working from a large language model draft lowers cognitive load or merely redistributes it. Thirty-six translators, twenty translation trainees and sixteen practising professionals, rendered nine English popular-science and technical passages into Chinese under three within-subject conditions. These were unaided human translation (HT), post-editing of a static neural machine translation draft (NMT-PE), and post-editing of a GPT-4o draft with an interactive prompt pane (GAI-PE). Pupil diameter change, fixation duration, gaze distribution, keystroke logs, NASA-TLX ratings and MQM scores were recorded. GAI-PE cut task time by 44.2 per cent against HT and by 17.1 per cent against NMT-PE. Its MQM scores were higher than those of NMT-PE, while the difference from HT did not reach significance. Pupil diameter change was nevertheless largest in GAI-PE at 0.095 mm, against 0.083 mm in HT and 0.064 mm in NMT-PE, while gaze on the source text fell to 38.3 per cent, the lowest of the three. Subjective ratings moved the other way and ranked HT as the most demanding condition. In an exploratory analysis, self-reported AI-use frequency was associated with a smaller pupil response in GAI-PE, whereas English proficiency showed no such association. These observations are consistent with an increased demand for verification, though the generative condition also gave access to an interactive prompt pane, so the differences cannot be attributed to the model alone.'}