Through the Eyes of the Viewer: The Cognitive Load of LLM-Generated vs. Professional Arabic Subtitles

As streaming platforms adopt artificial intelligence (AI)-powered subtitle systems to satisfy global demand for instant localization, the cognitive impact of these automated translations on viewers remains largely unexplored. This study used a web-based eye-tracking protocol to compare the cognitive load that GPT-4o-generated Arabic subtitles impose with that of professional human translations among 82 native Arabic speakers who viewed a 10 min episode ("Syria") from the BBC comedy drama series State of the Union. Participants were randomly assigned to view the same episode with either professionally produced Arabic subtitles (Amazon Prime's human translations) or machine-generated GPT-4o Arabic subtitles. In a between-subjects design, with English proficiency entered as a moderator, we collected fixation count, mean fixation duration, gaze distribution, and attention concentration (K-coefficient) as indices of cognitive processing. GPT-4o subtitles raised cognitive load on every metric; viewers produced 48% more fixations in the subtitle area, recorded 56% longer fixation durations, and spent 81.5% more time reading the automated subtitles than the professional subtitles. The subtitle area K-coefficient tripled (0.10 to 0.30), a shift from ambient scanning to focal processing. Viewers with advanced English proficiency showed the largest disruptions, which indicates that higher linguistic competence increases sensitivity to subtle translation shortcomings. These results challenge claims that large language models (LLMs) lighten viewer burden; despite fluent surface quality, GPT-4o subtitles demand far more cognitive resources than expert human subtitles and therefore reinforce the need for human oversight in audiovisual translation (AVT) and media accessibility.