Generative artificial intelligence in ergonomic risk assessment: A comparative analysis of REBA scores by human and AI raters

BackgroundWork-related musculoskeletal disorders are a major occupational health concern, and REBA is widely used to evaluate postural risk. The role of generative artificial intelligence (AI) in REBA scoring remains insufficiently examined.ObjectiveTo explore REBA scoring differences between a certified human ergonomist and three generative AI systems under controlled multimodal input conditions.MethodsSeventy-seven anonymized posture images from four task categories were evaluated by a human reference evaluator and ChatGPT, Gemini, and Wrtn. Identical images and standardized posture descriptors were provided to the AI systems. REBA scores were compared using repeated-measures ANOVA with Bonferroni-adjusted post-hoc comparisons.ResultsSignificant rater, body-part, and rater × body-part effects were observed. Wrtn generated comparatively higher trunk and wrist scores and lower scores for the lower arm, Gemini generated comparatively higher Final REBA scores, and ChatGPT produced comparatively lower scores across several body regions.ConclusionsGenerative AI systems showed systematic body-region-specific differences in REBA scoring relative to the human reference evaluator. These exploratory findings indicate the need for further validation and expert oversight before AI-assisted REBA assessment is used in workplace practice.