Effective human–AI interaction in manufacturing requires that AI systems do more than deliver information—they must adapt to the operator's cognitive state, augmenting rather than overloading human performance. This study presents the first comparative empirical investigation of mental workload across three distinct human–AI interaction modalities—Virtual Reality (VR), an AI-integrated humanoid social robot (Furhat + GPT-4o), and video instruction—evaluated during both learning and operational phases of an industrial assembly task (n = 36). Using a triangulated measurement protocol combining subjective ratings (RSME), physiological monitoring (heart rate variability via PPG), and performance outcomes (completion time, error rate), we identify a critical dissociation between perceived and physiological cognitive load: VR produced the lowest perceived effort during learning (RSME median = 25.50) yet the worst operational performance (completion time median = 490 s; error rate median = 8), while video yielded consistently low workload and the highest accuracy. The humanoid robot occupied a distinct middle position, generating the highest HRV ratios during learning (median = 1.02)—physiological evidence of social engagement not captured by self-report—alongside moderate performance outcomes comparable to video. Crucially, VR's low perceived effort masked a hidden physiological cost that subjective measures entirely failed to detect, demonstrating that self-report alone is insufficient to evaluate human–AI interaction quality. The findings surface a fundamental design principle for human–AI symbiosis: different interaction modalities engage distinct cognitive and autonomic processes, and only triangulated measurement can reveal this. Together, the results provide empirically grounded guidance for designing adaptive human–AI training systems in Industry 5.0 manufacturing.