Abstract
This conceptual review examines the integration of artificial intelligence (AI) tools into visual effects (VFX) compositing workflows across three dimensions: (i) the current technical capabilities of AI tools, including generative compositing, deep image harmonization, and neural rendering; (ii) the cognitive science and human factors foundations of human perceptual judgment; and (iii) a proposed conceptual taxonomy organizing the gap between these two domains. Drawing on situation awareness theory, trust in automation research, the ironies of automation, and studies of tacit knowledge in visual expertise, the review structures compositing processes into a five-level taxonomy: Automated Visual Processing (L1), Physical Rendering Consistency (L2), Temporal-Sequence Consistency (L3), Semantic-Narrative Alignment (L4), and Human Perceptual Decision (L5). The taxonomy is explicitly positioned as a domain-specific manifestation of established human-automation interaction frameworks, notably Sheridan's levels of automation and supervisory control, adapted to the unique aesthetic and narrative dimensions of VFX compositing. The evidence reviewed suggests that while generative AI and neural rendering tools expand automation reach at Levels 1–2, high-level perceptual decisions requiring temporal, semantic, and narrative consistency (L4–L5) remain predominantly human-dependent under current paradigms. However, this may reflect temporal contingency rather than structural inevitability. A critical implication, underexplored in the literature, is the deskilling risk: if AI absorbs L1–L3 tasks, junior compositors may not develop the tacit perceptual expertise required for L4–L5 judgments. Rather than eliminating perceptual expertise, current AI technologies appear to redistribute the compositor's role from technical executor to perceptual supervisor, with significant implications for training, trust calibration, and professional development.