Is general intelligence (g) a reflective construct, representing a latent causal entity underlying subtest performance, or a formative construct, better understood as an aggregate variable shaped by and summarizing variation across subtests? Genetically informative data provide a framework for testing whether a construct is reflective or formative by comparing common pathway and independent pathways structural equation models (SEMs). Previous studies using biometric SEMs have predominantly supported the reflective model, with phenotypic g mediating the effects of additive genetic and environmental influences on lower level abilities. In the current study, four large genetically informed datasets (three from the US and one from the UK) were analyzed to test three competing SEM models — common pathway, independent pathways, and merged — using Confirmatory Factor Analysis (CFA). Genetic g was estimated in each sample as a latent variable derived from polygenic scores indexing educational attainment and cognitive abilities. The models were compared as follows: the common pathway model, consistent with a reflective g, included a direct path from genetic g to phenotypic g; the independent pathways model, consistent with a formative g, featured indirect paths from genetic g to phenotypic g via subtests; and the merged model incorporated both direct and indirect paths. Across all four datasets, the merged model consistently provided the best fit (based on goodness-of-fit and parsimony criteria). Phenotypic g mediated between 31% and 81% of the effects of genetic g on subtests. These findings suggest that g functions as both a reflective and formative entity.