Background: Hispanic/Latino (HL) populations bear a disproportionately high burden of type 2 diabetes (T2D). The ability to predict T2D genetic risk using polygenic risk scores (PRS) offers great promise for improved screening and prevention. However, there are a number of complications related to the accurate inference of genetic risk across HL populations with distinct ancestry profiles. We investigated how ancestry affects the inference of T2D genetic risk using PRS in diverse HL populations from Colombia and the United States (US). In Colombia, we compared T2D genetic risk for the Mestizo population of Antioquia to the Afro-Colombian population of Chocó, and in the US, we compared European-American versus Mexican-American populations. Methods: Whole genome sequences and genotypes from the 1000 Genomes Project and the ChocoGen Research Project were used for genetic ancestry inference and for T2D polygenic risk score (PRS) calculation. Continental ancestry fractions for HL genomes were inferred via comparison with African, European, and Native American reference genomes, and PRS were calculated using T2D risk variants taken from multiple genome-wide association studies (GWAS) conducted on cohorts with diverse ancestries. A correction for ancestry bias in T2D risk inference based on the frequencies of ancestral versus derived alleles was developed and applied to PRS calculations in the HL populations studied here. Results: T2D genetic risk in Colombian and US HL populations is positively correlated with African and Native American ancestry and negatively correlated with European ancestry. The Afro-Colombian population of Chocó has higher predicted T2D risk than Antioquia, and the Mexican-American population has higher predicted risk than the European-American population. The inferred relative risk of T2D is robust to differences in the ancestry of the GWAS cohorts used for variant discovery. For trans-ethnic GWAS, population-specific variants and variants with same direction effects across populations yield consistent results. Nevertheless, the control for bias in T2D risk prediction confirms that explicit consideration of genetic ancestry can yield more reliable cross-population genetic risk inferences. Conclusions: T2D associations that replicate across populations provide for more reliable risk inference, and modeling population-specific frequencies of ancestral and derived risk alleles can help control for biases in PRS estimation.