Study region: Central Asia (CA) Study focus: This study characterized the uncertainty of a machine learning (ML)-derived evapotranspiration (ET) product, which has quasi-observational properties (ETML) and examined its applicability across Eurasia. We used Extended Triple Collocation (ETC) to quantify the error variance in ETML and benchmarked it against five physically based Penman models and Eddy Covariance (EC) FLUXNET ET data spanning 1983-2018. The main goal was to determine the uncertainty propagated by the ETML transfer learning framework and to evaluate the product's utility for ungauged stations, focusing on performance across different climates and plant functional types (PFT) in CA. New hydrogeological insights from the region: This study introduces a novel application of ETC for uncertainty characterization of ETML in a data-scarce region, confirming its efficacy in ungauged arid and semi-arid areas. Our results highlight the region-dependent error variance and accuracy of ETML under different PFTs and climate zones. ETML's temporal patterns align closely with the EC method, with a slight variance of +/- 0.03 mm yr-1 , though discrepancies appear during low ET periods (1985, 1992-1997) and high ET periods (2013-2016). ETML indicated an overall uncertainty of +/- 0.137 mm day-1. The ETML transfer framework is a notable advancement in CA's water resource management which can effectively generalize EC-based ET to target domains, though cautious application is advised.