Based on a set of climate simulations utilizing two kinds of Earth system models (ESMs) in which observed ocean hydrographic data are assimilated using exactly the same data assimilation procedure, we have clarified that the successful simulation of the observed air-sea CO2 flux variations in the equatorial Pacific is tightly linked to the reproducibility of coupled physical air-sea processes. When an ESM with a weaker ENSO (El Nino-Southern Oscillations) amplitude than that of the observations was used for historical simulations with ocean data assimilation, the observed equatorial anticorrelated relationship between the sea surface temperature (SST) and the air-sea CO2 flux on interannual to decadal timescales could not be represented. The simulated CO2 flux anomalies were upward (downward) during El Nino (La Nina) periods in the equatorial Pacific. The reason for this was that the non-negligible correction term in the governing equation of ocean temperature, which was added via the ocean data assimilation procedure, caused an anomalous, spurious equatorial upwelling (downwelling) during El Nino (La Nina) periods, which brought more (less) subsurface layer water rich in dissolved inorganic carbon (DIC) to the surface layer. On the other hand, in the historical simulations where the observational data were assimilated into the other ESM with a more realistic ENSO representation, the correction term associated with the assimilation procedure remained small enough so as not to disturb an anomalous advection-diffusion balance for the equatorial ocean temperature. Consequently, spurious vertical transport of DIC and the resultant positively correlated SST and air-sea CO2 flux variations did not occur. Thus, the reproducibility of the tropical air-sea CO2 flux variability with data assimilation can be significantly attributed to the reproducibility of ENSO in an ESM. Our results suggest that, when using data assimilation to initialize ESMs for carbon cycle predictions, the reproducibility of the internal climate variations in the model itself is of great importance.