Integration of NDVI data into ecological and biogeochemical modeling has placed more stringent requirements on the accuracy and stability of the measurement. We compare two recent AVHRR NDVI data sets included as part of ISLSCP Initiative II: (1) the Fourier-Adjusted, Sensor and Solar zenith angle corrected, Interpolated, Reconstructed (FASIR) monthly time series and (2) the Global Inventory Modeling and Mapping Studies (GIMMS) monthly time series. Although both started with nearly identical composited AVHRR GAC data sets, each data set has been processed differently to reduce sensor, atmospheric, and illumination effects that vary over time. We find that the resulting absolute NDVI data records differ substantially and consistently for large parts of the globe. These differences also propagate into the NDVI anomaly record (e.g., deviations from monthly or annual means) particularly in the 1984--1985, 1994 periods. To assess the effect of these differences on predictions of land surface CO2 fluxes, the fraction of absorbed photosynthetically active radiation (fPAR) was calculated from each record, and used to drive a biogeochemical model (CASA). On a global basis, calculated net ecosystem exchange shows large variability inherited from the NDVI records. However, these variations do not match global CO2 fluxes derived from atmospheric inversion of CO2 concentration measurements. We conclude that other processes (burning, physiologic response to stress) are likely responsible for major anomalies in the observed global land net carbon fluxes to the atmosphere during the period 1982--1998.