Some studies have qualitatively proved that nonuniform profiles along the light path in turbulent flows can cause temperature measurement inaccuracies in Laser Absorption spectroscopy (LAS), based on the analysis of Beer-Lambert’s law. In this work, we attempt to further analyze this nonuniformity effect quantitatively from the viewpoint of data analysis. Ten thousand synthetic CO2 absorption spectra are respectively generated from uniform profiles and five discrete sections nonuniform profiles. Sixteen machine learning/deep learning models are trained on the spectra of uniform profiles to estimate the (average) temperature. The top three models, i.e., VGG13, XGBoost, and GPR with constant kernel, are used to estimate the average temperature of nonuniform-profile spectra. Furthermore, the model sensitivities are examined for spectral twins, which have similar spectra appearance, but totally different temperature/concentration along the light path. The results demonstrate that while all models work well on uniformprofile spectra, the three best models cannot provide accurate average temperature estimation for nonuniform-profile spectra. The maximum of the absolute error can reach up to 942 K and the corresponding root mean squared errors (RMSE) are all above 200 K. Moreover, the top three models have weak sensitivity when it comes to distinguishing temperature differences between spectra twins, particularly when more stringent spectra twin selection criteria are used.