Context. Stellar carbon abundances are crucial for tracing the star formation history and predicting the near-infrared emission of galaxies. It is still a complex task to derive accurate carbon abundance estimates for a wide variety of stars because it is hard to measure it based on atomic and molecular lines. It therefore remains challenging to include the abundance in stellar population models. Aims. We analyse the carbon abundance determination for the large empirical X-shooter Spectral Library that is commonly used as a benchmark for the development of stellar population models. Methods. We analysed the carbon abundance over strong molecular CH bands in the G-band region. We used the automated spectrum synthesis code GAUGUIN and adopted two different grids of separate reference synthetic spectra, each with the same coverage of the [C/Fe] abundance. We carried out a detailed comparison of the two grids to evaluate the accuracy and the model dependence of the measured [C/Fe] abundances. Results. We obtained a large and precise unbiased [C/Fe] abundance catalogue (~200 stars) from the two theoretical grids, well distributed in the Hertzsprung-Russell diagram and with no a trend with the stellar parameters. We also measured compatible values from each independent CH band, with a high-quality [C/Fe] abundance estimate for both dwarfs and giants. We always observed a dispersed flat trend around [C/Fe] ~ 0.0 dex throughout the entire metallicity regime we covered (–5 < [Fe/H] < +0.5 dex). This agrees with some previous studies. However, we found variations up to |Δ[C/Fe]| ~ 0.8 dex in the [C/Fe] composition of the star depending on the adopted grid. We found no such difference in the α-element measurements. This behaviour implies that the [C/Fe] abundance estimate strongly depends on the model. Conclusions. Potential sources of error might be associated with the use of spectral synthesis methods to derive stellar carbon abundances in the CH 4300 Å band. Intrinsic small differences in the synthetic models for this crowded and blended region may induce a large disparity in the precise abundance estimate for any stellar type, which leads to inadvertently inaccurate carbon measurements.