Geothermal springs are controlled by deep faults, which are the major causes of earthquakes. A series of geochemical indicators of hot springs, such as carbon dioxide content (CO2), carbon isotope and noble gas isotope etc., have been used for earthquake monitoring and early warning at home and abroad. Given that previous studies have failed to establish an inevitable link between the carbon isotopic variations of hot springs and seismic activities, in this study, mass spectrometry techniques were used to analyze the carbon isotopes of gaseous CO2 and noble gas isotope in hot springs, thermal water, and newly formed travertines from western Sichuan before and after an Ms4.8 earthquake (i.e., 7.14 earthquake). The results showed that the carbon isotope level of the hot spring system increased significantly before the earthquake, i.e., δ13CCO2 = -2‰ to 3.3‰, δ13CDIC = 2.2‰ to 6.6‰ (DIC = dissolved inorganic carbon in geothermal water), R/RA = 0.90–1.83, and δ13Ctravertine = 1.76‰ to 5.75‰. These indicators decreased markedly 1 week after the earthquake, whereas the helium isotope ratio of R/RA did not exhibit a noticeable change (δ13CCO2 = -3.9‰ to 0.4‰, δ13CDIC = 1.6‰ to 4.1‰, R/RA = 0.68–1.87). Combined with previous thermodynamic calculation and experimental results on carbon isotope fractionation of the CO2–calcium carbonate system, we suggest that the high carbon isotope characteristics of hot springs are caused by carbon isotope fractionation following the release of CO2 from fault carbonate decomposition owing to tectonic activity, not by a mixture of different carbon sources. A high carbon isotopic delivery path in geothermal water, from released CO2 to DIC and then to travertine, was observed. The new mechanism proposed in this paper reveals the intrinsic relations between the changes of the hot spring geochemical properties and deep tectonic activities, and the variations of these geochemical parameters before and after earthquakes may assist in capturing seismic precursors. A post-peak earthquake warning model has thus been preliminarily proposed, however, long-term monitoring data are required to reduce some uncertainties.