Nicolas Metzl et al. present 36 years of measurements of carbonate system properties in the Indian Southern Ocean, an area important for oceanic CO2 uptake. They investigate multi-decadal trends and their causes. Among other results, they find the accumulation of anthropogenic CO2 to be the dominant control on carbon and acidification trends in their study area, modulated by natural decadal variability. Metzl et al. use the observed trends to extrapolate into the future, including estimates of the time of crossing undersaturation thresholds important for marine life. The findings are presented clearly and convincingly. I certainly recommend to publish this work. I would like to add that the sustained efforts to maintain high-quality oceanic measurements over such a long time frame already represent a great value for the research community in itself, far beyond what can be discussed in a single paper. 

Minor comments meant to further improve clarity:

Title: consider mentioning "Kerguelen" as a more easily recognizable geographic term

Line 59: Rödenbeck et al (2022) note that interannual variability before the 1990s may be underestimated because there are hardly any pCO2 data to add variations not captured by the extrapolation based on relationships to predictor variables

Lines 252-253: Why would the lower salinity speak against being representative?

Line 258: Just to clarify: Will the data not be present in any upcoming GLODAPv2.20xx release?

Line 433: The seasonal cycle is actually difficult to see from Fig 2a

Line 435: Somewhat unclear: do you mean the annual CO2 flux 1985-1998, or the winter one?

Line 494: It sounds as if this rate has been calculated from just the difference of 2 values. If so, why not from a linear fit?

Fig 5: I was wondering whether the break point between the 2 trends (orange) wasn't mainly induced by the large values in 1998 and 2000? Visually, the dip in 2008-2010 actually seems to be in the range of the general variability since 2002. Is is really justified to suggest a break in trends?

Lines 682-683: The change is termed "very small", but isn’t it actually outside the given uncertainty range?

Line 694: can you briefly indicate why lower O₂ explains lower Cant?

Line 716: Maybe replace "filtered" by “excluded” or "discarded"

Line 764: missing "uatm"

Line 765: If I understand correctly, a correlation between Chl-a and fCO2 is built into FFNN. Is such a correlation also seen directly in the data? I'm asking because if not, how do you know it is a real feature and not an extrapolation artifact? 

Lines 965-981: It took me quite a while to understand how to read Fig 12, though in the end everything makes good sense. Maybe there is a way to further help readers? I cannot offer a good suggestion either, unfortunately.

Line 1066: Consider writing "CT(1962)+Cant(t)" in both caption and figure legends, because otherwise the sum "CT+Cant" is quite confusing

Line 1096-1097: Couldn’t this hypothesis be checked by the FFNN, by calculating the FFNN response to a counterfactually constant SST predictor?

Line 1110: "undersaturation"?

Line 1312: "BGC-Argo"

see pdf attached.