In this study, the authors investigate the observed record-breaking SST events by comparing with the expected rate for a trend-free random variable (TFRV). The authors find the asymmetric nature of the high and low records and reveal islands of cooling in the North Atlantic and Southern Ocean. The record-breaking theory is interesting; however, the assumption of the theory may not be enough reliable and the results may be sensitive to the length of a time series. Given these issues, I would recommend that this paper is not suitable for publication in this high-rank journal. More specific comments are listed as below.

Major comments:

1. The results of this study depend mainly on the comparison with the TFRV model. However, the climate system is clearly not a TFRV. Significant trends in SST can be detected even without the influence of human-induced greenhouse gases. As Deser et al.(2013) and Wallace et al. (2015), the internal variability is important for multi-decadal trends in climate variables, which are independent of human activities.

The authors use only a trend-free model for comparison, which is more of a hypothesis testing tool to test observed trends. and is of little implication for the climate community to understand the observations and climate change.

It is suggested that considering internal trends of the climate variables, such as add the trend distribution of Pre-industrial experiments from CMIP5/6 into the record-breaking statistics and comparing it with observed data, may yield more valuable results.

2. The results may depend to a large extent on the sample size. As shown in Equation 2, the broken k-record varies with the length of the time series. In other words, the results may be sensitive to the length of the time series and not robust.

The authors explore the SST trends over the last 75 years and discuss the spatial characteristics. They use a trend-free random variable technique along with standard methods to test statistical significance. The number of high records clearly outweighs the number of low records, with the majority of the SST grid points showing a warming trend. While I consider the paper sound from a methodological point of view, I’m not sure whether ACP is the ideal outlet given its statistical nature. The results do merit publication, but in order to make it work for ACP I would suggest incorporating CMIP6 models to add more analytical (i.e. physical) content to the discussion of the observed trend or their differences for that matter. I therefore suggest major revisions to make it fit for ACP.

Overall, I have three suggestions which may help to improve the manuscript:

1 Use CMIP6 pre-industrial control runs (random 75 year chunks) in order to compare the results when applying the TFRV technique. In order to obtain meaningful results (or learn something), I suggest selecting a dedicated CMIP6 subsample with a transient climate response (TCR) similar to observations, i.e. 1.3-1.8K. This way models with unrealistically high internal variability and/or too rapid southern ocean warming are likely omitted, which helps to confine the ‘true’ range of low-frequency internal variability. Taking all CMIP6 at face value without subsampling won’t help us to distill the magnitude of the unforced trends.

2 I suggest expanding the analysis to other SST datasets such as HadSST4, Kadow et al. 2020 (https://www.nature.com/articles/s41561-020-0582-5.pdf) or the (preliminary) version of HadISST2. They do tend to have small difference even after 1946 (between 1950-1980 in particular), hence it would be insightful to test them separately with the same method. It will certainly increase the robustness of the results. Expanding the analysis further by using the same subsample of CMIP6 SSP2-4.5 simulations between 1946-2020 would have the potential to add further insight into the forced nature of the signal and the physical reasons for missing warming trends in some ocean regions.

3 It might be worthwhile applying an ENSO correction to the data. The method provided in Foster and Rahmstorf 2011 (https://iopscience.iop.org/article/10.1088/1748-9326/6/4/044022/pdf) using multiple regression to filter out all natural factors (volcanoes, solar variability and ENSO), could offer some more interesting insights, both as far as trends as well as high and low record statistics at grid point level are concerned.