It's not just US data, they do a similar thing in Australia too. You can compare the raw data to their ACORN-SAT2.1 dataset where they make "udjustments" to "correct" the recorded temperature data. Effectively what they end up doing is spreading the urban heat island effect that is recorded in metropolitan weather stations out to the rural stations, add to that the transition to automated weather stations that use almost instantaneous temperature measurements rather than the min/max thermometers that would have a reaction time of several minutes. This means that even very brief spikes of temperature get picked up where they otherwise wouldn't, leading to something like a 0.5-0.7C jump in the max temperature recorded. This can be observed where they have overlapping datasets of old and new. Basically, the temperature data is a fucking mess.