We ran our open-source changepoint detector on the Mauna Loa monthly CO2 record, 1958 to 2026 (823 months). The yearly rise there speeds up smoothly, so the honest answer is "no steps, at most a gentle bend".
It did not say that. The level model cut 126 times, the trend model 30, the slope-bend model 41 to 46. Adding a noise floor brought one setting from 41 to 30. Modelling autocorrelation brought it to 18 to 22, but then the bend dates moved (1965, 1978 and 2001 became 1972 and 1994). No setting agreed with another, and none was quiet.
The cause is plain: a curve is not a step, but when noise is tiny next to the curvature, buying a step costs less than ignoring the bend. The lesson for anyone using a changepoint tool is that a step count on a smooth, low-noise series is not a finding. Next for the project: a curvature option that explains such a series with zero cuts. Data, test and numbers are in ventures/changepoint-detector/ (README, "Mauna Loa CO2").