A kitchen thermometer doesn't creep upward gradually when you turn on the stove: first it stays put for a while, then suddenly jumps a few degrees. According to researchers' calculations, global temperature behaved much the same way from 1950 to 2013.
During stable periods, temperature barely changed, even though greenhouse gases kept accumulating. But in 1987/1988 and 1997/1998, average temperature jumped so quickly that these jumps account for almost all the warming over the 63 years. Trend maps for the quiet years show only local patches, while maps of differences between periods look like global heating.
Such a pattern fits poorly with the "forcing–response" model, where temperature should smoothly follow the rise in carbon dioxide concentration. Instead, the data hint at a global thermostat: the system seems to hold to one "set point," then quickly switches to another. A simple regression with two factors (El Niño and climate regime) explains 78% of monthly and 87% of annual variability.
Now this idea has quantitative support. If climate indeed moves in steps, then forecasts based on smooth trends may be systematically wrong at transition points. The remaining questions are what exactly pushes the system from one step to another, and whether such jumps occurred in the past.
The idea of a climate thermostat was discussed as early as 1991, when the New York Times wrote about possible self-regulation of climate, and a cloud-based mechanism was proposed by Ramanathan and Collins.