Historically, meteorological observations have been made for operational forecasting rather than long-term monitoring purposes, so that there have been numerous changes in instrumentation and procedures. Hence to create climate quality datasets requires the identification, estimation, and removal of many nonclimatic biases from the historical data. Construction of a number of new tropospheric temperature climate datasets has highlighted previously unrecognized uncertainty in multidecadal temperature trends aloft. The choice of dataset can even change the sign of upper-air trends relative to those reported at the surface. So structural uncertainty introduced unintentionally through dataset construction choices is important and needs to be understood and mitigated. A number of ways that this could be addressed for historical records are discussed, as is the question of How it needs to be reduced through future coordinated observing systems with long-term monitoring as a driver, enabling explicit calculation, and removal of nonclimatic biases. Although upper-air temperature records are used to illustrate the arguments, it is strongly believed that the findings are applicable to all long-term climate datasets and variables. A full characterization of observational uncertainty is as vitally important as recent intensive efforts to understand climate model uncertainties if the goal to rigorously reduce the uncertainty regarding both past and future climate changes is to be achieved.