In a dense passage in a thick book (On Organizational Learning, 2nd ed.), Chris Argyris makes a point that should deeply concern social scientists including, perhaps especially, economists and organization theorists.
It is important for social sicentists to study double-loop change because if they focus only on single-loop change, they may unwittingly become servants of the status quo…
This consequence holds negative outcomes for social science as a science. It is becoming evident that there may be a paradox embedded in the goal that social science should be descriptive of the world as it is. If social scientists aspire to study individuals and systems as they are, they will inevitably fall short of their goal: a complete description of things as they are would have to include a valid description of the capacity to make significant changes, and of the mechanisms by which these changes will occur. Knowledge of these mechanisms will also produce valid generalizations about constraints to double-loop organizational change. Such significant changes require changes in the organizational governing variables and master programs, that is, double-loop changes. But double-loop changes cannot occur without unfreezing the models of organizational structures and processes now in good currency. These models, in turn, cannot be unfrozen without a model of a significantly different organizational state of affairs: otherwise, toward what is the organization to change? If these models are genuinely new, then they do not now exist. if they do not now exist, then their invention and their use is an act of proscription, a normative stance. Yet if the logic is correct, the normative stance is needed to get at the inner nature of the present double-loop features and potentials of the organization. Hence, a full description of the world as is requires the intervention of stimuli from a world that presently is largely theoretical. (70)
By “double-loop change” or “double-loop learning”, Argyris means “questioning or altering the underlying values of the system” (68). He uses the term in opposition to “single-loop learning” which “is designed to identify and correct errors so that the job gets done and the action remains within stated policy guidelines” (151-2). In single-loop learning, “the underlying program is not questioned” (151).
To paraphrase: the potential range of social system behavior can be known only by construction, not by description. Purely descriptive approaches to social science will underestimate the range of the possible and, to the extent that these descriptions are used by agents within the system to shape its development, description may ultimately constrain the possible.