Mary Douglas

In the AI Story

Douglas was trained at Oxford in the British social anthropology tradition, and her early fieldwork among the Lele of Kasai (now Democratic Republic of Congo) gave her the comparative material that informed all her later theoretical work. The central insight of her career was that apparently irrational beliefs about purity and pollution turn out to be remarkably coherent when read as statements about social organization rather than as claims about empirical danger.

The collaboration with Wildavsky extended this insight to contemporary industrial societies. Environmental movements, public health panics, and regulatory debates all exhibited the same cultural patterns Douglas had observed among the Lele: the danger perceived was the danger that mattered for the social form at risk. This was controversial precisely because industrial societies typically flatter themselves that their risk perceptions are based on evidence rather than culture. Douglas and Wildavsky's argument was that this self-understanding was itself culturally constructed, and that cultural theory applied to the Sierra Club and the Union of Concerned Scientists as much as to any small-scale society anthropologists had ever studied.

Douglas's relationship to AI was prospective — she died in 2007, before the transformer architecture — but her framework applies with unusual force. The AI alignment discourse exhibits all the classification dynamics Douglas spent her career studying: the drawing of boundaries between safe and dangerous capabilities, the rituals of evaluation that certify some systems as acceptable, the purity concerns that animate debates about which training data should be included. A Douglas-inflected reading of AI safety would treat it as an anthropological phenomenon first, a technical one second, and would ask what the AI safety community's risk portfolio reveals about its cultural structure.

Her late work, particularly How Institutions Think (1986), extended the framework in directions that now seem prescient for the governance of algorithmic systems. Institutions, she argued, classify the world before individuals can classify it for themselves; the boundaries available for thought are shaped by the institutional arrangements within which thought occurs. Applied to AI, this implies that the categories through which we perceive AI risk are not given by the technology but by the institutional arrangements through which we encounter it.

Origin

Born Margaret Mary Tew in 1921 in San Remo, Italy, to British parents, Douglas trained in social anthropology at Oxford under E.E. Evans-Pritchard. Her fieldwork among the Lele of Kasai in the 1950s produced the material for Purity and Danger (1966), the book that established her reputation.

She spent most of her career at University College London, with visiting positions at Russell Sage Foundation and Northwestern. The collaboration with Wildavsky began in the late 1970s and continued until Wildavsky's death in 1993.

Key Ideas

Classification is social. The categories through which a society distinguishes clean from unclean, safe from dangerous, reveal the society's organizational form.

Risk is accounting. Identifying dangers is a way of expressing what a society values, not simply a way of reporting what harms it.

Grid and group. The two-dimensional framework that classifies social organizations also classifies their characteristic risk portfolios.

Institutions think first. The categories available to individual thought are shaped by the institutional arrangements within which thought occurs.

Cultural theory is universal. Industrial societies are not exempt from the classification dynamics that anthropologists documented in small-scale societies; they are further cases.