Paper short abstract:

Revisiting Sacks' 'Sociological Description' in light of Love's 'descriptive turn', this paper explores the problems different kinds of observer have in describing what AI does and how using two examples - Google's AlphaGo on stage and a commercial natural language processing algorithm at work.

Paper long abstract:

Heather Love and colleagues, in a series of much discussed papers, advocate an approach to analytical work that would focus on surface rather than hidden orders and 'build better descriptions' by drawing on particular kinds of sociological research as a source of proximate inspiration. Reflecting on Love and colleagues' proposals, we return to the work of one of those they cite, Harvey Sacks, to ask 'which surfaces?' and 'better for what?' as a way of contributing to the debate they sought to initiate. We use a peculiar 'descriptive assemblage' proposed by Sacks to explore the 'descriptive politics' of contemporary AI. In 'Sociological Description', Sacks imagines a 'commentator machine' composed of two parts: a doing part and a describing part. This machinery does things while providing simultaneous commentaries on those doings. We are interested in the kind of commentator machine contemporary AI might be, i.e., in what the saying and doing parts are and how their relations can be resolved, and in exploring the problems different kinds of observer of AI have in describing it. AI and machine learning technologies are often said to speak for themselves, the proof of their efficacy displayed in what they do. Reviewing two examples involving 'higher' and 'lower' profile AI - the gameplay of Google's AlphaGo and the work of a commercial NLP algorithm - we examine the descriptive 'meetings' involved, the troubles they reveal and what we can learn from them when it comes to describing what AI does and how.