In this paper we attempt to provide a formal assessment of AI interpretability based on an observation of machine learning development sessions as 'algorithmic encounters'. We begin by considering the unbalance between the existing lacuna of metrics to assess AI interpretability in relation to the abundance of established approaches for measuring accuracy. To this respect, we introduce an approach based on interactional sociology to the study of AI interpretability as the patterning of talk between different forms of machine learning models utilisation: reproductive and consumptive utilisation. We then provide a detailed scenario of how this approach could be designed into a formal analysis of AI interpretability, based on the example of the testing of rival models to predict football player transfer value.
|Publication status||In preparation - 2021|
- symbolic interactionism
- artificial intelligence
- gaming encounters