Diachronic Interpretability & Automated Systems

Abstract

This paper argues it is time we aim beyond ‘merely synchronic’ interpretability of automated systems. That most of the debate to date has been content to limit itself to that synchronic -and hence individualistic- level is not surprising: our understanding of what interpretative practices entail is still heavily influenced by a long tradition of text-based interpretation. In the latter context, there is no need to distinguish between synchronic and diachronic interpretability. If one has the former, one has the latter too. That this is not necessarily the case with the interpretability of automated systems is down to the impact of such systems on our continued capacity for normative agency. This capacity can be lost. By making any effort of critical engagement redundant, reliance upon non-ambiguous, highly efficient code-based systems of administration and behavior management will affect the extent to which we are made to flex our ‘normative muscles’. What if we enjoy the comforts of automated, simplified practical reasoning a bit too much, a bit too long? What was born out of efficiency and accessibility concerns may become a `normative holiday’ which we are unable to put an end to, for want of being able to mobilise normative muscles that have become atrophied through lack of exercise. This paper concludes with concrete ways of implementing interpretability requirements that are more likely to preserve diachronic (rather than merely synchronic) interpretability, and hence avoid what I call ‘the Wall-E scenario’.