"We Would Never Write That Down"

This paper draws attention to new complexities of deploying artificial intelligence (AI) sensitive contexts, such as welfare allocation. AI is increasingly used in public administration with the promise improving decision-making through predictive modelling. To accurately predict, it needs all agreed criteria part decisions, formal and informal. empirically explores informal classifications by caseworkers make unemployed seekers 'fit' into categories applied a Danish job centre. Our findings show that these are documentable, hence traceable AI. However, caseworkers, they at odds stable explanations assumed any bureaucratic recording system involve negotiated situated judgments people's character. Thus, for moral reasons, find them ill-suited representation purposes choose not write down. As result, although classification work crucial centre's activities, denuded real-world (and real work) character this context. an important finding CSCW only about whether can 'do' particular previous research has argued. shows problems may also be caused unwillingness provide data systems. It purpose present empirical results research, followed discussion implications AI-supported practice research.