When judgment varies: the potential of AI to reduce noise in social work assessments

Research across Europe demonstrates variability in social workers’ assessments of similar child vulnerability cases. This variability calls into question the quality of professional judgments and the legitimacy of social services, potentially undermining trust in public institutions and the perception of fairness at individual, professional, and societal levels. This article introduces the concept of ‘noise’ (Kahneman et al., 2021) to the field of social work, defined as unwanted variability in judgments. We understand noise as a consequence of the unique complexity embedded in child vulnerability assessments, which requires contextual understanding. Acknowledging that professional discretionary space is essential for addressing this complexity, we also recognise that predictability in assessments is equally crucial for ensuring fairness and institutional trust. To examine whether generative artificial intelligence (AI) can mitigate noise while preserving the discretionary space, we conducted a survey experiment where one group of social workers were exposed to AI-generated assessment before assessing a vignette and another group simply assessed without any AI-generated input. Results showed that the mean assessment of concern was equal between groups, meaning that the AI-generated input did not introduce new statistical bias. At the same time noise was significantly reduced among social workers exposed to the AI assessment.