The Need for Assessment of the Cascade Effects from Sequencing AI Systems in Policing

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

This position paper examines the use of multimodal AI in policing in England and Wales focusing on the risks that arise when the output of one system becomes the input to another. Drawing on 25 stakeholder interviews across the criminal justice system, we find that outputs from AI systems used for transcription, classification, summarisation, or information retrieval are used as inputs to downstream tools. We argue that this sequencing creates a distinct inclusion and governance problem: errors and biases introduced at early stages may be propagated, amplified, or transformed downstream, with particular risks for people whose speech, language, behaviour, accent, disability-related communication patterns, or distress diverge from normative assumptions embedded in training and evaluation data. Even minor inaccuracies, such as transcription errors, may distort evidential narratives and affect legal outcomes. Using examples from AI-assisted transcription and witness statement drafting, we identify the need for chain-level assurance methods that make provenance, uncertainty, and error propagation visible to practitioners. We conclude by proposing a research agenda for inclusive assessment of chained multimodal AI in high-stakes public-sector settings.