Responsible AI in criminal justice: LLMs in policing and risks to case progression

LLMs are being considered for 15 different policing tasks, ranging from evidence analysis to report writing, but the paper identifies 17 distinct risks associated with their use that could undermine criminal cases and justice outcomes.

The risks have real consequences for case progression: The researchers documented over 40 examples showing how LLM failures could directly impact whether cases move forward, are dismissed, or result in wrongful outcomes—making this not just a technical concern but a justice system problem.

Many risks are manageable through established good practices, but addressing them requires deliberate effort and resources; the paper emphasizes this won't happen automatically and needs proactive system-wide planning rather than waiting for problems to emerge.

The research is grounded in England and Wales' actual legal and policing systems, meaning the identified risks reflect real constraints and requirements of how criminal justice actually works, rather than theoretical scenarios—making the findings directly applicable to policy decisions.

There's an urgent need for system-wide coordination: The paper stresses that individual organizations implementing LLMs in policing won't solve these problems alone; defining impacts and benefits requires coordination across the entire criminal justice system to prevent fragmented, risky deployments.