Replacing the Role-Player: What AI-Enabled Simulation Has and Hasn't Trained.
Maintaining advantage in critical incidents on human judgement, adaptability and decision-making under pressure. Increasigly, AI-enabled simulation has been promoted as the means of developing those capabilities at scale. This paper presents a structured review of peer-reviewed studies from 2019 to 2026 in which artificial intelligence was embedded within simulation-based exercises for law enforcement. The review extends the author's earlier review of simulation-based training in policing (2014–2024) by focusing specifically on the AI component and on what it is claimed to train. Publication is heavily recent, with fifteen studies meeting the inclusion criteria,. Each study is coded for the function performed by the AI and for the outcome actually measured. The paper argues that AI is currently being adopted to make interpersonal training scalable and consistent rather than to develop judgement under pressure, that a simulator which removes the role-player has solved a resourcing constraint rather than a cognitive one, and that acquisition decisions are being made against evidence that does not test the capability being bought. It proposes a minimum set of outcome measures for AI-enabled simulation studies so that claims made for these systems can be compared, validated and accredited.