Artificial Intelligence in Community Safety and Policing

The growing presence of Artificial Intelligence (AI) in society has prompted police agencies to explore the potential benefits in the police/community safety milieu. It is apparent that the adoption of AI has significant potential within the police environment. However, the implementation of AI has prospective hazards which must be understood and managed.

What is artificial intelligence?

Artificial intelligence (AI) is a transformative technology, but also a moving target. The term has shifted over time, from symbolic AI to machine learning, to generative AI, agentic AI and sometimes even artificial general intelligence or superintelligence. AI systems are machine systems that, broadly speaking, perceive, learn and act. They infer from inputs how to generate outputs such as predictions, content, recommendations, actions or decisions, with varying degrees of autonomy and adaptiveness. What unifies current AI more than any single architecture is that modern systems learn from experience represented by data. (UN Preliminary Report on AI, 20026.)

Much current discussion focuses upon the very broad foundation models, general-purpose AI. These comprehensive models concurrently perform a variety of complex tasks. However, the systems currently explored or implemented in policing are task-specific, narrow AI applications which are designed for performing specific tasks within a police/community safety environment.

Current use

Currently, police agencies in the UK, US, and Canada employ AI for eight principal tasks.

  • Analysis of data – tabulation of reported incidents, disposition, outcomes, and progress of cases through the court system.
  • Coordination of multi-agency incident response or investigations – the development of a matrix of resources/agencies engaged in a response to a complex incident or investigation and the tracking, monitoring, and subsequent review of activity.
  • Monitor public spaces – facial recognition, licence plate recognition, accelerated analysis of video evidence, such as dashcam and doorbell.
  • Triage calls for service from the public – call logging, reporting, redirection, risk assessment, multi-language translation.
  • Streamline operations – predictive policing and deployment based upon incident data.
  • Evidence and found property – registered, special storage requirements, multiagency linking to property issues in reported cases, date triggering for matters such as court and disposition.
  • Stream administration – communication between personnel, meeting preparation, public and media communications content, personnel records management, and HR issue analysis. 
  • Prepare and file reports – based upon formats which delineate required information and in conjunction with incident reporting from personnel or recorded body-warn camera livestream storage, dovetailing between various media.

 

However, the introduction of AI strategies has identified several concerns.

  • Evidence contamination – AI can produce erroneous, but appearing accurate, text based upon evidence which has been “misheard” or “misread”. For example, during two simultaneous sources of input, such as a radio broadcast in a vehicle and conversation input to a body worn camera. AI misconstruing an accent can result in erroneous text.
  • Biased data sources – AI is trained utilising existing data. These may already reflect a positive or negative bias based upon past perspectives.
  • Absence of a complete picture – AI tools are limited by the evidence which is entered or permitted to have access to. AI is unable to provide a situational context which is often crucial to police evidence.
  • Deference to a ‘higher information power’ – personnel may defer to AI, which, it is assumed, possesses superior powers of assessment and analytics. This confers legitimacy on AI probabilistic outputs. Personnel may perceive AI as having access to information, interpretation and conclusions not possessed by a manager. Personnel may be less inclined to consult a manager for assistance, guidance and decision.
  • Lack of transparency – perception of AI as a ‘black box’ inaccessible to management for review and to defence lawyers to challenge police evidence. Inaccessibility can be supported by vendor-imposed limitations.

 

Provincial Police Acts prescribe the role of the police, and adherence to the Acts is the touchstone of policing and protection of civil liberties.  Consequently, AI cannot replace the decision making of personnel at any strata of the organization. AI is merely an assist to achieve the attainment of policing objectives.

Guardrails

Although police agencies are developing policies and practises for the use of AI, the current paucity of legislation or guidelines means that police boards and executives should develop guidelines and guardrails based upon good practices.

  • Advocacy – to provincial and federal governments for legislation and standards as a framework for agency strategies.
  • Oversight – establish governance principles to introduce and monitor the use of AI.
  • Single organizational point of responsibility – it is essential that a human is ultimately responsible for the research, strategic planning, development of policy and guidelines, the introduction, maintenance, and the ongoing review of all AI-related outcomes. Similarly, responsibility for monitoring AI use and outcomes should be cascaded down through the organization commensurate with organizational levels of responsibility. For example, all incident reports in which AI has been used must be reviewed by a supervisor. Further, it is essential that a human makes the ultimate decision, based on reasonable grounds, to initiate police action.
  • Measured introduction – as AI is explored, initial introductory steps should be cautious, modest, piloted, monitored, and objectively assessed.
  • Disclosure – for court and for security purposes, case files should track the use of AI and the conclusions drawn based upon that use. To avoid cross contamination or unintended data exchange, staff should be required to disclose and discuss their use of personal AI when it interacts with workplace AI.

Conclusion

Although largely at an exploratory level in policing, evidence, so far, suggests that task-specific AI can deliver measurable benefits. However, the task must be well defined, accurate data are available, and institutions which deploy AI incorporate effective governance, measured and calculated introduction, and monitoring. Informed control by humans is essential at the principal junctures in AI-generated processes. In policing, supervisors, managers, and executive must actively oversee operations. Decisions which impact personal liberty and civil rights must be made by a human. As with all policing tasks AI generated work must be monitored, reviewed and periodically audited. AI is merely an assist to achieve the attainment of policing objectives.