Governing with Artificial Intelligence

AI in law enforcement and disaster risk management

AI tools are used across various public bodies with a role in governance of critical risks, including disaster management, law enforcement and intersecting areas such as counterterrorism, customs and border management. Relevant agencies use AI technologies to enhance the scale, speed and accuracy of capabilities for anticipatory analysis, real-time surveillance and monitoring and administrative processes while lowering operational costs. This can enhance investigation and crisis management capabilities, optimise resource allocation and improve response times.

Because AI use has high-impact potential in this space, especially with regard to law enforcement, these entities are considered high-risk AI end users (OECD, 2020[254]) and face additional considerations when tapping into AI for public good and safety. Above all other types of public institutions, citizens place the highest levels of trust in police (OECD, 2024[46]). Relevant agencies need to uphold this trust by ensuring AI is adopted in a trustworthy manner that mitigates ethical challenges and risks to the protection of personal rights, including through vigilance regarding how underlying data are sourced, maintained and applied (OECD, 2020[255]; 2019[191]; 2022[256]). Lessons can be derived from existing use cases, and as discussed below, some governments have put in place heightened accountability expectations for AI use by these actors.

Current state of play

Identifying criminal suspects and missing persons

AI-powered facial recognition technology enhances capability to match real-time or recorded images from surveillance footage, social media and other sources against a stock of image databases. This may be used to assist in identifying criminal suspects, persons of interest and missing persons. This has also shown to have some deterrent effect on persons who might otherwise attempt to enter areas where access is prohibited to them. For example, in many countries closed circuit television (CCTV) footage is combined with facial recognition technology to enable real-time checking of spectators entering or leaving major sports venues without stopping the activities inside the facility (Hutchins and Andrejevic, 2021[257]). In the United States, the Next Generation Identification system, operated by the Federal Bureau of Investigations (FBI), has implemented AI-powered systems to deliver more accurate biometrics capabilities, including name matching, fingerprint matching, facial recognition, iris recognition and latent print matching. The system generates ranked lists of potential matches, which are then manually reviewed by biometric analysts to ensure accuracy. This process has proven effective in aiding investigations, providing leads in cases where other identifiers are either degraded or missing (US GAO, 2024[258]).

Also with regard to biometrics, the use of facial recognition technologies has raised concerns over infringement of privacy in many jurisdictions (OECD, 2024[259]) The need for a robust legal and policy framework that places limitations on interventions based on such analysis is discussed further below under “Managing risks and challenges”.

In addition to biometric identification, AI algorithms help analysts sift through vast amounts of data to identify patterns and anomalies that may indicate a higher or lower likelihood of a specific person’s involvement, or to uncover criminal activity that otherwise would have gone unnoticed. This includes analysing social media posts, financial transactions and signals intelligence for evidence of illicit behaviour that has already occurred.

Anticipating criminal behaviour and identifying risk factors

By analysing the psychological profiles, behaviour and past actions of individuals, AI systems can anticipate potential future criminal behaviour and offenses (Hung and Yen, 2020[260]). Sources of information to derive these predictions include social networks, online transactions, public records, existing law enforcement documentation and CCTV footage, among others. Such applications have been used for monitoring known offenders or identifying previously unknown individuals who may be planning terrorist attacks or other violent crimes (McKendrick, 2019[261]). They are also deployed for high frequency non-violent offenses, and identifying suspicious behaviour that stands out as a statistical anomaly (Wastupranata, Kong and Wang, 2024[262]).

AI-powered systems have also been used in border management. For example, the US Customs and Border Protection (CBP) has deployed a Passenger Security Assessment Model that to identify potential risks in passenger crossings, focusing on various indicators of concern. This model is designed to support CBP personnel in quickly recognising crossings that may warrant additional scrutiny. (US DHS, 2024[263]) A further example is CBP’s use of a ML system to identify potentially suspicious patterns in a border-crossing histories of vehicles, which has led to 240 seizures of thousands of kilograms of illicit drugs (US DHS, 2024[264]).

AI tools can be used to analyse social networks and relationships to identify individuals who may be at higher risk of becoming involved in criminal activity, as a perpetrator or a victim. This can help in targeting interventions and community policing efforts. AI technologies are designed to provide investigative assistance, for example, by alerting police regarding whom to watch before crimes are committed.

There are many relevant examples around the world. This can be seen in small-scale uses. For instance, in Japan, AI shoplifting detection software detects potentially suspicious body language on security camera footage and alerts authorities about potential thieves through a smartphone application (Belova, 2020[265]). They can also take the form of larger-scale operations, such as the UK Durham Constabulary’s Harm Assessment Risk Tool (HART). Predictive policing has raised controversy, as discussed below.

Box 5.53. Germany’s risk assessment tool for Islamist terrorism

The Risk Assessment for Islamist Terrorism (RADAR-iTE) is a standardised algorithmic risk assessment tool introduced by the German police throughout the country in 2017. The tool is specifically designed for use in state protection (Staatsschutz) contexts to evaluate the risk of politically motivated serious violent religiously motivated terrorist acts committed by individuals already known to law enforcement.

RADAR-iTE involves a standardised case preparation process in which risk and protective factors of an individual are assessed, assigning them to a two-tiered risk category (moderate or high). This enables prioritisation of individuals and supports the efficient allocation of police resources. The assessment relies on observable behaviours and information already available to or legally obtainable by police officers, rather than on attributes such as ideology or religiosity. The tool uses a risk assessment form with standardised questions and answer categories, but it is not automated; the evaluation of the features always requires the professional assessment of trained specialists.

RADAR-iTE facilitates prioritisation decisions by identifying the presence of risk and protective factors. Individuals assessed as high-risk are then discussed in case conferences managed by the Federal Criminal Police Office (Bundeskriminalamt, or BKA) within the Joint Counter-Terrorism Centre's (GTAZ) Risk Management Working Group (AG RIMA). The tool is designed to enhance the structuring and documentation of biographical histories of known individuals within the militant-Salafist spectrum, improving risk assessment transparency and comparability across Germany. It serves as a tool to assist, but not replace, individual case assessments, complementing standard measures applied to potential threats and relevant individuals in police state protection.

Source: https://www.bka.de/DE/UnsereAufgaben/Deliktsbereiche/PMK/Radar/radar_node.html, https://penal.org/sites/default/files/files/A-02-23.pdf.

Predicting times and locations at risk of criminal activity

In addition to predicting who may be at risk of committing or being a victim of crime, predictive policing algorithms analyse historical crime data to help predict when and where crime may be more likely to occur. These algorithms take into account various factors, such as time of day, weather conditions and socioeconomic data. Some police departments use these analyses to focus patrols in identified hotspots, helping agencies to determine patrol routes and deploy resources more strategically.

Some anticipatory policing systems use real-time data to adjust predictions and recommendations dynamically both in law enforcement and emergency management contexts. Several countries use CCTV surveillance, live incident reports and sensor data to refine predictions in real time. In Korea, for example, the Internet and Security Agency (KISA) and the National Police Agency (NPA) collaborated to develop technology aimed at helping potential stalking victims recognise and avoid danger. This system detects precursors of stalking crimes, issuing warnings to potential victims and alerting the police. Its goal is to apprehend offenders before any harm is done. This application uses AI technology to analyse information from surveillance cameras in real time (which were previously available only for post-crime apprehension of the offenders), making it possible to identify signs of stalking crimes in advance and inform the victims, thereby preventing any harm. (OECD, 2025[266])

Enhancing disaster risk identification and anticipation

AI is transforming how countries identify and anticipate risks across sectors. ML algorithms are being applied to vast datasets to detect early warning signals for a wide range of critical risks, from natural hazards to public health threats. Box 5.54 highlights how AI is used in a Canadian province to anticipate wildfire risks and pre-position assets. Greece is also using AI for early fire detection in forests and groves, as well as for flood prevention.1

Box 5.54. Anticipating wildfire risks in Canada

The Province of Alberta in Canada uses artificial intelligence to its support decisions on allocating resources and pre-positioning assets to combat wildfires. The innovative system predicts the likelihood of wildfires breaking out the following day across protected forests, which cover more than half of the province.

Trained on a decade's worth of fire, weather and ecological data, the AI system incorporates daily weather forecasts and factors such as increased human activity during long weekends. The system has a 50% success rate in predicting wildfires and is accurate 80% of the time when it indicates a fire is likely. Crucially, the tool does not replace human decision-making but supports staff.

Developed through collaboration between the public and private sectors, the AI system continues to improve each year as new wildfire data is integrated.

Source: https://www.oafc.on.ca/about/announcements/fire-news-headlines/how-alberta-uses-artificial-intelligence-wildland.

In Guatemala, AI analytical techniques applied to earth observation and street-level imagery have been used to identify structures susceptible to collapse in the event of an earthquake. The approach was able to identify approximately 85% of the buildings identified as vulnerable based on traditional civil engineering assessments using AI to identify damaged areas (GFDRR, 2018[267]; Tilon et al., 2020[268]).

Earth observation, connected devices and AI tools are also being used to improve the forecasting of wildfire risk (Nelson, He and Moore, 2024[269]). Earth observation imagery provides regular information on vegetation growth to ascertain fuel loads that drive wildfire potential. In Australia, a satellite with infrared technologies that is specifically calibrated to monitor the dominant local vegetation (eucalyptus and shrub) is under development to provide improved wildfire forecasting accuracy (Amos, 2020[270]).

Brazil created the Geospatial Transmission Management System (GGT System), which uses AI and satellite image processing to inspect transmission lines and therefore avoid wildfires (Box 5.55).

Box 5.55. AI and satellite images processing to protect against wildfires in Brazil

Wildfires are one of the main causes of blackouts in Brazil. To monitor the thousands of Brazilian Transmission Lines (TLs), Brazilian Electricity Regulatory Agency (ANEEL) created the Geospatial Transmission Management System (GGT System), which consists of a tool that uses AI and satellite images processing, to preventively inspect the maintenance of TLs against wildfires. The GGT System uses AI to process satellite images and identify all spans of transmission lines affected by fires. It automatically detects variations in the amounts of vegetation and can detect vegetation suppression activities carried out in the field by the transmission companies without the need for displacement by ANEEL’s inspection teams.

The system also automatically generates monitoring reports containing diagnostics on the adjustments of maintenance performed by the transmission companies in the safety lanes of transmission lines. This allows ANEEL’s inspection teams to act preventively and with enough assertiveness to mitigate the risks of blackouts caused by wildfires throughout Brazil. The transmission companies use the information generated by GGT and the alerts issued by ANEEL to improve their transmission line maintenance processes and reduce the risks of blackouts caused by wildfires. From its launch in 2017 to 2022, there was an 89% decrease in shutdowns caused by wildfires on the monitored transmission lines, although the number of annual wildfires has increased throughout the country.

Source: https://oecd-opsi.org/innovations/ai-and-geospatial-data-to-protect-against-wildfires.

In the United States, for example, a major energy services provider for the states of North and South Carolina used AI techniques ahead of Hurricane Florence in 2018 to determine where to pre-position the technicians needed to quickly restore power if transmission lines were damaged (Seto, 2018[271]).

AI can process vast amounts of data from diverse sources, such as medical records, social media and environmental monitoring devices, to detect patterns and anomalies that may indicate the onset of a disease, enabling authorities to take proactive measures. (El Morr et al., 2024[272]) AI is also being deployed in applications that address human-induced risks, not only those arising from natural hazards. Box 5.56 illustrates how AI can be used to advance risk anticipation and law enforcement.

Box 5.56. Korea’s AI-powered CCTV to protect against crime and emergencies

In Korea, as urban areas face increasing surveillance demands, cities are deploying AI-based selective monitoring systems to enhance safety, emergency response and crime prevention. These systems use existing CCTV infrastructure and advanced deep learning algorithms to automatically detect and classify objects, people and events, including accidents, fires and unusual pedestrian behaviour (e.g. loitering, fighting unauthorised access).

By prioritising relevant footage, AI-powered CCTV monitoring reduces reliance on manual surveillance, enabling real-time incident detection and rapid response. Key capabilities include multi-keyword trajectory analysis for object tracking, high-speed video search and intelligent event recognition to support law enforcement and disaster management. To ensure privacy, the system anonymises sensitive information, such as faces and license plates, and restricts video export through encryption.

Source: https://smartcity.go.kr/wp-content/uploads/2024/09/EN02.AI-Based-Selective-Monitoring-System.pdf.

Improving risk assessment and predictive modelling

Advanced AI systems are enabling more precise risk assessments by integrating varied and complex datasets. Predictive modelling powered by AI can simulate various risk scenarios, allowing decision-makers to assess potential impacts and plan mitigation strategies. These systems use data from multiple sources, such as climate records, socioeconomic indicators and infrastructure networks, to generate more comprehensive risk profiles. AI's ability to process large volumes of data enhances the accuracy of risk models, particularly in highly dynamic environments such as disasters. By identifying correlations and causal relationships between different variables, AI systems can provide more nuanced insights into the likelihood and severity of future risks. This improved understanding supports evidence-based decision-making and resource allocation. AI can also help modellers to run faster, less resource intensive and more accurate models. For example, the European Centre for Medium-Range Weather Forecasts (ECMWF) uses the Artificial Intelligence Forecasting System (AIFS) to enhance numerical weather prediction. The AIFS delivers up to 20% greater accuracy than leading physics-based models in several areas, including the tracking of tropical cyclones (ECMWF, 2025[273]). AI systems can be continuously refined as new data becomes available, ensuring that risk assessments remain up to date. This adaptability is particularly valuable in addressing emerging risks, such as those associated with cyberattacks.

Box 5.57. Singapore’s data-driven virtual twin

Singapore developed Virtual Singapore, the world’s first digital twin of a country, to support data-driven urban planning and enhance disaster management. Led by the Singapore Land Authority (SLA), this initiative uses advanced 3D modelling and real-time data integration to create a highly detailed digital replica of the city-state. Laser-scanning technology, deployed through aircraft and vehicles, captures precise terrain and surface information, mapping above-ground structures, underground utilities, environmental conditions and population movements. As a key part of Singapore’s Smart Nation strategy, Virtual Singapore provides policymakers and researchers with an interactive platform to test scenarios, anticipate risks, optimise services and enhance resilience to future emergencies.

The platform enables users to simulate the impact of new projects, assess transportation flows and model environmental factors (such as urban heat islands) and risks (such as floods). By facilitating cross-sector collaboration, Virtual Singapore supports better resource allocation, enhances decision-making, and fosters innovation in both the public and private sectors. Government agencies can use the platform to improve infrastructure resilience, while businesses and researchers use its data to develop smart city solutions. Additionally, the model plays a crucial role in emergency and disaster management, helping authorities plan responses to extreme weather events and optimise crisis coordination. With ongoing updates incorporating new AI approaches, IoT and new data sources, Virtual Singapore continues to evolve as a dynamic tool for shaping the future of urban living.

Source: https://oecd-opsi.org/innovations/virtual-twin-singapore.

Enhancing risk communication

AI tools are enhancing how risk information is communicated and understood by decision-makers and the public. By translating complex data into visual formats and accessible language, AI makes risk assessments more comprehensible. Interactive dashboards, chatbots and virtual assistants provide tailored information to different audiences, fostering better risk awareness and preparedness. (UNU Institute for Environment and Human Security, 2024[274])

AI can also personalise risk communication, delivering targeted messages to individuals based on their location, vulnerability and preferences. This approach helps ensure that information reaches those who need it most, especially during emergencies (Zhao et al., 2025[275]). Moreover, AI-powered sentiment analysis tools can gauge public perceptions of risks, enabling authorities to adjust their communication strategies accordingly. Real-time risk communication platforms improve crisis management by providing updates on evolving situations and guidance on protective measures. This dynamic information flow enhances public trust and encourages more proactive risk mitigation behaviour (Box 5.58).

Box 5.58. The United States’ hazard mitigation planning assistant

In the United States, the Federal Emergency Management Agency (FEMA) is currently exploring the use of a generative AI solution, called the Planning Assistant for Resilient Communities (PARC). This new system creates efficiencies for the hazard mitigation planning process for local governments, including underserved communities. Hazard mitigation plans are a foundational step that communities can take to build their resilience; however, they can be lengthy to produce and challenging for communities that lack resources to do so. PARC will specifically support State, Local, Tribal, and Territorial (SLTT) governments’ understanding of how to craft a plan that identifies risks and mitigation strategies as well as generate draft plan elements — from publicly available, well researched sources — that governments could customise to meet their needs. This could lead to more communities being able to submit grant applications for funding, making them more resilient and reducing disaster risks. 

Source: https://www.dhs.gov/ai/use-case-inventory/fema.

Improving disaster and crisis response

AI is improving situational awareness during disasters by processing satellite imagery, sensor data and social media updates in real time. This information helps emergency services map affected areas, prioritise resource allocation and coordinate response efforts. AI-powered decision support systems can recommend optimal evacuation routes, identify critical infrastructure at risk and predict the spread of hazards. AI can also monitor changes in a disaster response operating environment. By analysing live data feeds, AI tools can provide updates on road closures, power outages, and transport disruptions. These insights help authorities and communities adapt to changing conditions and maintain continuity of essential services.

AI can help with complex data gathering directly from affected populations. In Indonesia, residents of Jakarta have used social media to share flood updates, prompted by Bencana Bot, a chatbot from the Indonesian Disaster Mapping Foundation (PetaBencana.id). The chatbot invites users to report flood details, which are then mapped in real time and accessible online for free. The Jakarta Disaster Mitigation Agency also updates the map with official data. Immediately following large-scale flooding in the city, the platform has seen 2000% surge in activity within 12 hours, helping residents avoid dangerous areas and make informed safety decisions. (PetaBencana.id, 2021[276])

Response agencies are also using change detection algorithms. These can be applied to earth observation imagery to quickly identify impacted geographic regions as well as damage to individual buildings and infrastructure.

Enhancing post-disaster needs assessment and recovery/reconstruction

AI is enhancing post-disaster needs assessment by integrating data from satellite imagery, social media and administrative records to produce more rapid and comprehensive damage and needs assessments. These systems can automatically identify damaged infrastructure, assess the number of displaced individuals and estimate recovery costs. This automation accelerates the assessment process, allowing recovery efforts to begin sooner.

AI-powered decision support systems help authorities prioritise recovery projects by identifying the most urgent needs and optimising resource allocation. These systems can model different recovery scenarios, enabling decision-makers to evaluate trade-offs between rebuilding infrastructure, restoring livelihoods and enhancing resilience.

AI supports transparent and accountable reconstruction by tracking the progress of recovery projects and identifying delays or irregularities. This oversight helps ensure that resources are used efficiently and that vulnerable communities receive timely assistance. AI can also help to optimise the deployment of resources, such as assigning officials to specific areas or tasks based on predicted needs. These technologies may also be used for public safety purposes that fall outside predictive contexts, such as how to improve response times and overall efficient use of limited resources. A notable example from the US Federal Emergency Management Agency (FEMA) can be found in Box 5.59.

Box 5.59. Accelerating post-hurricane recovery with AI in the United States

In the aftermath of Hurricane Ian, which struck the state of Florida in 2022, the U.S. Federal Emergency Management Agency (FEMA) deployed an AI tool known as the Geospatial Damage Assessments (GDA) model. Traditionally, post-disaster damage assessments relied on manual inspections, a process that could take weeks and often delayed the delivery of financial aid to victims. The GDA model transformed this approach by automating damage classification.

This innovation used aerial imagery, satellite data and ML to rapidly assess structural damage across affected areas, significantly improving the speed and accuracy of disaster response efforts. It identified and categorised structural damage levels across over 1 million structures impacted. By integrating additional data — such as wind speeds, building characteristics and proximity to the hurricane’s path — the model provided a more comprehensive understanding of damage severity. It reduced the number of structures requiring human review from over 1 million to just 77 000, cutting assessment times from weeks to days. Within 72 hours of Hurricane Ian’s landfall, FEMA had insights into the extent of damage across affected regions, enabling faster resource allocation and recovery planning.

Source: https://gis-fema.hub.arcgis.com/pages/geospatial-da-training.

Managing risks and challenges

AI tools can enhance government capabilities related to law enforcement, disaster risk and crisis management presents clear benefits; however, their deployment should take into account several risks and challenges. An analysis of 61 cancelled public service automation initiatives across various countries found that half of the cancelled systems were related to policing and law enforcement (in part due to discontinuation of facial recognition), often as a result of critical reviews by government agencies, media or civil society (Redden et al.[229]). The top risks and challenges identified are discussed below, which governments should consider in their development and deployment of AI tools.

Associated risks

• Inadequate or skewed data in AI systems

• Misuse or questionable use of AI, resulting in surveillance and privacy concerns

• Lack of transparency and explainability

• Impact in public trust if perceived as unfair or invasive

• Over-reliance on AI

While the use of AI for certain productivity gains, such as automated completion of routine tasks and paperwork, are generally low risk and non-controversial, the use of AI in predictive policing raises several significant public policy considerations. These should be considered and addressed to ensure such uses provide credible analysis and include safeguards to ensure responsible use of their capabilities.

If AI systems rely upon inadequate or skewed data, it could lead to inaccurate or adverse outcomes for some individuals or groups. With regard to law enforcement and disaster risk management, this could, for instance, result in unlawful discrimination (to the extent a country has such laws). Policymakers should seek to ensure that these systems are rigorously tested and validated before deployment. They should require rates of true and false positives to be recorded and made available to scrutiny. Research using various methods for systems validation has proliferated in recent years (Stettinger, Weissensteiner and Khastgir, 2024[277]).

AI experts have highlighted invasive surveillance and privacy infringements that undermine the ability of individuals to freely exercise their human rights and freedoms as one of the most pressing AI risks (OECD, 2024[29]). The extensive data collection required for predictive policing risks unintentional violation of rights to personal privacy and data protection. The use of AI for surveillance purposes can lead to intrusive monitoring of individuals, potentially violating privacy rights. Policymakers should seek to promote public safety while also ensuring protection of individuals through privacy and data protection regulations to prevent misuse and overreach.

In bringing together the risks of bias and privacy infringements, remote biometric identification (i.e. facial recognition) has proven particularly controversial because the systems involved may carry an inherent technological bias (e.g. discriminating by racial or ethnic origins) (OECD, 2020[278]). AI’s use in public spaces may also undermine the ability of individuals to freely exercise applicable rights related to privacy, data protection and freedom of expression and assembly, according to some civil society organisations (Access Now, 2022[279]; CSOs, 2021[280]), EU enforcement bodies (EDPB-EDPS, 2021[281]), and some UN Special Rapporteurs (Khan et al., 2022[282]). Several issues related to the use of facial recognition in this context are discussed in Box 5.60.

Box 5.60. Facial recognition privacy and bias concerns

As facial recognition technology has matured, it has become increasingly capable of identifying faces in a crowd, matching images from sources ranging from CCTV footage to police databases to provide real-time surveillance, and identifying criminal suspects or missing people, among other applications. The privacy concern is that public surveillance technology enables governments to gather vast amounts of information about citizens for one purpose (without their consent) that could be used for different, unauthorised and unlawful purposes.

AI tools to enhance facial recognition technology, when trained on inadequate or skewed datasets, may reduce the accuracy of identification for some groups.

Cases have also emerged of inappropriate procedures by police forces leading to the use of poor-quality input data, substantially weakening the accuracy of facial recognition software. For example, police forces in the United States have sought to match drawings of suspects, poor quality CCTV stills, computer-enhanced images and even a picture of a suspect’s celebrity doppelganger to image databases. These examples underscore the importance of clear rules on precisely how the software should be used and to clarify whether a match is sufficient grounds for an arrest.

While much attention has focused on cases where the technology malfunctions (e.g. false positives) or is misused, some privacy advocates stress that general use of underlying surveillance or monitoring technology undermines the ability to freely exercise rights by diminishing the expectation of anonymity in public spaces, which is a condition for the exercise of rights such as freedom of assembly. For example, EU data protection authorities have issued opinions stating that the use of such systems in public spaces is incompatible with the right to data protection, under EU law, as scanning all people passing through these spaces interferes with their rights under EU law. Accordingly, in the EU, the EU AI Act only allows the use of “real-time” remote biometric identification systems in publicly accessible spaces for the purposes of law enforcement if such use is authorised under a member state’s national law and subject to adequate safeguards. In particular, each use must receive prior authorisation from a judicial authority or an independent administrative authority, and it must be notified to the relevant national data protection authority. Other economies draw a different balance in securing the benefits of while addressing the risks related to these technologies.

Source : (OECD, 2020[255]), https://medium.com/data-science/how-ethical-is-facial-recognition-technology-8104db2cb81b, www.bbc.co.uk/news/technology-47117299, http://ainowinstitute.org/publications/after-a-year-of-tech-scandals-our-10-recommendations-for-ai , www.flawedfacedata.com, www.americaunderwatch.com, www.bbc.co.uk/news/technology-48222017, https://edri.org/our-work/remote-biometric-identification-a-technical-legal-guide, https://eur-lex.europa.eu/eli/reg/2024/1689/oj.

Many predictive AI systems are proprietary and/or rely on “black box” models, making it difficult for the public and policymakers to understand fully why some operational decisions in the field are made. This lack of transparency can undermine trust in law enforcement. AI tools extrapolate results from statistical correlations between different data sources they aggregate. When designing and using these decision-support systems, developers and investigators should reflect on the relevance of using specific socio-demographic variables in predictive models. In the case of a popular AI tool used by the French Gendarmerie, the predictive model uses 15 socio-demographic variables which, according to the developers, are strongly correlated with crime. However, there is no transparency about the nature of these variables and no attempt to demonstrate a genuine cause-and-effect relationship (Lecorps and Tissandier, 2022[283]).

Finally, there is a risk that law enforcement agencies may become overly reliant on AI, neglecting other critical aspects of policing such as community engagement and traditional investigative methods. Evaluative studies of alternative AI tools used in predictive policing should be conducted to identify comparative strengths in terms of crime occurring in forecasted areas and number of successful police interventions (Birks, Townsley and Hart, 2023[284]). Relatedly, the use of AI in policing and disaster risk management raises ethical questions about the delegation of moral decision-making to machines. Policymakers should consider the moral implications of using AI to make potentially life-altering decisions in the context of law enforcement and disaster risk management. Ensuring that human oversight and ultimate determinations are maintained, and that AI is used to augment rather than replace human judgment is crucial.

Implementation challenges

• High costs of AI adoption and scaling and demonstrating results and return on investment

• Skills gaps

• Inflexible or outdated legal and regulatory environments

Implementing AI technologies in policing requires significant investment. Most police departments are local agencies covering a small geographic area with small populations. Functionalities built into standard workplace software and geographic information systems can support many predictive methods. However, the most expensive AI tools, software packages or computers may exceed the affordability (and perhaps needs) of local predictive policing programmes. Policymakers should analyse the relative costs to benefits and seek to ensure that resources are allocated effectively. Law enforcement personnel need proper training to effectively use and understand AI tools. Policymakers should seek to ensure that adequate training programs are in place (EUCPN, 2022[285]), which is often not the case.

Current laws may not adequately address the unique challenges posed by AI in policing. Policymakers need to develop and update legal frameworks to regulate the use of AI in law enforcement. Establishing industry standards and best practices for the development and deployment of predictive policing technologies is essential to ensure consistency and quality.

Untapped potential and way forward

AI tools continue to offer increasing opportunities to enhance the capabilities of law enforcement and emergency management agencies by improving public safety and responsiveness. The use of these tools across a broader range of public bodies with a role in governance of critical risks is expected to continue to increase as the underlying technologies advance and training of personnel becomes more prevalent. Anecdotes on the benefits of AI tools in relevant contexts proliferate, so too have narratives about irresponsible uses. Continued empirical research is needed to strengthen the evidence base both about measurable benefits and the drawbacks of using AI in these contexts. This will help calibrate its uptake where increased effectiveness and efficiencies can be demonstrated. Or on the contrary, it will help support more stringent oversight due to adverse impacts that erroneous results may have on human lives.

The ethical, legal and societal implications of AI adoption demand careful consideration and proactive measures to ensure that these technologies are deployed in a manner that is trustworthy and respectful of individual rights and freedoms. Policymakers should engage with communities to build trust and ensure that their concerns about AI deployment are addressed. Clear communication about how AI in policing and in disaster risk management — including its goals, benefits and limitations — is vital to maintaining public confidence (Evans, 2022[286]). Addressing the challenges outlined above requires a comprehensive and multi-stakeholder collaboration among end-user agencies, policymakers, technologists and the communities concerned. Research that draws on such consultations may expand the evidence of measurable benefits of AI use as well as deepen the understanding of unintended consequences.

Governments and law enforcement agencies should strengthen their understanding of AI’s limitations in predictive policing to reduce crime. Advanced statistical models and computational power are neither necessary nor sufficient to reduce crime. Algorithms can help predict the risk of future events but not the events themselves. While AI can simplify the search for patterns, the tools extrapolate from past events and are only as accurate as the underlying data used to make them. Humans are the most important elements in the predictive policing process; they needed to find and gather pertinent data, prepare it for analysis, review and interpret the results of these analyses to weed out incorrect conclusions, and offer recommendations for how to proceed based on analytical findings (Perry et al., 2013[287]).

The use of AI in law enforcement is one of the most challenging application areas for the technology. It provides potential high benefits to national security and public safety but also potential adverse impacts. For instance, biased algorithms may lead to false positives that result in erroneous arrests and imprisonment of innocent persons, or false negatives that result in overlooking unlawful activity (Petra et al., 2024[288]). Such actions strengthen the need to develop consultative mechanisms aimed garneting public support for deployment of AI tools within conditions and modalities of accepted guardrails. Such actions introduce significant uncertainty regarding the path forward for AI in this area.

Note

← 1.  See https://www.youtube.com/watch?v=JzLhcQijsz4, https://epiteliki.civilprotection.gov.gr/sites/default/files/2024-01/%CE%94%CE%B9%CE%B1%CE%BA%CE%AE%CF%81%CF%85%CE%BE%CE%B7_%CE%A0%CF%85%CF%81%CE%B1%CE%BD%CE%AF%CF%87%CE%BD%CE%B5%CF%85%CF%83%CE%B7_24PROC014095623_03.01.2024.pdf and https://civilprotection.gov.gr/en/deltia-tupou/b-kikilias-i-tehniti-noimosyni-stin-ypiresia-tis-asfaleias-ton-politon (in Greek).