The Perils of Predictive Policing: Critically Analysing the Human Rights Implications of Predictive Policing Technology

Each article has been preserved in its original submitted form. No edits have been made to grammar, spelling or AGLC4 referencing.

By Elizabeth Browne

“A senior local officer explained: ‘We can run that tool now and it will tell us—like the kid might be 15—it tells how many crimes he is going to commit before he is 21 based on that, and [has] a 95% accuracy. It has been tested.’”[1]

Predictive policing (PP) represents one of the newest and most controversial developments in law enforcement. Using artificial intelligence (AI), PP analyses historical crime data to predict where, and by whom, crime will occur. Whilst proponents of PP argue that this technology allows police to effectively concentrate their limited resources in problematic areas, it also significantly undermines the right to non-discrimination by reproducing systemic racial biases in law enforcement under the label of an ‘objective’ framework. Thus, this essay contends that PP poses an unjustifiable risk to human rights in Australia, lacking the efficacy that would justify its widespread implementation. This essay firstly analyses the research surrounding PP, arguing that there is insufficient data to support its reliability and accuracy. Following this, it argues that PP risks further entrenching racial bias, particularly against First Nations peoples, within Australian policing given that PP has been shown to disproportionately label minority communities as crime ‘hotspots.’ This essay finally concludes that whilst PP and related AI technologies will likely bleed into law enforcement, human rights violations can be prevented by implementing legislative and political restrictions on AI in law enforcement. Thus, as AI gradually leaks into Australia’s justice system through technologies like PP, limitations on their usage must be urgently implemented to protect human rights.

I THE ARGUMENT FOR PREDICTIVE POLICING: FASTER AND BETTER CRIME PREVENTION?

Policing strategies involving the prediction of where crime will occur, and by whom, predate AI, with so-called ‘intelligence-led’ policing emerging in Australia in the late 1990s.[2] Boasting a supposedly “objective…decision-making framework,” current PP algorithms build upon older ‘intelligence’ models, introducing a largely computerised process that supposedly reduces human bias in policing whilst also better concentrating police resources in historically problematic areas.[3] Given research shows that crime often follows predictable spatiotemporal patterns—for example, a burglary in one neighbourhood increases the risk of subsequent burglary in that same neighbourhood—PP amalgamates known criminal patterns with historical data to identify potentially problematic areas and people.[4] Such technologies have already been adopted by police departments across Europe and the United States, and there is evidence to suggest that Australian police have already implemented such strategies.[5] One of the most notable examples of this technology is ‘Geolitica’ (formally known as PredPol), which emerged out of California around 2011, and was subsequently adopted in varying forms across the United States and Europe.[6]

However, despite its international adoption, research has not unanimously found it to be effective at crime prevention. Indeed, some studies have endorsed the potential of PP; Italy’s algorithm KEYCRIME predicted crime hotspots in Milan with a 70% accuracy rate, whilst the Dutch Police Academy found that PP positively improved management of police.[7] Similarly, certain evaluations of Geolitica/Predpol found it to reduce crime by 7.4% to 40%.[8] That being said, such findings are strongly counterbalanced by opposing evidence. For example, one large-scale analysis of over 23,631 Geolitica predictions (one of the largest and most influential PP algorithms) found that its predictions had less than a one percent accuracy rate.[9] Likewise, a subsequent test of the same algorithm found an accuracy rate of only around 4%.[10] As a result of this high possibility of inaccuracy, PP programs have already been shut down or re-evaluated across Europe and the United States.[11] Thus, PP does not currently represent a consistently reliable method of predicting crime hotspots.

II PP AND THE RIGHT TO NON-DISCRIMINATION

The possible inaccuracy of PP is unacceptable given its significant threat to rights like the right to non-discrimination; rather than eliminating human bias from policing, it perpetuates it under the guise of an ‘objective’ system. In particular, PP risks entrenching further racial bias against minorities, including First Nations peoples, within law enforcement, violating Australia’s non-discrimination obligations within the United Nations Declaration of Human Rights, and the United Nations Declaration on the Rights of Indigenous people (UNDRIP).[12] This is because most PP algorithms disproportionately identify minority communities as crime ‘hotspots’ because such groups are overrepresented in older crime data due to over-policing and systemic bias.[13] For example, 82.3% of possible criminals identified by the Metropolitan Police’s ‘gang violence matrix’ belonged to an ethnic minority, with 74.6% of them being of African-Caribbean descent.[14] Likewise, the Los Angeles police’s ‘Suspicious Activity Reporting’ program found that 30% of reports to data fusion centres implicated Black-identifying individuals, despite Black-identifying individuals comprising less than 10% of the population.[15] Thus, as a result of skewed historical data that disproportionately implicates minorities, PP reproduces systemic biases under the guise of an ‘objective’ system, fuelling racially motivated stop-and-searches and over-surveillance.

Applying PP’s possible racial bias to Australia, it is very likely that one of the most targeted groups by PP will be First Nations peoples. Notably, Australia’s Indigenous peoples are currently one of the most incarcerated peoples in the world and are routinely over-policed, violating their rights to self-determination in UNDRIP.[16] For example, Aboriginal and Torres Strait Islander people disproportionately come into contact with the justice system, with 17% of searches by the NSW Police involving Indigenous peoples despite them comprising only 3.4% of the population.[17] Furthermore, existing ‘intelligence’ and risk-assessment tools within Australia already disproportionately subject Indigenous peoples to intensive monitoring, with minimal evidence to suggest this is either effective or warranted.[18] Thus, existing crime data fed to any Australian AI system will, as with Black communities in the United States and England, likely implicate Indigenous Australians as a result. Considering the already conflicting data supporting its efficacy, PP’s possible breach of the right to non-discrimination (amongst other rights, such as those laid out in UNDRIP) is an unjustifiable one.

III PROTECTING HUMAN RIGHTS IN THE DAWN OF AI AND PP

Hartzog et al. argue that despite many ethical concerns surrounding AI and the digitisation of the justice system, the world has already “enter[ed]…a new era where large portions of the law enforcement process may be automated…with little to no human oversight or intervention”; its ubiquity means that it is perhaps unreasonable to aim to eliminate all AI from policing.[19] This is supported by the fact that AI is already being used to police in most Australian states; for example, the Western Australia Police are currently trialling an AI facial recognition program, whilst the NSW Police has openly disclosed their research into PP and other AI-driven crime prevention strategies.[20] Within Victoria, the police have already begun to use AI to reformat community-submitted crime reports.[21] Consequently, the use of AI and related PP technologies in Australian policing is likely inevitable—the issue is thus one of when it is adopted, rather than if it is.

Consequently, it is necessary that the public demand that the government implement legislative restrictions on PP to prevent human rights breaches. Where the European Union’s radical ‘AI Act’ effectively restricted PP to the identification of areas, rather than individuals (limiting the use of personal data in AI-based policing), Australia not only lacks legislation restricting AI usage in law enforcement, but also lacks a coherent and legally enforceable human rights framework.[22] This leaves Australians, especially those of First Nations backgrounds, extremely vulnerable to human rights breaches by these new policing technologies. Thus, by demanding the government implement legislation restricting AI usage in policing and requiring transparency surrounding policing strategies, human rights can be better protected in the face of digitised law enforcement. Alongside this, even without an AI Act like the European Union’s, requiring even some transparency surrounding policing practices may indirectly protect human rights by exposing rights violations, thereby placing political pressure on police to comply. Such restrictions on PP represent a powerful means of protecting human rights in this new, computerised era of policing.

IV CONCLUSION

Thus, as AI bleeds into every sector of life, strict limits must be placed on its usage in policing, particularly given the likely inaccurate and biased nature of PP. Whilst there is limited data suggesting it may offer some benefits for crime prevention, such research is greatly limited and points to the necessity of improving AI before implementing it within Australia. This is especially important given the Australian police’s fraught history of over-policing and profiling minorities such as First Nations peoples, meaning that any AI system implemented would likely rely on already racist crime data. As the influence of AI on law enforcement becomes indomitable, it is hence essential that legislative restrictions be urgently placed on PP. Hence, whilst it may be impossible to entirely eliminate AI from law enforcement, it is important that its usage be at least confined in order to protect the rights of all Australians.


[1] Lois Weber, Research Brief #17: Risk-Based Policing and Belonging (Research Brief, July 2020), quoted in Adelle Ulbrick, Predictive Policing and Young People: Discriminatory Impacts of Pre-Emptive and Racialised Policing in Victoria (Report, Flemington and Kensington Community Legal Centre, 2021) 8.

[2] Marilyn Chilvers and Don Weatherburn, ‘Operation and Crime Review Panels: Their Impact on Break and Enter’ (2001) Crime and Justice Statistics Bureau Brief 1 (NSW Bureau of Crime Statistics and Research).

[3] Jerry H Ratcliffe, Intelligence-Led Policing (Willan, 2008) 5.

[4] Michael Townsley, Ross Homel and Jacqueline Chaseling, ‘Infectious Burglaries: A Test of the Near Repeat Hypothesis’ (2003) 43(3) British Journal of Criminology 615.

[5] Ulbrick (n 1) 8.

[6] Simon Egbert and Elena Esposito, ‘Algorithmic Crime Prevention: From Abstract Police to Precision Policing’ (2024) 34(6) Policing and Society 521, 523.

[7] Bas Mali, Carla Bronkhorst-Giesen and Mariëlle Den Hengst, Predictive Policing: Lessen voor de Toekomst (Politieacademie, 2017), quoted in Ulbrick (n 1) 371; Andrey Bogomolov et al, ‘Once Upon a Crime: Towards Crime Prediction from Demographics and Mobile Data’ (Conference Paper, ACM International Conference on Multimodal Interaction, 2014), quoted in Ulbrick (n 1) 371.

[8] George Mohler et al, ‘Randomized Control Field Trials of Predictive Policing’ (2015) 110(512) Journal of the American Statistical Association 1399; Matthew Fielding and Vincent Jones, ‘Disrupting the Optimal Forager: Predictive Risk Mapping and Domestic Burglary Reduction in Trafford, Greater Manchester’ (2012) 14(1) International Journal of Police Science & Management 30; Jerry H Ratcliffe et al, ‘The Philadelphia Predictive Policing Experiment’ (2021) 17(1) Journal of Experimental Criminology 15.

[9] Aaron Sankin and Surya Mattu, ‘Predictive Policing Software Terrible at Predicting Crimes’, The Markup (online, 2 October 2023) [3] <https://themarkup.org/prediction-bias/2023/10/02/predictive-policing-software-terrible-at-predicting-crimes>.

[10] Christopher E Church and Amanda J Fairchild, ‘In Search of a Silver Bullet: Child Welfare’s Embrace of Predictive Analytics’ (2017) 68(1) Juvenile & Family Court Journal 71.

[11] Wim Hardyns and Anneleen Rummens, ‘Predictive Policing as a New Tool for Law Enforcement? Recent Developments and Challenges’ (2018) 24(3) European Journal on Criminal Policy and Research 201.

[12] Kim Rubenstein, Parliament of Australia, Parliamentary Joint Committee on Human Rights, Inquiry into Australia’s Human Rights Framework, Committee Hansard, 20 October 2023, 43; Universal Declaration of Human Rights, GA Res 217A (III), UN Doc A/810 (10 December 1948); United Nations Declaration on the Rights of Indigenous Peoples, GA Res 61/295, UN Doc A/RES/61/295 (2 October 2007, adopted 13 September 2007).

[13] Brian Jefferson, ‘Predictable Policing: Predictive Crime Mapping and Geographies of Policing and Race’ (2018) 108(1) Annals of the American Association of Geographers 1.

[14] Mayor’s Office for Policing and Crime, Review of the Metropolitan Police Service Gangs Matrix (Report, 2018) 25 <https://www.london.gov.uk/sites/default/files/gangs_matrix_review_-_final.pdf>.

[15] Los Angeles Police Commission Office of the Inspector General, Review of Suspicious Activity Reports, Fiscal Year 2013/2014 (Report, 23 January 2015) 3 <https://www.oig.lacity.org/_files/ugd/b2dd23_197536b31f834289bc1cf0c94b32a40e.pdf>.

[16] Indigenous Law and Justice Hub, Submission No 157 to Parliamentary Joint Committee on Human Rights, Inquiry into Australia’s Human Rights Framework, 7, quoted in Parliamentary Joint Committee on Human Rights, Inquiry into Australia’s Human Rights Framework (n 12) 32; UNDRIP (n 12) art 3.

[17] Justice and Equity Centre, ‘Contemporary Policing of First Nations People’, Justice and Equity Centre (Web Page, 19 June 2026) <https://jec.org.au/focus-areas/first-nations-justice/towards-truth/contemporary-policing-of-first-nations-people/>.

[18] Vicki Sentas and Camilla Pandolfini, Policing Young People in NSW: A Study of the Suspect Targeting Management Plan (Report, Youth Justice Coalition, 2017).

[19] Woodrow Hartzog et al, ‘Inefficiently Automated Law Enforcement’ [2015] Michigan State Law Review 1763, 1789.

[20] Cam Wilson and Andy Burns, ‘Privacy concerns raised over Australian-first live AI police scanning trial’, ABC News (online, 11 August 2026) <https://www.abc.net.au/news/2026-08-11/ai-police-face-screening-trial-sparks-privacy-concern/107009644>.

[21] Ry Crozier, ‘Vic Police Uses AI to Reformat Community-Submitted Crime Reports’, iTnews (online, 14 July 2025) <https://www.itnews.com.au/news/vic-police-uses-ai-to-reformat-community-submitted-crime-reports-618334>.

[22] Paul De Hert and Vagelis Papakonstantinou, ‘The New Police and Criminal Justice Data Protection Directive: A First Analysis’ (2016) 7(1) New Journal of European Criminal Law 7; Scott Walker, ‘An End to the Patchwork Quilt? The Guarded Potential of the PJCHR’s Inquiry into Australia’s Human Rights Framework’ (2024) 30(1) Australian Journal of Human Rights 99.

Share your thoughts