Predictive Policing and the Risk of Algorithmic Bias

Predictive Policing and the Risk of Algorithmic Bias

When the Future Becomes a Police Target

Predictive policing promises to make law enforcement “smarter”: algorithms analyze past crime data to forecast who is likely to offend, where crime is likely to occur, or which victims are most at risk. On paper, this seems attractive: limited police resources could be deployed more efficiently, and crime could be prevented rather than merely reacted to.

Yet behind this narrative lies a serious concern: algorithmic bias. Predictive systems are not neutral oracles. They are built from data that reflect historical practices, including over-policing of certain neighborhoods or groups, systemic discrimination, and incomplete reporting of crime. When such data are fed into predictive models, there is a real danger that past injustice becomes encoded as “objective” risk, legitimizing and amplifying unequal treatment.

This article explores how predictive policing works, where algorithmic bias arises, and what is at stake for criminal law, fundamental rights, and the future of policing.

---

1. What Is Predictive Policing?

“Predictive policing” is an umbrella term for various technologies that use data and algorithms to forecast crime-related risks. Three broad types can be distinguished:

1. Place-based prediction
   • Identifying geographic areas where crime is statistically more likely in the near future.
   • Often uses historical crime reports, sometimes combined with contextual data (time, weather, events).
2. Person-based prediction
   • Estimating the risk that specific individuals will commit crimes (offender risk) or become victims (victim risk).
   • Includes risk assessment tools used in pre-trial decisions, sentencing, probation and parole.
3. Network and group analysis
   • Mapping social networks to identify “key actors” (e.g. gang members, associates, influencers) for targeted policing.

In all of these, the basic idea is: use data to anticipate risk, then intervene earlier – by increased patrols, targeted surveillance, social programmes, or preventive arrests.

---

2. How Data Create Algorithmic Bias

Predictive policing systems are trained on data that look numerical and clean but are often deeply shaped by social and institutional biases. Key problems include:

• Policing bias in the data
  • Crime data reflect where police choose to patrol and whom they choose to stop, not just where crime “objectively” occurs.
  • Over-policing of certain communities (e.g. minority or low-income neighborhoods) generates more recorded incidents there, even if actual offending rates are similar elsewhere.
• Under-reporting and hidden crime
  • Many offenses (e.g. domestic violence, sexual offenses, white-collar crime) are under-reported.
  • This skews data towards offenses that are more visible and more frequently policed, leading algorithms to over-focus on those categories.
• Historical discrimination
  • Past policies (e.g. stop-and-frisk, discriminatory drug enforcement) can produce datasets where certain ethnic or social groups appear as “high risk”, simply because they were more heavily controlled, not because they are inherently more criminal.

When algorithms treat this data as ground truth, they risk recycling biased patterns as predictions, embedding discrimination into mathematical form.

---

3. Feedback Loops: Reinforcing Over-Policing

A central danger of predictive policing is the feedback loop:

1. Police historically patrol Area A more than Area B.
2. More crime is recorded in Area A (because more officers are present to detect it).
3. The algorithm learns that Area A is a “hotspot” for crime.
4. The system recommends even more patrols in Area A.
5. More stops, searches and arrests occur in Area A, feeding more data into the system.

Over time, the algorithm may self-validate a biased pattern: “we find more crime where we look more”. This creates the illusion of objective confirmation, even though the original pattern stems from policing choices, not neutral reality.

Such feedback loops can:

• Entrench racial or socioeconomic disparities,
• Normalize continuous surveillance of specific communities,
• Make it extremely difficult for those communities to ever “escape” the label of being high risk.

---

4. Technical Sources of Bias: Features, Proxies and Labels

Algorithmic bias in predictive policing is not only about data quantity. It also arises from which variables are used and how outcomes are defined:

• Problematic features and proxies
  • Variables such as neighbourhood, housing type, or prior contact with law enforcement may operate as proxies for race, ethnicity or poverty, even if these protected attributes are not explicitly included.
• Biased labels
  • Many models use “arrest” or “charge” as the outcome variable (the “label” to be predicted), not actual offending.
  • If arrests are already biased, the model is literally optimized to reproduce biased arrest patterns, not objective risk.
• Imbalanced and noisy data
  • Errors, missing values and inconsistent recording practices can disproportionately affect certain groups and skew predictions.

Technical adjustments (re-sampling, fairness metrics, feature selection) can mitigate some issues, but they cannot fully neutralize structural bias in how crime is policed and recorded.

---

5. Legal and Human Rights Concerns

Predictive policing raises deep concerns for criminal law principles and human rights:

• Non-discrimination and equality
  • If certain ethnic or social groups are systematically labelled as high risk, they may face disproportionate stops, searches, and police encounters.
  • Even when intent to discriminate is absent, indirect discrimination can occur through algorithmic criteria.
• Presumption of innocence
  • Risk scores can blur the line between suspicion based on concrete facts and suspicion based on statistical group profiles.
  • Individuals may be treated as quasi-offenders not for what they have done, but for what an algorithm thinks they might do.
• Privacy and surveillance
  • Predictive systems often rely on large-scale data aggregation: criminal records, social media, financial or mobility data.
  • This can erode privacy and support the expansion of preventive, data-driven surveillance.
• Due process and fair trial
  • When risk scores influence bail, sentencing or parole decisions, defendants must have a fair opportunity to challenge the underlying logic and data.
  • Black-box models with proprietary code and trade secrets can severely undermine this right.

---

6. Transparency and Explainability: Can We Understand the Algorithm?

Many predictive policing tools are complex and opaque:

• Some rely on proprietary software, where vendors refuse to disclose model details.
• Others use machine-learning methods that are technically difficult to interpret even when the code is available.

From a legal perspective, this raises questions:

• How can courts assess whether an algorithmic tool is reliable and unbiased if its workings cannot be scrutinized?
• Can an officer base “reasonable suspicion” or “probable cause” on a risk score they do not understand?
• How can affected individuals contest decisions if they cannot know why they were labelled high risk?

Transparency and explainability are therefore not merely technical virtues; they are constitutional and human-rights safeguards.

---

7. Predictive Policing and Criminal Procedure

Predictive policing also challenges criminal procedure law at various stages:

• Pre-contact stage – determining where to patrol and whom to monitor; algorithms can shape who is likely to be stopped at all.
• Reasonable suspicion / probable cause – officers may rely, explicitly or implicitly, on risk maps and scores to justify stops, searches or arrests.
• Pre-trial release and sentencing – risk assessment algorithms may be used to argue for detention or harsher sentences.
• Evidence and admissibility – courts must decide whether and how algorithmic outputs can be used as evidence, and under what conditions.

If predictive tools are treated as authoritative oracles, they risk becoming “invisible witnesses” whose credibility is assumed rather than tested. Criminal procedure must insist that algorithmic tools be subject to the same adversarial scrutiny as any other expert evidence.

---

8. Towards Safer and Fairer Use: Safeguards and Limits

The question is not necessarily whether predictive policing should be abolished entirely, but under what conditions, if any, it can be justified. Key safeguards include:

• Clear legal basis and democratic debate
  • Use of predictive tools should be grounded in legislation, with open democratic discussion of their purposes and limits.
• Impact assessments and bias audits
  • Independent assessments of disparate impact on protected groups; regular audits to detect and correct bias.
• Data governance
  • Careful curation of training data, including efforts to mitigate legacy bias and represent under-reported harms.
• Human oversight and discretion
  • Algorithms should support, not replace, human judgment.
  • Officers must be trained to critically interpret outputs, not blindly follow them.
• Transparency and contestability
  • Public information about the use of predictive tools, their objectives and general logic.
  • Meaningful ways for individuals and communities to challenge their deployment and outcomes.
• Substantive limits
  • Some uses – such as individual-level predictions leading to invasive interventions without concrete evidence – may be incompatible with fundamental rights and should be prohibited.

---

Conclusion: Risk Scores Cannot Replace Justice

Predictive policing presents itself as a neutral, data-driven solution to crime. In reality, it often translates old biases into new mathematical forms, creating a veneer of objectivity around unequal treatment. Algorithmic bias in policing is not just a technical glitch; it is a challenge to core principles of criminal law – equality, presumption of innocence, due process and accountability.

If predictive tools are to be used at all, they must operate within a robust framework of law, oversight and public scrutiny, with clear recognition that risk scores cannot substitute the moral and legal judgment at the heart of just policing. The future of criminal justice should not be surrendered to algorithms; it should be shaped by societies willing to use technology without abandoning their commitment to fairness and human dignity.