Unmasking Algorithmic Bias in Predictive Policing: A Paper Explainer

Imagine you are an urban planner holding a magic map that predicts where crime will occur, directing patrol cars to those areas. This is no science fiction—"predictive policing" is already deployed in dozens of US cities. But this map may be a "biased mirror."

1. The Algorithm's "Tinted Glasses"

Consider a thought experiment: Community A was heavily patrolled over the past decade, producing more arrest records; Community B was rarely patrolled and has fewer records. An AI learning from this data concludes: "Community A has high crime—send more police!"

This is a feedback loop—the more you patrol, the more arrests you make, and the data increasingly "proves" you need more patrols. It is like someone staring into a funhouse mirror, believing the distorted image and dieting, with the dieting seemingly "confirming" the mirror.

2. How Researchers Dissected the Bias

Researchers Pronob Kumar Barman and Pronoy Kumar Barman of Jagannath University, Bangladesh, built a framework to quantify this bias.

Tool 1: Generative Adversarial Network (GAN)

• Generator: a "forger" trying to draw realistic crime hotspot maps
• Discriminator: an "appraiser" trying to tell real from fake

Through adversarial training, the generator learns to produce patrol points nearly indistinguishable from real data.

Tool 2: Noisy-OR Detection Model

A probabilistic model computing the probability of discovering crime when police patrol an area, accounting for the probability that crime actually occurs, police presence, and the chance of witnessing or reporting.

Tool 3: Four Fairness Metrics

1. Disparate Impact Ratio (DIR): DIR = P(detected | Black) / P(detected | White). Under the legal "four-fifths rule," DIR below 0.8 signals systematic underestimation of Black communities; above 1.25 signals overestimation. 2. Demographic parity difference: the raw gap in detection probability between groups. 3. Gini coefficient: borrowed from economics; 0 means perfect equality, 1 total inequality. 4. Bias Amplification Score (BAS): combines the above, penalizing cases with both directional bias and high inequality.

3. Striking Numbers

The study analyzed:

• Baltimore: 2017–2019, over 145,000 cases
• Chicago: 2022, over 233,000 cases

Baltimore: Dramatic Year-to-Year Swings

| Year | Average DIR | Meaning | |------|-------------|---------| | 2017 | 0.95 | Near parity, slight overestimation of Black residents | | 2018 | 0.079 | Severe underestimation of Black residents | | 2019 | 15,714 | Extreme overestimation of Black residents |

In 2019's "detection mode," white residents nearly vanished from the police radar—patrols were directed to Black neighborhoods while white communities received almost none.

Chicago: The Opposite Bias

Chicago's 2022 average DIR was just 0.22—systematic underestimation of crime in Black neighborhoods. The key insight: the direction of bias is not inherent to the algorithm but determined by the spatial distribution of training data.

4. Two Policing Modes Compared

• Detection mode (AI-driven patrol allocation): wildly unstable DIR (0.04 to 35,582), prone to feedback loops.
• Report mode (responding to citizen reports): relatively stable DIR (0.61 to 1.22). Citizens' "ground truth" acts as a corrective.

Community participation may be a firewall against algorithmic bias.

5. CTGAN Debiasing: Hope or Trap?

Using a Conditional Tabular GAN (CTGAN) to synthesize data for under-represented Black neighborhoods, the Black detection rate rose from 3.44% to 4.93%—but the white rate collapsed from 6.70% to 1.59%. DIR flipped from 0.513 (underestimating Black residents) to 3.106 (overestimating them).

Like a seesaw: under a fixed policing budget, more attention for one group necessarily means less for another.

6. Structural Roots of Bias

The study found strong socioeconomic associations:

• Correlation between a community's white share and detection rate: r = 0.83
• Correlation between a community's Black share and detection rate: r = -0.81

This is no statistical fluke but the structural legacy of decades of segregation, economic deprivation, and enforcement bias. The algorithm did not create these inequalities—it merely learned the distorted reflection in the mirror.

7. What Can Be Done?

Three policy recommendations from the study:

1. Annual audits: bias evolves over time; do not check only once at deployment. 2. Resource reallocation: pure data debiasing is insufficient—real policy and resource adjustments are needed. 3. Strengthen community reporting channels: citizen participation is the antidote to algorithmic feedback loops.

8. Deeper Reflection

The problem of predictive policing is fundamentally a collision between technological optimism and social complexity. We treat data as "objective," forgetting that data is itself a product of social processes.

As the researchers put it: historical crime data encodes the "spatial footprint of past enforcement behavior," not the "true spatial distribution of crime." When we train future algorithms on past bias, we get not prediction but self-fulfilling prophecy.

---

Paper information:

• Title: Unmasking Algorithmic Bias in Predictive Policing
• Authors: Pronob Kumar Barman, Pronoy Kumar Barman
• arXiv: 2603.18987
• Published: March 20, 2026