Technology
The Impact of Algorithmic Bias in Predictive Policing
Quick fact
A 2016 study by ProPublica found that predictive policing software (COMPAS) was twice as likely to falsely flag African-American defendants as future criminals compared to white defendants, illustrating algorithmic bias in action.
Why this is interesting
Imagine a police department using a crystal ball to predict crimes—except the ball was crafted from a history of biased decisions. Would you trust it to be fair?
Read the full explanation
Understanding The Impact of Algorithmic Bias in Predictive Policing
Predictive policing algorithms are trained on historical crime data to identify areas or individuals at high risk of criminal activity. The core idea is that if we know where crime has happened, we can anticipate where it will happen next. However, the crucial flaw is that this historical data is not a neutral reflection of reality—it's a product of past policing practices, which may include racial profiling, over-policing of certain neighborhoods, and under-reporting of crimes in others. When an algorithm learns from such skewed data, it effectively codifies those biases. The system then produces predictions that lead officers to patrol those already-targeted areas even more. More police presence leads to more arrests and reports in those areas, which in turn is fed back into the algorithm as 'evidence' of criminality, creating a self-fulfilling prophecy. This is the feedback loop: biased data → biased predictions → biased policing → more biased data. As a result, rather than objectively reducing crime, the algorithm can systematically over-police minority communities, reinforcing existing social inequalities.
A deeper explanation
The mechanism of algorithmic bias in predictive policing is rooted in how machine learning models optimize outcomes based on training data. The model's goal is typically to predict the likelihood of a crime at a location or for an individual. The algorithm learns patterns (like 'crime is frequent in area X') from the historical data. If the data is biased—for example, if area X was heavily policed in the past, leading to many recorded incidents—the model will learn that area X is high-risk. This is not a malicious choice but a statistical consequence. The problem intensifies because the model's own predictions drive police deployment, which generates new data in the same locations, creating a positive feedback loop. The model 'sees' more crime where it is already looking, reinforcing its initial bias. Moreover, the model can suffer from disparate impact, meaning that even if the algorithm is accurate overall, it may have different false-positive and false-negative rates across demographic groups. For example, it might flag more Black individuals as high-risk even when their actual offending rates are similar to others, due to the biased data. This violates the principle of procedural justice, which emphasizes fairness in the process of law enforcement. Understanding this mechanism highlights that algorithmic fairness is not just a technical problem to be solved with better math; it's a systemic issue that requires careful consideration of historical context, data quality, and the societal impact of automated decision-making.