Nowhere is the intersection of race and AI more consequential, or more contested, than in policing and surveillance.

Facial recognition technology has become one of the most scrutinised AI applications precisely because its racial disparities are dramatic and its consequences potentially severe. Research by Joy Buolamwini and Timnit Gebru, published in 2018, found that commercially deployed facial recognition systems misclassified darker-skinned women at error rates up to 34.7%, compared to error rates of less than 1% for lighter-skinned men. The systems had been trained predominantly on datasets of lighter-skinned faces, and their performance reflected that training.

These error rates matter in the abstract. They matter catastrophically when facial recognition is used to identify criminal suspects. In 2020, Robert Williams, a Black man in Detroit, was arrested based on a facial recognition match that was incorrect. He was detained for 30 hours before investigators acknowledged the error. His was not an isolated case. The American Civil Liberties Union has documented multiple cases of wrongful arrests based on facial recognition misidentification, all involving Black individuals.

The deployment of facial recognition technology is itself racially unequal. Studies of US cities have consistently found that AI surveillance infrastructure, cameras, facial recognition systems, predictive analytics, is deployed more densely in communities of colour than in predominantly white communities. The justification offered is typically that these communities have higher crime rates. CRT would note that these communities also have higher policing rates, which produces more recorded crime, which justifies more policing, in a self-reinforcing cycle.

Predictive policing amplifies this dynamic. Algorithms that predict where crime will occur are trained on historical crime data, which is itself a record of where policing has historically been concentrated. The algorithm does not predict crime. It predicts prior policing. Deploying more police to the predicted locations generates more arrests, more data, and more predictions, of the same locations.

This is what CRT calls a feedback loop of structural racism. The discriminatory pattern is not introduced by the algorithm. It is already present in the historical data. The algorithm encodes it, legitimises it with mathematical authority, and makes it harder to challenge because it appears objective.

The consequences are not only individual, wrongful arrests, damaged lives, but structural. Racialised surveillance normalises the treatment of entire communities as inherently suspect. It produces a different relationship between the state and communities of colour than the relationship between the state and white communities. This difference is not incidental to how AI policing systems work. It is part of what they do.

Reflection question: Who decides where AI surveillance is deployed? What assumptions about safety, risk, and community are embedded in those decisions?