Predictive Policing: $300M and Racial Bias in 2026

São Paulo expanded the use of cameras with facial recognition and behavior analysis in 2026. Rio de Janeiro did the same. Together, the two public security departments injected R$ 300 million into predictive policing pilot projects (São Paulo Public Security Department, Annual Report 2026; Rio de Janeiro Public Security Department, Annual Report 2026). The promise is simple: use computer vision and statistical models to anticipate where—and when—crime will happen.

The effectiveness numbers exist. The problem is that the technology still carries a past of biases and an uncertain future regarding privacy.

The mathematics of anticipation: what the data shows

The most cited report of the year comes from the Urban Institute. The organization tracked 14 American cities that implemented predictive policing between 2023 and 2025. The conclusion: a reduction of up to 20% in property crimes, such as theft and vehicle robbery (Urban Institute, "Predictive Policing in American Cities: A Three-Year Assessment", 2026). The result is not uniform—some cities saw a 5% drop, others came close to 30%.

The market follows the enthusiasm. The AI-powered public safety software sector is expected to move US$ 12.8 billion in 2026 (MarketsandMarkets, "AI in Public Safety Software Market Report", 2026). Companies like ShotSpotter, Axon, and Motorola Solutions dominate the segment, selling everything from acoustic gunshot sensors to platforms that cross-reference historical police report data with real-time images.

The technical logic is straightforward. Computer vision algorithms process thousands of hours of video. They identify patterns: crowds at unusual times, vehicles driving in zigzag patterns, people returning to the same spot multiple times. The system cross-references this with the region's criminal history and generates a risk map. The officer receives the information on a tablet before heading out on patrol.

The Brazilian side: billion-dollar investment and local doubts

In Brazil, the official discourse is one of modernization. São Paulo and Rio de Janeiro spent R$ 300 million combined in 2026 to expand their camera networks (São Paulo Public Security Department, Annual Report 2026; Rio de Janeiro Public Security Department, Annual Report 2026). The justification is the same as in American cities: optimize patrols and respond faster.

But the local application exposes a problem that international reports already point out. Predictive models are trained with historical incident data. And historical incident data reflects decades of biased policing. If the police have always patrolled a neighborhood more, crime records in that neighborhood will be higher. The algorithm learns that it is a "risk area." The cycle feeds itself.

The AI Now Institute warns that predictive systems tend to project historical biases from training data, which can reinforce inequalities in policing (AI Now Institute, "Algorithmic Accountability in Public Safety", 2026). The warning is backed by research. The same institution analyzed predictive systems in operation in the US and Europe. Conclusion: 70% of them show measurable racial bias, reinforcing historical inequalities in policing (AI Now Institute, "Algorithmic Accountability in Public Safety", 2026). In Brazil, the debate is even more sensitive. The profile of the population stopped by police is already the target of constant criticism from social movements.

Accuracy vs. privacy: the dilemma no one has solved

There is another silent cost: mass surveillance. Cameras that identify faces, license plates, and behaviors generate a gigantic database about ordinary citizens. Most have never committed a crime. Still, their daily commutes, work schedules, and social circles are recorded on public servers or those of contracted companies.

Brazilian legislation treats the topic in a fragmented way. The LGPD protects personal data but opens exceptions for public security. In practice, citizens do not know what is filmed, where it is filmed, or how long the image is stored. Transparency is the exception, not the rule.

Companies in the sector try to respond. Axon, for example, publishes ethical impact reports on its technologies. IBM reduced its facial recognition operations after pressure from civil groups. But self-regulation has limits. The economic incentive is strong: the US$ 12.8 billion market (MarketsandMarkets, 2026) grows every year, and competition rewards those who deliver quick results, not those who raise difficult questions.

The future of patrol: human or algorithmic?

The central question is not whether the technology works. In part, it does. The 20% reduction in property crimes (Urban Institute, 2026) is relevant data. The question is what model of public security we want to build.

There is a possible path. Predictive systems can be externally audited. Training data can be corrected to reduce biases. Cameras can have clear image retention policies and restricted access. Transparency can become a contract requirement, not a courtesy.

But that requires political will and social pressure. The Brazilian cities that invested R$ 300 million (São Paulo Public Security Department, Annual Report 2026; Rio de Janeiro Public Security Department, Annual Report 2026) need to answer: who audits these algorithms? Who guarantees that an innocent citizen filmed today will not be harmed tomorrow by a classification error?

Conclusion

AI-powered cameras are a powerful tool. They reduce crime, optimize resources, and respond to a legitimate demand for security. But the effectiveness numbers come with an uncomfortable warning: 70% of the systems analyzed have racial bias (AI Now Institute, 2026). And the billion-dollar investment in Brazil—R$ 300 million in São Paulo and Rio alone (São Paulo Public Security Department, Annual Report 2026; Rio de Janeiro Public Security Department, Annual Report 2026)—demands more than technological enthusiasm. It demands external oversight, independent auditing, and a public debate about the line between protection and surveillance. Without that, the risk is trading street violence for a new form of injustice, silent and algorithmic.